TL;DR: Consensus expects aggregate hyperscaler free cash flow to decline through 2027, then roughly triple its 2024 peak by 2030. Q2 strengthened the case for the decline, but the rebound’s composition matters more than its size: cash flow generated from utility-like compute deserves a utility multiple, not a software one. Yet forecasts still assume that all four companies recover in proportion. The earnings instead offered the first clear evidence of a structural separation. Hyperscalers are becoming the agentic economy’s landlords, financiers, and counterparties. As part of this metamorphosis, the value migrates toward the stack’s two ends: the silicon and power below, and the irreplaceable customer-facing context above. The models and the harness in the middle risk being commoditized or subsidized in the contest for those positions. The question is not how large the 2030 cash-flow bars become, but which are achievable, at what margins and from which defensible layers.

Subscribe now

The four largest hyperscalers reported earnings within nine days of one another. All four beat revenue expectations and raised or extended capital-spending guidance. Together, their guidance now points to $720-745 billion in Capex in 2026, while Goldman Sachs argues that the 2027 consensus remains too conservative, with a base case approaching $1.1 trillion.

Yet the market delivered four different verdicts. Microsoft recorded the largest single-day increase in market value in stock-market history, nearly $450 billion, after a quarter whose optics were flattered by an accounting change that extended the estimated useful life of its data centers from 15 to 25 years. The market punished Alphabet, which delivered the strongest operating earnings, by 7%. Amazon rose 15% despite reporting the first negative trailing free cash flow of its AI investment cycle. Meta fell 8% even as its advertising machine converted AI spending into auction yield more directly than any of its peers.

When it comes to the hyperscalers, the ROI question is usually framed too narrowly: how much AI revenue is being generated relative to Capex, and how quickly will free cash flow recover? But a dollar invested in commodity compute, proprietary silicon, a frontier model, or an irreplaceable customer substrate does not produce the same margin, durability, or multiple.

The consensus free-cash-flow curve obscures that distinction. It shows combined free cash flow declining through 2027, then roughly tripling its 2024 peak by 2030, with all four companies recovering in broadly proportional terms. But cash flow earned from infrastructure volume is not equivalent to cash flow generated by proprietary silicon, enterprise identity, or an irreplaceable commerce, search, or attention graph.

That the market has stopped pricing the Big Four as one trade is progress. In May, I argued in The Hyperscaler Reckoning that the commoditization of compute and the erosion of application interfaces would not affect the four companies equally because their Capex was not buying the same strategic positions. But look at how the symmetry broke: investors sorted primarily on near-term monetization - who could show the clearest revenue against the spending - rather than on who owns the layers that make those returns durable. By that logic, they punished the company with arguably one of the strongest structural positions and most rewarded a print whose reported economics were flattered by an accounting change. The sorting has begun, but on the wrong axis.

The right axis is structural: which layers the Capex strengthens, where scarcity persists, and where competition drives margins down. Answering those questions requires reviewing the consolidated companies and separating the different businesses within them.

I apply here the Orchestration Economics framework that divides each hyperscaler into four layers:

• Layer 0 is silicon: the chips and physical inputs, including Google’s TPU and Amazon’s Trainium.

• Layer 1 is intelligence: the frontier models and the temporary capability premiums they command.

• Layer 2 is the compute substrate and the harness: the infrastructure that serves the models and the orchestration runtime that turns them into working agents. These two functions currently sit together but may develop very different economics.

• Layer 3 is the proprietary context in which intent originates and outcomes are executed: Search, Office, and enterprise identity; Amazon’s store and fulfillment network; and Meta’s attention graph.

The question is therefore not whether the combined 2030 free cash flow bar is achievable. It is which company’s portion of that bar is real, at what margin and with what durability. Answering that requires taking the companies apart, layer by layer. I’ll begin with where the spending is landing and how it is being financed.

The buildout becomes infrastructure

The quarter confirmed a projection from the AGNT Manifesto: the buildout would consume most of the hyperscalers’ operating cash flow, pushing free cash flow toward zero and, for some, below it.

Amazon’s trailing 12-month free cash flow turned negative at minus $7.6 billion. Alphabet’s record $44.9 billion of Q2 Capex exceeded operating cash flow. Meta’s quarterly free cash flow fell 91%, with some sell-side models projecting negative cash flow through 2027. Microsoft remains the exception, generating $19.6 billion, although 23% less than a year earlier.

The more revealing change is not that free cash flow has fallen. It is that the buildout has begun to outgrow the cash generated by the businesses financing it.

Alphabet raised $84.75 billion of equity while suspending buybacks. Meta halted repurchases, issued $25 billion in bonds, and disclosed roughly $52 billion in maximum exposure and guarantees across its off-balance-sheet data-center vehicles before its newest venture with BlackRock has even been quantified. Amazon’s long-term debt nearly doubled in six months. According to FactSet, debt now finances 32% of trailing Capex across the hyperscaler complex (including Oracle), up from 9% in fiscal 2024. The companies are not running out of capital, but a buildout expected to finance itself through operating cash increasingly depends on equity, debt, private credit, leases and guarantees.

Spending continues to rise for two reasons: demand exceeds supply, and the inputs themselves are getting more expensive.

All four companies described themselves as supply constrained. Jassy said that even at $220 billion in Capex, Amazon would lack enough capacity to meet demand through 2027. AWS, Microsoft and Google Cloud together report nearly $1.7 trillion of commitments. Meanwhile, Amazon attributed its latest $20 billion increase in Capex guidance to memory costs. Memory scarcity raises both the prices hyperscalers collect and the capital required to create new supply.

Then there is accounting, which exposes the different clocks inside the buildout. Microsoft is extending the estimated useful life of certain data-center assets from 15 to 25 years, affecting depreciation and moving approximately $15 billion out of reported calendar-2026 Capex without changing the contracted capacity. In the same week, Jassy cited a sub-three-year payback on servers Amazon depreciates over five years, installed inside shells that may operate for decades.

Those figures measure different things, but that is the point. The AI buildout contains chips that may become obsolete within years, servers expected to repay their cost before retirement, and buildings, power systems and leases that remain on the balance sheet for decades. It no longer resembles the financial architecture of an asset-light software business. It increasingly resembles telecommunications, power, and other capital-intensive infrastructure.

The strongest objection to this thesis is AWS itself. Cloud was called a commodity in 2012 and never became one. AWS sustained operating margins above 30% because applications accumulated data, dependencies, and operating history that made moving expensive, risky, and slow. Perhaps AI infrastructure is simply cloud again, one order of magnitude larger.

Jassy gave that defense its clearest form yet on Amazon’s Q2 earnings call. The physical shell is built once and hosts successive generations of improving equipment. Servers are ordered months rather than years ahead of demand, repay their cost in less than three years and serve capacity largely contracted on five-year terms. Demand visibility reduces the danger of empty data centers, while rapid payback limits the capital exposed to obsolescence. AI margins, he said, are tracking core cloud margins at the equivalent stage of development.

It is a serious argument. But it answers the risk of stranded capacity more convincingly than the risk of falling prices.

Model-serving workloads are more standardized and portable than the applications of the SaaS era. Models can be served across multiple clouds, and open-weight models can increasingly be deployed wherever the economics are most attractive.

The buyers are different, too. Traditional cloud sold to millions of enterprises, most with limited bargaining power. A disproportionate share of frontier AI demand comes from a small number of laboratories that negotiate at enormous scale and increasingly participate in the design of the silicon they consume. Anthropic buys capacity from all three major clouds. OpenAI ended Azure’s exclusivity. The suppliers financing the infrastructure are also financing its largest customers.

Most importantly, today’s scarcity rents are attracting the capital that can eventually eliminate them. Hyperscaler balance sheets, special-purpose vehicles and private-credit funds are financing new capacity beyond the incumbents’ historical rationing discipline. Jassy’s contracts can protect utilization. They cannot guarantee the price at which each new generation of equipment will be sold.

That is the distinction between the ROI question and the layer question. The first asks whether current contracts allow today’s servers to repay their cost. Amazon has made a persuasive case that they do. The second asks whether the pricing power behind those contracts survives once capacity expands, and workloads become more portable. The terminal multiple turns on the second.

The reported backlogs illustrate why the distinction matters. Microsoft has disclosed $678 billion of contracted commitments, AWS $496 billion, and Google Cloud $514 billion. These figures provide powerful evidence of demand and significantly reduce near-term utilization risk. But the agreements are economically layer-agnostic. The same committed dollar might ultimately purchase a commodity GPU hour, proprietary silicon, a managed model service, or a higher-margin orchestration product. None of the three companies discloses enough of the mix to determine where the backlog’s eventual margins will reside.

Counterparty quality also matters. Amazon has committed a $20 billion facility to Anthropic tied to compute delivery, while roughly $4 of its $5.75 in headline quarterly EPS came from the appreciation of its Anthropic stake. The demand is real but economically interdependent: Amazon finances a customer whose commitments support the backlog that justifies its Capex, while the customer’s appreciation flows back through Amazon’s income statement. More to come on that next week.

Backlog proves the infrastructure will be used, but not which layer captures its value. For that, the analysis must move above compute to the models and harnesses sold through it.

The middle gets cheaper

Kimi K3 forces an honest scoring of Layer 1, the “intelligence” itself.

Released with full open weights, Moonshot AI’s model ranked fourth on the Artificial Analysis index and led several agent-relevant automation benchmarks. Its exact position will change; that is the finding. The frontier moved twice in one quarter: frontier margin is a moving window, not a seat any company permanently owns.

Anthropic and OpenAI currently hold that window. Google is the only hyperscaler consistently contesting it, aided by the integration economics of TPU and Gemini. But K3 compresses the economic space below the frontier: buyers can increasingly pay for genuinely scarce capability or begin from an improving open-weight baseline. For three of the four hyperscalers, Layer 1 is becoming less of a race to win and more of a procurement problem.

That moves the contest to the harness.

Nadella stated the doctrine explicitly on Microsoft’s earnings call: “You got to keep your harness separate from the model…any model at any given time is swappable.” Within days of K3’s release, Microsoft was reportedly evaluating it for use inside Copilot - a switch that reports suggest could save it as much as $600 million a year in inference costs currently paid to OpenAI and Anthropic. If it substitutes a cheaper model without changing Copilot’s price, the spread accrues to Microsoft initially.

Here is where I revise my May position, based on the evidence that has emerged since.

The conclusion at the time was that the spread would widen mechanically as models became cheaper, causing margin to pool in the harness. That is plausible in the near term. But it is not a sufficient three-year thesis. A spread persists only if the layer capturing it possesses a barrier that competitors cannot reproduce, bypass, or subsidize away.

Generic harness functionality does not meet that test. Laboratories are training, planning, and tool use into models; open runtimes reproduce previously proprietary features; and Nvidia, MCP, and free agent SDKs subsidize orchestration to strengthen adjacent businesses. Hyperscalers, labs, and chip companies can all treat the harness as a customer acquisition channel rather than a standalone profit pool.

The deeper problem is that the frontier itself absorbs the harness. Each model generation internalizes more of what orchestration used to do: planning, tool selection, memory management, multi-step execution. What required an elaborate scaffold around last year’s model ships inside next year’s. A layer the models are steadily absorbing cannot hold a moat - whatever the harness does well becomes a training target for the next release.

What the models cannot absorb is authority. Who authorized an agent, with what permissions, spending whose money, accountable to whom - these are property-rights questions, not intelligence questions, and they grow more acute as agents grow more capable, not less. Identity, permissioning, governance and verified outcomes sit within the enterprise’s control structure, outside the weights.

An orchestration layer that holds that authority - and the state that accumulates around it: persistent memory, evaluation history, learned operating policies, verified outcome data - stops behaving like a harness and begins behaving like a platform. In the language of this framework, it migrates economically from Layer 2 toward Layer 3. The product category has not changed. The source of its defensibility has.

Margin moves out of undifferentiated models as open-weight performance improves. It eventually moves out of generic compute as scarcity normalizes. Value can pass through the harness, but it stays only where orchestration holds authority, and the proprietary state and context that accumulate around it. That is precisely Microsoft’s play, aided by the rise of open-source models in the enterprise.

The ends start printing

If the middle faces margin compression, durable value should pool at the two ends of the stack - a barbell. Q2 supplied the first evidence that it is.

Let’s start below, at the silicon layer.

Google delivered TPU systems directly into customer data centers for the first time - recognizing its first, still-modest revenue from external TPU sales, with the bulk of contracted revenue expected from 2027 - while a merchant ecosystem formed around them: a Blackstone joint venture selling TPU compute, Anthropic expanding its commitment by a further 3.5 gigawatts and a reported $36 billion debt package financing its purchases. Layer 0 is developing its own customer base and capital markets.

Amazon’s custom silicon is appearing through contracts rather than direct sales. The associated chips business reached a $25 billion annualized run rate, while agreements with OpenAI and Anthropic tied performance obligations to the Trainium roadmap. Google and Amazon therefore possess an option Microsoft and Meta lack: their Capex can create an externally monetized silicon architecture rather than remaining solely an internal cost.

That option does not guarantee excess returns, but it means the same Capex dollar buys different positions across the four companies.

The other end of the barbell appears in Layer 3.

In May, I argued that moats located beneath the interface could survive a change in operator, while moats whose value resided principally in the interface were more exposed. Amazon supplied the clearest evidence. Shoppers engaging with sponsored prompts inside its agentic shopping flow converted 48% more often and spent 21% more than shoppers who did not click one.

The result is early, but its direction matters: when an agent replaces the shopping interface, Amazon’s selection, fulfillment, payments, and trust become inputs the agent needs. The interface changes while the substrate - and its monetization - survives.

Meta’s attention substrate produced a related result. Ad revenue rose 27%, while one million businesses reportedly transact through its agents each week. Zuckerberg described a model in which businesses “only pay us when we achieve results for them”, allowing Meta to auction compute as it auctions advertising. This is not a cloud proposition: it applies Meta’s audience, demand signals and optimization history to agentic outcomes rather than impressions.

Google is more ambiguous. Search revenue grew 17%, but management’s insistence that AI features still send “billions of clicks” to websites reveals the pressure point: the click remains central as the product moves toward answers and actions. Google’s defense is the index, commercial intent, advertiser demand, and transaction rails beneath the interface - not the preservation of the interface itself.

Google is therefore building transaction rails alongside the interface that may eventually reduce the importance of the click. Its advantage is not necessarily that the traditional search experience survives unchanged. It is that the index, commercial intent, advertiser demand and payment relationships beneath Search can be reorganized around an agent-mediated transaction. Look at what is happening inside the search box itself: AI Mode now generates working mini-apps on the fly - dashboards, trackers, custom tools built from a natural-language prompt - and connects directly to third-party apps to complete tasks in place. The box that once returned links is learning to produce software and execute transactions, keeping intent origination and its monetization inside Google’s substrate even as the click fades.

Microsoft illustrates the more difficult side of the dividing line.

Microsoft 365 seat growth has slowed to 6%, with an increasing share of revenue growth coming from price. GitHub Copilot and Copilot Cowork, the products most directly exposed to autonomous software work, have both moved beyond pure per-seat pricing. GitHub Copilot revenue accelerated more than 60% quarter over quarter as usage-based elements expanded.

Nadella described the shift like this: “…we are also evolving our business model beyond per seat to per seat plus consumption.”

That sentence reflects the argument I made in May about the inverse correlation between the per-seat model and agentic success. Agents can produce additional work without adding human employees. As agentic systems become more capable, the amount of activity flowing through a product can rise while the number of seats remains flat - or even declines. A pricing model tied only to human headcount cannot fully capture the value generated by synthetic labor.

It is too early to call this a retreat from seat pricing. Rather, it seems more like a managed fallback, executed while seat optics still look strong. It is what a rational incumbent does when it believes the old model’s clock is running. Microsoft may be layering a new revenue model on top of an enduring subscription franchise rather than replacing it. But the direction is clear. The company is beginning to search for a unit of value better suited to a workforce whose productive capacity can expand without occupying another seat.

Share

Four metamorphoses, four multiples

The hyperscalers entered this investment cycle as a single AI trade. They are emerging with increasingly different anatomies.

Google holds the broadest collection of defensible positions. Cloud grew 82% at a 35.6% operating margin; Gemini keeps Google in the frontier race; direct TPU sales create a merchant-silicon option; and Search continues to compound even as its traditional interface comes under pressure. Google owns both ends of the barbell plus the frontier option. The market sold it down 7%.

Amazon combines proprietary silicon below, an agent-amplified commerce and fulfillment substrate above, and the current compute harvest between them. Its principal caveat is concentration: a substantial part of its backlog is anchored by laboratories that Amazon finances, supplies, and marks through its own income statement. The demand is real, but more economically interdependent than the headline backlog suggests.

Microsoft is attempting the most unusual transformation. Agent 365 extends Entra’s identity system from human employees to synthetic ones, covering registration, permissions, metering, and auditing. Those functions govern authority and accountability - who permitted an agent to do what - and become more important as agents grow more autonomous.

Microsoft is trying to rebuild its application franchise one layer down. Agent 365 is an attempt to make Microsoft the system through which enterprises manage synthetic workers, as Office managed human ones. The test is whether agents built on other companies’ runtimes register into Entra. If they do, Microsoft controls a durable Layer 3 position. If not, it risks being left with an exposed work interface above and a contracted compute utility below.

Meta should not be judged as an aspiring external compute provider. Its infrastructure program is an internal investment in the attention and advertising substrate it already owns. AI is already improving targeting and pricing; the next step is to extend the advertising model from selling access to attention to charging for completed outcomes. The attention substrate is proven. Whether the resulting outcome market can justify the capital required remains an assertion.

The market’s ordering - Microsoft and Amazon rewarded, Google and Meta punished - therefore mixed structural judgment with quarterly optics. Amazon’s reward was directionally supported, and Meta’s punishment reflected a legitimate gap between current spending and future returns. But Google’s sell-off discounted the company with the broadest layer positions, while Microsoft’s record gain priced in a control-plane outcome that has not yet materialized.

The consensus free-cash-flow chart repeats the same analytical error at the other end of the forecast. The trough keeps moving as Capex estimates rise. The terminal bars assume something close to today’s scarcity margins survives the capacity arriving by 2029. And all four companies recover roughly in proportion.

The J-curve is not wrong so much as undifferentiated. If it lands, it will land at four different heights, at four different margins and deserving four different multiples. It is the dispersion within the curve that matters, not just the curve itself.

A hyperscaler is no longer a single company with a single cash flow. It is a portfolio of layer positions operating on different clocks: a compute harvest whose scarcity premium may expire, merchant-silicon options whose markets are forming, a frontier position that resets quarter by quarter, and Layer 3 substrates already producing returns.

The stack is coming apart in the financing, accounting, contracts, and disclosures. The multiples are the last thing still consolidated. They will not stay that way.

DISCLAIMER: The views and opinions expressed here are those of the author alone and are based on publicly available information. They do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Past performance is not indicative of future results. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

TLDR - Stripe is reportedly in talks to buy OpenRouter for around $10 billion, about 8× its most recent valuation. Nobody pays that for a take-rate API aggregator. Open-weight models are multiplying, and intelligence is getting cheap; what stays scarce is the choice of which model to use, weighed on price, task-fit, latency and jurisdiction, and then turned into an enforceable transaction. OpenRouter sits at the moment of selection and sees what the whole market is buying. Stripe sits at the moment of settlement and supplies the operating context: wallets, metering, mandates, finality. Together they could be the transaction layer of the machine economy, where agents hire minds the way humans hire freelancers. Orchestration control points like that don't move often.

Subscribe now

Last week, the Wall Street Journal reported that online payment giant Stripe, currently valued at around $159 billion, is in talks to acquire OpenRouter, a company of a few dozen people that routes developer requests across hundreds of AI models, in a deal reportedly worth close to $10 billion.

OpenRouter raised its Series B in May at a valuation of $1.3 billion, with investors including Databricks, which also then reportedly made its own bid to buy the company, according to The Information. A repricing of almost 8 times suggests that Stripe is not simply valuing OpenRouter as an API aggregator that collects roughly 5 percent of the inference spending passing through it. [Note: The talks remain unconfirmed and may collapse or close on different terms].

As I wrote last week, the string of open-weight model releases culminating with the release of Kimi K3, which became the first open-weight model to beat the closed frontier on an independently run leaderboard, demonstrated how quickly Chinese labs are compressing the frontier release cycle.

Suddenly, a debate over open source and open weights went from thoughtful ponderings of theoretical situations to overtones of a crusade. Anthropic’s leadership accused Chinese companies of IP theft, and the U.S. government seemed to hint at some possible action to limit access to open-source models. This prompted a remarkable industry counter-attack led by Nvidia CEO Jensen Huang, who used his first post on X to share an open letter defending open weights as a foundation of American AI leadership.

The letter has since been signed by some twenty-five companies including Microsoft, Meta, IBM, OpenAI, and Google. Anthropic was initially conspicuous by its absence. On Monday, Anthropic CEO Dario Amodei published a statement emphasizing that the company had never sought an open-source ban while explaining his reasons for not signing the statement: “I don’t agree with the letter’s assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers. It seems at least as likely to me that the opposite will be true.”

Lost amid the protests and counterprotests is the reality on the ground that can be tracked on OpenRouter: Chinese-origin open models have gone from less than 2 percent of traffic in late 2024 to a weekly peak of 46 percent by mid-2026. During the same period, US models’ share of that same traffic fell from roughly 70 percent to 30 percent.

For OpenAI, Anthropic and Google, the threat is not that open models replace the frontier everywhere, but that routers increasingly reserve their expensive models for the hardest tasks while diverting the far greater volume of routine work toward cheaper alternatives.

This inversion goes beyond denoting a change inside the competitive ranking of AI models to the entire economic architecture around intelligence. As capable models proliferate and inference prices fall, scarcity migrates away from producing intelligence and toward deciding which intelligence to use, under which constraints, and how to turn that choice into an accountable economic transaction.

In other words, scarcity may be migrating toward the things a company like OpenRouter does. But whether that position is worth anything depends on a key question: can model selection eventually be reduced to commodity plumbing, or does it remain a defensible judgment? If the market settles and prices stabilize, model capabilities become predictable, and the optimal choice for each task can be written into a fixed set of rules, then routing becomes a feature that can be replicated, open-sourced, or bundled away. OpenRouter would be useful infrastructure, but hardly a $10 billion company.

That valuation makes sense only as a bet that the model market will remain volatile enough to prevent the optimal choice from becoming fixed. And thanks to the rise of open source and weights, we now see hundreds of models improving and repricing at different speeds, with their relative performance changing across different tasks, latency requirements, and jurisdictions. The present volatility means the potential advantage lies in the accumulated evidence of which models users choose and why. The more OpenRouter observes, the better situated it becomes to route the next request.

But it still does not explain why Stripe would pay such a premium to own it.

To understand that, we must turn to the core thesis of Orchestration Economics, which holds that AI demand will increasingly come from machines that are becoming actors, eventually leading to agents purchasing cognition on behalf of other machines. Picture millions of small, continuous decisions in which model quality, price, latency, jurisdiction, and spending authority must be reconciled in real time. In that economy, model selection becomes the margin.

In this Agentic Era scenario, Stripe doesn’t simply view OpenRouter as a thin bit of API plumbing that will add a 5 percent toll on rapidly deflating inference to the income statement. Instead, it’s more likely that Stripe sees an opportunity to fuse the moment an agent chooses which mind to hire with the infrastructure that completes the transaction. OpenRouter supplies the market-wide flow required to make the routing decision. Stripe supplies the operational context that makes its consequences enforceable and monetizable.

In this article, I want to examine why routing becomes strategically valuable as intelligence grows abundant, whether OpenRouter’s view across a volatile model market can form a defensible position, and why Stripe may be uniquely placed to monetize it. The larger question is whether connecting the selection of intelligence directly to the payment infrastructure behind it creates the control point through which the emerging machine economy will operate.

What OpenRouter Does, and Why AI Model Routing Matters

OpenRouter is one API endpoint, compatible with the interface every developer already knows, standing in front of four hundred models from sixty inference providers. When you send a request, the router selects where it runs, handles failover when a provider degrades, compares price and latency across the field, and consolidates the whole mess into a single bill.

Eight million developers use OpenRouter. Volume has grown from five trillion to roughly twenty-five trillion tokens a week in six months, a pace that annualizes to a quadrillion tokens. The business model charges roughly 5 percent on the inference flowing through. The company owns no GPUs and trains no models, and is purely an intermediary.

Described that way, routing sounds like a handy tool. And in 2023, it was. Open source redefined that role by changing the supply of intelligence. The simplistic explanation that “open models caught up” misses the larger structural transformation.

Open source caused one axis of variance to converge while detonating three others:

First, price variance exploded. Near-equivalent capability now trades across a 10x–40x range: self-hosted open-weight inference at a few cents per million tokens against $30 at the proprietary frontier, repricing weekly. When quality was scarce, price dispersion didn’t matter. You paid what the capable model cost. Now that quality is abundant, price dispersion is the entire game.

Then, task variance persisted along a predictable axis. The convergence is real but not uniform. Open models match or beat proprietary systems wherever verification is cheap: code that compiles, math that checks, and retrieval that grounds. The proprietary premium survives where verification is expensive: long-horizon reasoning, multi-turn business judgment, the weakly-verifiable professional work where Fable 5 still reigns. Distance-to-verifier, not benchmark scores, now governs which model wins which task. Nobody publishes that mapping, task by task, week by week. It has to be learned from live traffic.

Finally, jurisdictional variance has become a major factor. Which model may legally serve which customer, in which jurisdiction, with which data. Two years ago, that question did not exist as a routing input. After the release of Kimi K3, the furious open-model debate, the open letter, the hints at restrictions, and the long-term traffic migration toward open models, it is a first-order constraint that changes with the news cycle. Routing has acquired a foreign policy variable with an economic cost.

None of this makes models commodities, and the argument doesn’t require them to be. What a router commoditizes is procurement, the act of buying intelligence, rather than capability itself. Quality, reliability, and enterprise trust still carry a premium, which is exactly why the choice is hard. The inflection made quality abundant and the choice space four-dimensional: price, task-fit, latency, jurisdiction, across four hundred models whose relative positions never stop moving.

In this framing, the deal would potentially be a merger of two routers.

OpenRouter routes intelligence across models and inference providers. Stripe routes money across businesses, customers, and financial networks. Combined, they would form one interface that selects the model, observes consumption, meters the tokens, prices the call, and settles the transaction. Whether that interface is worth $10 billion depends on whether the selection it performs remains defensible or gets commoditized.

When AI Model Routing Becomes Valuable: Human Demand vs. Machine Demand

OpenAI’s first large-scale study of Codex usage gives an early view of how quickly agentic demand can multiply. More than 10 percent of users now operate at least three agents concurrently during a typical week, while OpenAI’s most intensive users generate approximately 71 cumulative hours of agent runtime in a single day by running multiple workflows in parallel.

The study does not yet describe an autonomous machine economy, but it shows the transition that makes one economically consequential: AI demand expanding from discrete human prompts into continuous, concurrent streams of machine work.

Yet as this agentic demand is beginning to explode, the strategic value of routing depends on who is buying the intelligence. Inference demand is splitting into two economies with radically different unit economics:

The first economy is human-priced: agents drafting equity research, adjudicating claims, and resolving tickets. This is work that displaces labor billed by the hour. The empirical record here is blunt. The first large-scale study of agents in production [See: Mapping Agents in Production; Revised June 4, 2026], which we have repeatedly used as a key datapoint in this publication, found teams overwhelmingly defaulting to the most capable proprietary models, because inference cost is a rounding error against the expert the agent augments. Only three of twenty case studies used open models at all, and those under cost or regulatory duress. In the human economy, model selection is close to a solved problem: use the most capable model available. Cost is usually noise.

The second economy is machine-priced: This one is new. In late January, Moonshot shipped swarm orchestration trained into the weights of K2.5. Days later, Moltbook, a social platform for autonomous agents, registered over a million agents within seventy-two hours. A subsequent investigation concluded that most of those accounts were scripted shells, and skeptics declared the episode theatre. They corrected the sociology and missed the economics. Fake or not, the accounts burned real compute, and the serious evidence was never the spectacle anyway. It is the infrastructure going in beneath it. Amazon now wires programmable agent wallets, with session-level spending limits, directly into its agent runtime. The infrastructure is being assembled for agents to become economic actors, even if the scale and timing of that demand remain uncertain.

In the agent world, the unit economics invert, because nobody’s salary anchors the value of a machine-to-machine interaction. Compute is not noise against labor. Compute is the cost of goods sold for every orchestrated workflow. At $30 per million tokens, a million-agent swarm is an impossibility. At open-weight prices, it runs continuously. Agent populations can double in days, with no evenings and no weekends. The open-source inflection is not simply a supply-side event that changed how routing works. It is the demand-side event that created the economy in which routing matters. In the human economy, model cost is often noise. In the machine economy, it is the margin.

The question is whether this makes routing a defensible position or just a useful feature. The open-model inflection has transformed model selection from a quality question (which model can do this at all?) into a portfolio decision: which point on a shifting four-dimensional frontier is optimal for this call, now, under these constraints? If that frontier eventually stabilizes, the answer can be encoded into a routing table and bundled into infrastructure. OpenRouter becomes valuable plumbing, but plumbing nonetheless.

The case for defensibility rests on the opposite possibility: that the frontier keeps moving too quickly for any static map to remain useful. In that environment, the better analogy is not indexing but market-making. The market-maker’s advantage does not lie principally in an algorithm that competitors can reproduce. It lies in seeing the flow. Whoever observes more activity across the market can respond more quickly as prices and preferences change.

OpenRouter sees twenty-five trillion tokens a week, providing a broad view of how the market purchases intelligence. Each lab sees demand for its own models. The router sees activity across the distribution. Its public rankings have already become something resembling the industry’s tape: a source of price discovery watched by the labs themselves. Routing policies refined against that flow should improve with volume. A system that has routed a quadrillion tokens has evidence that a new entrant does not.

OpenRouter’s position therefore depends on sustained volatility across the model frontier. Four hundred models are changing price and position across multiple dimensions, and every shift causes yesterday’s routing map to depreciate. That depreciation is both the weakness and the potential moat: the map never stays valuable for long, but only platforms with continuous, market-wide flow can keep it current. Stripe, in other words, would be buying an asset that is long model-frontier volatility. Judging by the release calendar that produced K3 and GLM-5.2, that volatility is currently well supplied.

Two things could break this thesis. First, meta-harnesses like Databricks’ newly open-sourced Omnigent could absorb model selection as a bundled feature, and workflows arrive at the router as pre-decomposed calls made upstream. The router would then slide toward invisible infrastructure.

Second, the model frontier could stabilize. If release cycles slow, prices converge, and performance by task becomes predictable, model positions settle into a printable table, and OpenRouter’s live information advantage loses much of its value. The two signals to watch are therefore where the routing decision originates and how quickly the frontier continues to move.

Even if routing is still valuable, however, that does not necessarily make it a good business. The machine economy runs, by construction, on the cheapest tokens available. A router earning 5 percent of prices deflating tenfold a year, on traffic selected for costing as little as possible, could win the intellectual argument and still lose the income statement. On OpenRouter today, a single proprietary provider accounts for roughly an eighth of tokens but nearly half of revenue, and that is exactly the traffic most able to go direct. Five per cent of nothing is nothing.

Answering that objection requires conceding its premise: the inference take may not be the real business.

What the routing position holds is the origination point: the moment a machine-demand workflow begins, where the agent chooses which mind to hire. Many of those workflows will ultimately produce an economic transaction whose value does not decline in lockstep with token prices. The larger opportunity is therefore not simply to collect a percentage of inference spending, but to link the purchase of cognition to metering, authorization, and settlement.

Which is where Stripe comes in.

The Absorption: How Stripe Turns OpenRouter Into Settlement Infrastructure

Routing may be strategically valuable without being sufficient, on its own, to support a durable business. OpenRouter can identify the moment at which an agent chooses which intelligence to use. Extracting the full value of that position requires the operational context to authorize the purchase and enforce its consequences.

To analyze OpenRouter’s position, we can see where it is situated along the Three Rings of the Agentic Enterprise.

OpenRouter, standing alone, is a Ring Two asset of unusual quality. It has genuine proximity to the machine-demand moment of choice and benefits from a flow advantage that compounds as more traffic passes through it. But its data-driven edge remains market context: a continuously updated picture of model demand that depreciates as the frontier moves and could theoretically be replicated by another player able to replicate a similar flow. It is a remarkable position that is also structurally exposed. The bear case described in the previous section is what valuable routing without a stronger mechanism for capturing that value looks like on an income statement. Each tremor pushes scarcity one ring outward, leaving a standalone router on the layer being compressed.

Stripe, standing alone, occupies the opposite position. It sits downstream of intent, executing instructions issued by upstream orchestrators, yet controls a Ring Three asset of the first rank: the operational context through which internet commerce becomes enforceable.

This is not passive data exhaust from which patterns must be inferred. Money moves through Stripe only when Stripe’s logic permits it. Radar’s fraud rules, spending mandates, KYC determinations, and dispute adjudications apply policy at the moment of transaction, across millions of merchants, linking each charge to an observable and often legally final outcome: charge, fraud outcome, chargeback, resolution, or recovery. The result is fifteen years of transaction-and-outcome history, tested in production against live adversaries and impossible to synthesize from training data. Stripe’s context is not a description of internet commerce. It is the enforced definition of what a legitimate transaction is.

Seen through this lens, Stripe’s acquisitions and product launches over the past two years no longer look eclectic. They resemble the components of a deliberately assembled machine-commerce stack. Metronome, acquired for roughly $1 billion, provides the infrastructure for usage-based billing, while Stripe’s backing of Tempo points toward sub-cent settlement at machine speed. Privy brings 100 million wallets, already connected to Amazon’s agent runtime, and Shared Payment Tokens and the Machine Payments Protocol—an open standard co-authored by Stripe and Tempo—give agents the authority and infrastructure to transact across payment methods.

In an indication of Stripe’s bullishness on the emerging potential for machine-to-machine payments, founder Patrick Collinson tweeted a progress report:

Stripe has begun bringing these pieces together through its AI Gateway, which provides access to multiple models, tracks usage for downstream billing, and is being developed to support real-time machine-to-machine payments. OpenRouter already uses Stripe for payment processing, invoicing, and tax, so the relationship is operational rather than hypothetical. An acquisition would turn that existing foothold into a market-scale routing position, adding the developer demand and cross-provider flow that Stripe could not quickly manufacture on its own. The machine-commerce layer described here is therefore not only an interpretation of Stripe’s assets. It is a product strategy the company has already begun to execute.

Yet connecting routing to payments does not, by itself, produce a coherent machine-commerce system. Settlement can prove that money moved, but not that the purchased inference was any good. Their connection emerges at the level of the individual transaction. Once each all is metered, bound to an authorized wallet, and settled in sub-cent increments, an inference call becomes an economic event that can be priced and cleared on its own.

This points to a new pricing model native to machine demand. Software evolved from seats to subscriptions and then to usage, progressively narrowing the distance between consumption and payment. Machine commerce could close that distance entirely: billing collapses into settlement, and cognition clears at the moment it is consumed. The thirty-day invoice belongs to a period when humans bought software on annual budgets. Machines purchase by the call, requiring the router to discover the price and the ledger to clear it within the same motion. The ledger cannot verify whether the model produced the right answer. It can verify that the agent was authorized to buy it, at an approved price and within a defined mandate, with explicit recourse if something goes wrong. Stripe does not eliminate the uncertainty surrounding inference quality. It makes that uncertainty economically governable by containing it within enforceable spending authority.

This is where the combination becomes more valuable than either component alone. OpenRouter contributes proximity to machine demand, while Stripe supplies the operational context needed to authorize and monetize it. That also explains the direction of the transaction: Stripe possesses the Ring Three context capable of absorbing OpenRouter’s Ring Two position and capturing the value that currently leaks away. The combined entity is something this framework has identified exactly once before, in a very different sector: an emerging orchestrator born holding an incumbent’s context moat.

A startup’s position with an incumbent’s memory

Orchestration positions are always contested as a pincer, the substrate climbing up and cognition climbing down toward the durable middle. The card networks are climbing from the rails with agentic protocols of their own. The labs are descending from the harness with wallets and commerce surfaces. Stripe–OpenRouter sits between them: above the networks, which move money but cannot route cognition. Below the harness, which expresses intent but cannot supply verified settlement, and which, as a player in the game, can never be its referee.

Cursor offers an early demonstration of how that contest may unfold. Its new router, trained on more than 600,000 live requests, uses the context of each coding task to select among models and claims savings of 30–50 percent for early enterprise customers, rising to 60 percent during broader A/B testing. This confirms that live flow can make routing valuable but also reveals the weakness in OpenRouter’s position: a vertical platform such as Cursor sees the task at its point of origin, with richer context than a horizontal router may ever receive.

In the language of market-making, Cursor is internalizing the order flow. OpenRouter may retain the wider tape across models and domains, but it runs the risk of losing the most valuable routing economics to the platforms where demand begins. Its stronger answer, and the logic of the Stripe combination, lies in the machine demand that originates outside any single vertical and in the settlement layer that Cursor does not control.

If intelligence becomes a metered input, who becomes its Visa, its Bloomberg terminal, its control plane? Three franchises, historically three different companies. Stripe’s answer is Stripe: the settlement rail through OpenRouter’s checkout, the tape through its rankings, the control plane through the mandates and metering already assembled around them.

Closing the Loop: Why Payments Are the Second Wedge After Code

A year ago, in a different market, I argued that coding was the wedge into orchestration. The labs’ obsession with code was more than a go-to-market tactic. The Coding Wedge is an entry point for building a harness that could eventually extend into every domain. The evidence since (Claude Code’s share of global public commits, agent teams building compilers, the harness generalizing from code to finance to law) has strengthened that claim.

It has also clarified why coding became the wedge. Code offers the cheapest verifier in the knowledge economy. It compiles, or it doesn’t. The test suite passes, or it fails. That feedback loop made autonomous capability demonstrable rather than asserted, enabled reinforcement learning against verifiable rewards, and allowed the labs to gain credibility in one domain and export the harness into others. The more general principle is that wedges into orchestration emerge where the distance between action and verification is shortest.

Outside code, few systems offer a verifier as powerful as money. A payment clears, or it does not. The ledger reconciles, or it does not. A dispute ends in a resolution that is observable and enforceable. Payments have the shortest distance-to-verifier of any domain outside code, arguably shorter, because finality is defined in law rather than in a test suite. The loop closes here.

The economy contains exactly two maximally verifiable substrates: code and money. The labs took the first and used it to build the harness for work. Stripe is taking the second to build the harness for machine commerce. Coding became the wedge into orchestrating machine work. Settlement could become the wedge into orchestrating machine commerce. The same structural move is unfolding one ring outward. And it could work because Stripe owns the mechanism that makes the outcome final.

The reported valuation is the clearest signal of Stripe’s ambition. If the reports hold, Stripe is considering paying nearly eight times OpenRouter’s valuation from ten weeks earlier, after spending two years and several billion dollars assembling the infrastructure for metering, wallets and settlement.

OpenRouter would connect that stack to the moment an agent selects which intelligence to purchase. Its current income statement cannot explain the price: a 5 percent take on inference spending becomes less attractive as token prices decline. The strategic value lies in the option to connect machine demand directly to an economic transaction, at the level of each individual call.

Whether this particular deal closes matters less than what the reported number has already reclassified. It suggests that the market is beginning, crudely and perhaps prematurely, to price a future in which selecting intelligence and paying for it collapse into a single motion. OpenRouter could determine where machine demand goes. Stripe could authorize the purchase and make it final.

Whoever succeeds in joining those two moments would provide the infrastructure for the machine economy. But more crucially, it would occupy one of the positions through which that economy is governed. That’s a potential combination that provides a durable advantage in the Agentic Era.

DISCLAIMER: The views and opinions expressed here are those of the author alone and are based on publicly available information. They do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Past performance is not indicative of future results. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

This is the latest excerpt from AGNT: The Orchestration Economics Manifesto - An Investment Framework for the Agentic Era. Each Thursday, I explore a major theme of the Manifesto and unpack the frameworks, adding extra context with more recent developments. Note: The figures and sequential references are taken directly from the larger Manifesto that was originally published in April 2026.

Subscribe now

The LLM battlefield has shifted. To understand the new frontline, however, one must understand how the terrain has transformed under the feet of the LLM giants since ChatGPT was first released in November 2022.

For much of the post-ChatGPT era, public attention remained fixed on model capability and generative AI. As 2025 began, agents were barely part of the discourse, let alone agentic AI. The benchmark to measure who was ahead in the generative AI race came down to model power. By the summer of 2025, conventional wisdom held that OpenAI had built a commanding lead in terms of model power, app downloads, compute deals, market share, and mindshare.

Less than one year later, conventional wisdom has been turned on its head. OpenAI is scrambling to catch up to Anthropic’s structural advantages. The models are commoditized table stakes, albeit extraordinarily expensive ones. And the LLM labs are leaving the model layer behind. They are building platform companies. The fight is now for the Orchestration Layer.

The distinction between these two layers is the difference between fragile and durable. Model intelligence answers the question: Can the AI do this task? Orchestration answers the question: Can the AI do this task, here, within this organization’s systems, rules, and workflows, reliably enough that the organization will depend on it?

The evolution of this competition was, of course, influenced by many of the technical and economic factors we have already defined. But the decisive shift that changed the trajectory of this battle and clarified the new agentic dynamic came from somewhere else, somewhere very specific. It came from coding.

Share

The Coding Wedge

Anthropic released an early version of Claude Code in February 2025 and then made it more generally available in May 2025. OpenAI unveiled Codex in May 2025. At first glance, these looked like developer tools. In fact, they marked a turning point in the enterprise adoption of agentic AI. The evidence was visible by mid-2025. In the “State of AI in the Enterprise” report by Menlo Ventures released in July 2025, Anthropic’s share of enterprise budgets had climbed to 32% from 24% since the start of 2024, while OpenAI’s had declined from 34% to 25%. By the end of 2025, Anthropic’s US market coding share had climbed further to 54% of the US coding market compared to 21% for OpenAI.

Again, the popular perception was that OpenAI was a juggernaut that was running away with the LLM prize. But in the trenches, it was Anthropic winning the battles for budgets and developers. As I analyzed this dynamic on behalf of clients to better understand what was driving this adoption, I gave it a name: The Coding Wedge.

Even now, months later, people view the impact of Claude Code with a facile, surface-level framing that is anchored in the previous paradigm. It’s a good product. Anthropic has strong marketing. It went viral. The brand resonated. Perhaps some or all of these are true at the margins. But these fail to recognize the core structural dynamic at work.

Software engineering was the first domain where autonomous agents crossed the threshold from experimentation to reliable production use. It is uniquely suited to this transition because it concentrates three properties that rarely coexist in other enterprise workflows:

Code lives in a highly structured environment. Every task requires sequencing, dependency management, decomposition, and recombination. Writing software trains models on the primitives of orchestration.

Code offers objective ground truth. It either compiles and runs, or it fails. That deterministic feedback creates a fast verification loop unavailable in most other domains. The model can act, test, correct, and improve against clear signals.

Code has unusually high economic leverage. Every productivity gain in software engineering compounds through the products that the team builds, the workflows it enables, and the businesses those products support.

That combination made it the natural entry point for every lab seeking to move from model provider to orchestration platform. The adoption data reinforces this. Cursor, an AI-native code editor, raised funding at a $29.3 billion valuation. Claude Code surpassed $2.5 billion in ARR, with business subscriptions quadrupling since the start of 2026. And 4% of all public commits on GitHub worldwide were authored by Claude Code as of February 2026, with as much as 20% projected by the end of 2026. The coding wedge is not theoretical. Coding is the fastest-growing category in enterprise AI.

Rather than being about code completion, the coding wedge is about controlling the first enterprise domain in which agentic systems became reliably deployable. The deeper strategic point is even more important: coding agents are not merely coding tools. They are general-purpose agent Harnesses disguised as developer products.

That is what the coding wedge really is: the first opening through which orchestration enters the company. The progression was not accidental. It was strategic. Win developers through the coding wedge, then expand to the broader enterprise through the trust, behavioral patterns, and organizational adoption that coding establishes.

However, as noted in Chapter 2, Anthropic had another powerful tool in its arsenal: MCP. Released as open source in November 2024, this protocol rapidly became the industry standard for agent-to-system integration, even adopted by OpenAI and Google. This gave Anthropic a structural edge. The axis of competition had rotated from “which model is best?” to “whose ecosystem is most deeply embedded?”

That rotation could not be reversed by a single model release, however impressive. Anthropic’s coding advantage also became a powerful development lever. First, Claude Code won developers. That became a wedge to win enterprise customers. But it also became a method for learning the architectural strengths and weaknesses of agentic AI to extend it to general knowledge work.

Two Architectures, Two Bets

The new terms of this rivalry were on full display in early February 2026 as demonstrated by the volleys exchanged between the “Great Model Powers” across business systems. Every significant product that OpenAI and Anthropic have shipped in the final months of 2025 and the beginning of 2026 revealed that neither company believes the model layer alone can sustain their soaring valuations.

On February 5, OpenAI announced the latest release of Codex, its coding platform. This represented a critical moment for the company as it sought to regain ground in the critical fight for the coding wedge. OpenAI acknowledged this directly at the launch of Codex, describing coding as the foundation for a much broader class of knowledge-work agents. As part of that effort to extend its value to developers, OpenAI concomitantly launched Frontier, an enterprise platform for deploying and managing AI agents across business systems.

Meanwhile, Anthropic had leveraged Claude Code to create Cowork, and then the vertical plugins that triggered the “SaaSpocalypse.” And then^, a couple of weeks later, it closed a $30 billion funding round at a valuation of $380 billion, more than double its $183 billion valuation from just five months earlier in September 2025, and the second-largest private funding round in technology history. Finally, in late May, Anthropic announced a $65 billion Series H funding round at a $965 billion valuation.

This all points to the real story.

Consider that $965 billion valuation. At $47 billion ARR, that represents a 20.5× multiple. That is notable because after the February funding, the reported $14 billion ARR implied a 27x multiple. So, revenue has reportedly grown fast enough that the headline multiple has compressed.

Still, if Anthropic is a model company selling intelligence in a market where the intelligence is converging, then this is an extraordinarily aggressive bet on a commoditizing asset. Especially in light of the recent release of Kimi K3. But if Anthropic is a platform company, selling orchestration, workflow coordination, and accumulated enterprise context, then 20.5x may be the entry price for a generational enterprise franchise.

Claude Code, Cowork, the vertical plugins, Frontier, and Codex are not model improvements. All of them are orchestration infrastructure: systems designed to coordinate AI agents across enterprise workflows, accumulate institutional context, and create the switching costs that justify platform multiples.

And yet, the two leading AI labs have arrived at the Orchestration Layer with structurally different visions. Both have real products, launched within weeks of each other, that embody different assumptions about how the enterprise AI market will evolve. Understanding the divergence is essential for valuing the companies.

Anthropic pursued the same logic from the other direction: win developers first, then use the trust, habits, and infrastructure built inside engineering teams to expand outward into the rest of the enterprise.

Agents on Frontier receive employee-like identities with scoped permissions. They connect to data warehouses, CRM systems, and internal applications through what OpenAI calls a “semantic layer for the enterprise.” They build institutional memory from their interactions over time.

The most critical thing to know is that they do not need to be OpenAI’s agents. Frontier manages agents from Anthropic and Google. About 10 days later, on February 15, 2026, OpenAI staged a minor coup by announcing the hiring of Steinberger, the man behind OpenClaw, “to drive the next generation of personal agents,” according to CEO Sam Altman.

Frontier is a Control Plane

Agents on Frontier receive employee-like identities with scoped permissions. They connect to data warehouses, CRM systems, and internal applications through what OpenAI calls a “semantic layer for the enterprise.” They build institutional memory from their interactions over time.

The design choices reveal strategic intent. Frontier’s agent identity management mirrors human HR systems: onboarding, scoped permissions, performance monitoring. OpenAI is not being metaphorical when it describes agents as “digital coworkers”. It is making an architectural claim. If AI agents operate as enterprise employees, the platform that manages them becomes as essential as the system that manages human employees. Frontier aspires to become the Workday for AI labor.

The most critical thing to know is that they do not need to be OpenAI’s agents. Frontier manages agents from Anthropic and Google. This seems to be generous. In fact, by welcoming competitors’ agents onto its platform, OpenAI concedes agent-layer competition in exchange for control-plane dominance. If enterprises standardize on Frontier for agent governance, OpenAI captures the value of orchestration regardless of which model powers any individual agent. The agents become interchangeable components. The control plane does not. This is horizontal platform logic. AWS applied to enterprise AI. It does not require OpenAI to have the best model. It requires OpenAI to have the best coordination infrastructure.

Cowork is an Agent that Became a Platform

Rather than building a management layer on top of agents, Anthropic has built outward from the agent itself. The sequence is systematic and each step compounds on the last.

Claude Code proved the model could carry out complex, multi-step enterprise tasks in the critical domain of software engineering. It generated $2.5 billion in revenue68. MCP standardized how agents connect to external systems and was adopted by competitors, including OpenAI and Google, establishing it as an emerging industry standard. Cowork extended orchestration beyond developers to knowledge workers. The vertical plugins created purpose-built entry points into specific business functions.

The strategic logic here is vertical integration. Anthropic is betting that agent quality remains meaningfully differentiated when embedded within orchestration. In that framing, the experience of using Claude to coordinate an intricate legal review or financial analysis is sufficiently superior that enterprises will build their workflows around it. This is closer to Apple’s logic: control the end-to-end experience, build an ecosystem around your product, and make execution quality the moat.

But there is a subtle departure from the Apple analogy. MCP is open-source. Anyone can implement it. This seems to undermine lock-in.

The answer lies in a finding from enterprise AI adoption research: only 11% of enterprise builders switched AI providers, even though technical substitution was trivial. The lock-in is not technical. It is organizational. When an enterprise builds workflows around Claude’s specific orchestration patterns, the cost of switching becomes operational. In this case, those patterns include its approach to multi-step reasoning, its tool-use conventions, and its plugin interfaces.

You can swap the model in an afternoon. You cannot re-tune thousands of enterprise workflows in an afternoon. Anthropic is making a layered bet: MCP creates ecosystem breadth (every system connects), Claude’s quality creates ecosystem depth (Claude is preferred within that ecosystem), and organizational adoption creates inertia (no one switches even when they could). A market phenomenon of behavioral lock-in despite technical substitutability becomes a deliberate business strategy.

The two architectures represent different bets about the speed and completeness of model commoditization. If intelligence commoditizes fully, Frontier wins. The platform that coordinates interchangeable agents captures the greatest economic value. OpenAI wins not because GPT is superior, but because Frontier is the control plane. It wins even if Claude is the better agent, because the control plane sits above the agent layer.

If intelligence retains meaningful differentiation when embedded in orchestration, Cowork wins. Enterprises do not want “any agent, well-managed”. They want the best agent, deeply embedded. And the switching costs compound with every workflow built around it.

What $965 Billion Requires

If orchestration, not intelligence, is what drives the labs’ valuations, three conditions must hold:

Orchestration lock-in must prove durable, not merely behavioral. Enterprises are not switching AI providers today. But this was measured during a period of rapid growth when no one had reason to test the limits of their commitment. The real test arrives when Frontier offers agent-agnostic orchestration at compelling economics, or when a competitor undercuts on price. MCP is open-source. It creates connectivity but not captivity. For $965 billion to be justified, the network effects must become self-reinforcing: more workflows generate richer context, produce better orchestration, and attract more workflows. This flywheel is architecturally plausible. It is not yet empirically confirmed at scale.

The coding wedge must compound. Claude Code’s $2.5 billion ARR validates the entry strategy. But the path from coding to legal, finance, HR, and general operations is the critical progression. If the vertical plugins drive cross-functional adoption, the platform thesis holds. If coding remains the dominant revenue line while other verticals grow incrementally, Anthropic is a remarkable single-function business at enormous scale. That is a different valuation than an enterprise platform.

The labs must accumulate enterprise context faster than incumbents can defend. The systems of record hold decades of accumulated institutional context: the workflow knowledge, compliance history, and operational understanding that orchestration depends upon. The labs are starting from zero. Frontier builds “institutional memory” from agent interactions. That memory is weeks old. The institutional memory embedded in a Fortune 500 company’s Salesforce deployment spans years.

The labs have speed and pliability. The incumbents have depth and irreplaceability. The labs must build enterprise context before model commoditization erodes the intelligence advantage that gives them the right to orchestrate. This is a race against the clock. The clock does not pause for fundraising announcements.

The Enterprise Tax

There is a dimension of this transformation that the market has yet to price: the expense of becoming an enterprise software company. The market narrative portrays AI labs as asset-light technology companies that are disrupting bloated incumbents. The reality is more complex.

To win the Orchestration Layer, the labs must build what every enterprise platform company before them has built: field sales organizations, customer success infrastructure, compliance certifications, vertical domain expertise, and the organizational capacity to manage thousands of enterprise relationships simultaneously.

OpenAI has embedded Forward Deployed Engineers within customer organizations. Frontier’s enterprise customers require SOC 2 Type II, ISO 27001, and a suite of related certifications. The EU AI Act begins enforcement in August 2026, introducing compliance requirements that did not exist when these companies were founded. Enterprise sales cycles sometimes stretch twelve months or longer for major deployments. These demand specialized teams that change margin assumptions for the labs.

The funding Anthropic raised in February 2026 was not exclusively for training runs. A meaningful portion of this capital will fund the construction of an enterprise go-to-market apparatus, including a sales force, compliance infrastructure, customer success organization, and vertical expertise. These are structural costs that permanently alter the business’s margin profile. A model company has research costs and compute costs. A platform company has all of these, plus go-to-market costs that scale with customer count rather than compute capacity.

This creates a tension that the labs’ financial disclosures will eventually make visible. The transition from model economics to platform economics requires near-term margin compression to build the switching costs and network effects that expand margins over time. The trajectory of that compression and subsequent expansion will tell investors more about the durability of these businesses than any benchmark or revenue growth rate.

Meanwhile, the labs face a clock that the incumbents do not. Model advantages erode quarterly as open source narrows the gap. If the labs do not establish orchestration lock-in before model commoditization is functionally complete, they risk becoming very expensive API utilities competing on price.

What This Means for the Public Markets

When the IPO prospectuses for OpenAI and Anthropic become public, they will present revenue growth, gross margins, and customer counts in the language familiar to technology investors. That language will be necessary but insufficient.

If this analysis is correct, if these companies are transitioning from model businesses to platform businesses, then the central valuation question is not the rate of growth but the nature of the growth. Revenue from model API consumption and revenue from platform orchestration may appear identical in a financial statement. They are not identical in their implications for margin trajectory, competitive durability, or terminal value.

The companies that complete this transition will justify platform multiples. Those that do not will eventually be priced as API providers in a commoditizing market, regardless of current growth.

The Single Points of Failure

The first is open-source compression. For now, enterprises overwhelmingly prefer closed-source frontier models today. The MAP study found that 85% of production agent teams build entirely in-house using direct model API calls, foregoing third-party orchestration frameworks. Open-source adoption is limited to edge cases: high-volume workloads where inference costs are prohibitive, or regulated environments that prohibit sending data to external providers.

Enterprise teams default to the best-performing closed-source model available, and runtime costs are negligible compared to the human experts the agents augment. This is encouraging for the labs. But it is a snapshot, not a guarantee.

Open-source models have continued to advance at the rate they had at the end of 2025 and the start of 2026. And as we saw this month with Kimi K3 from Moonshot AI, the performance gap has been closed. Is this parity sufficient for the 85% preference to erode rapidly, because the underlying driver is pragmatic performance selection, not structural loyalty? Enterprises test the top models and pick the best one. If an open-source model becomes the best, will they pick it instead?

Of course, this raises a second SPOF: Jurisdictional and continuity risk. The June export ban on Claude Fable 5 and Mythos 5 suggests that the regulatory picture is set to play an important role. And yet, the rules are far from clear. This has grown even fuzzier in the wake of the panic and debate following the release of Kimi K3, and the ensuing accusations by Anthropic that it was distilled from Fable, and rumblings from the U.S. government that it may look to restrict the use of some open-source models.

The third is the compute trap. This is the risk that Dario Amodei himself has articulated with remarkable candor. Asked on a podcast why Anthropic does not spend more aggressively on compute given its belief that a “country of geniuses in a data center” is imminent, Amodei was honest about the financial fragility of the model: “If my revenue is not $1 trillion, if it’s even $800 billion, there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt if I buy that much compute.”

By the same measure, if Anthropic underspends, it misses potentially massive growth opportunities. This is an unprecedented situation in enterprise technology. No company in the history of the software industry has operated at this level of capital intensity with this degree of revenue uncertainty over such a compressed timeline.

Anthropic’s annual burn rate in early 2026 was running at approximately $7-8 billion (one-third of revenue, per company guidance disclosed in confidential financials reported by The Wall Street Journal and Fortune in November 2025). The $65 billion Series H provides fresh runway. But the runway at these burn rates is measured in years, not decades. The compute commitments required to maintain frontier model performance only increase. Anthropic has committed approximately $80 billion in cloud-infrastructure spending to Amazon, Google, and Microsoft through 2029, with an additional 3.5 GW of compute capacity secured through Broadcom and Google partnerships beginning in 2027.

Of course, Anthropic then entered a compute agreement with SpaceX that allowed it to increase Claude Code and API limits. And yet, SpaceX, as part of its post-IPO plans, has moved to formally acquire AI-coding platform Cursor, a move that is effectively an attempt to re-run Anthropic’s Coding Wedge playbook.

The tension is structural. To win the Orchestration Layer, the labs need frontier-quality models. You cannot orchestrate enterprise workflows with a mediocre model. To maintain frontier models, they need massive and growing compute investments. To fund those investments, they need revenue growth to materialize on schedule.

If the orchestration transition takes longer than the compute commitments allow, if enterprise adoption moves at enterprise speed rather than startup speed, the financial model fractures. Intelligence is commoditizing. Context compounds. Compute burns. The race against all three will determine if the labs can move up the stack fast enough, and effectively enough, to capture enough of the enterprise to justify their precarious economic positions.

The views and opinions expressed in this publication are those of the author alone and are based on publicly available information. The expressed views and opinions do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Certain companies referenced may be current or former clients of, or counterparties to, the author or affiliated entities; such relationships will be disclosed where applicable. Past performance is not indicative of future results. To the fullest extent permitted by applicable law, the author does not accept any liability for any loss or damage arising from reliance on this content. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

TLDR: K3, unveiled on Thursday, is the largest open-weight model ever announced: 2.8 trillion parameters, with the weights themselves pledged for July 27. The model now ranks No. 1 on the Frontend Code Arena coding leaderboard, above Anthropic’s most powerful model, Claude Fable 5. Chinese labs have claimed parity all year, and the claims kept dying on standardized harnesses. This one did not. The consequences run through every layer of the stack: the capability premium has retreated to the top two models; everything beneath them now competes with a self-hostable peer at task-cost parity; and, most counterintuitively, frontier-scale open weights re-centralize inference into the datacenter rather than dispersing it to the edge. Within forty-eight hours of launch, demand had outrun Moonshot’s own compute.

Subscribe now

With the release of Kimi K2 Thinking last November, I asked whether the industry was getting the fundamental economics of compute and infrastructure costs all wrong.

The model had matched GPT-5 on key reasoning benchmarks at a reported $4.6 million in training costs. Yet unlike the DeepSeek moment of early 2025 that sent markets temporarily into a tailspin, almost nobody seemed to notice the implications of Kimi K2 at that time. I called it a “neutron bomb” that failed to detonate.

The detonation has now happened.

On July 16, Moonshot AI released Kimi K3, a 2.8-trillion-parameter sparse mixture-of-experts model that became the first open-source LLM to top any of the frontier benchmarks, even topping Anthropic’s Fable 5 in Arena’s Frontend Code rankings. The previous day, Thinking Machines shipped Inkling, Mira Murati’s first model: 975 billion parameters, open-weight under Apache 2.0. And on Sunday, Alibaba disclosed that Qwen3.8, with 2.4 trillion parameters, is now available in preview only and open-weight soon.

Unlike last year, Kimi K3 has provoked a dialogue over the past few days that frames this moment as almost existential, igniting fierce debates on some of the most central ideas at the heart of the current wave of AI transformation: East vs West; Governments vs Markets; Open vs Closed; Immigration vs Sovereignty. There are already rumors that the White House may move to curb access to open Chinese models, even as talk continues that it might also demand approval of any frontier models. And at a time when China is calling for openness!

With so much to unpack about this moment, I want to focus on the aspect of K3 that transcends these debates and may have the biggest impact going forward.

Despite the obvious signs, investors and enterprises continued to assume that meaningful differentiation existed across a broad tier of proprietary foundation models. They have invested historic amounts of equity and Capex around that thesis.

K3, Thinking Machines, and Qwen challenge that assumption. Once a frontier-adjacent model survives independent verification and becomes freely self-hostable, scarcity no longer resides in intelligence alone. It retreats to the handful of models that remain genuinely ahead. Everything below that level begins competing with an open substitute whose economics improve every time another cloud provider chooses to host it. Indeed, in just a few days, Moonshot had to limit access to the new Kimi due to overwhelming demand.

This shift will force a repricing across the entire AI stack. The immediate question is no longer whether open models can match the frontier. It is where value migrates once they nearly can.

The Orchestration Economics Manifesto maps this shift as geology, tracking six tremors between September 2024 and early 2026, each a phase shift in one layer of the stack that amplified the others, all moving along a single fault line. The shift from tools to goal-seeking agents. Intelligence and the Harness (routing, memory, orchestration code) are commoditizing fast and becoming self-improving, but real scarcity, which defines durable value and moats, migrates to the two true bottlenecks: scarce compute and operational context (the real-world data, outcomes, and judgment that labs can’t automatically generate or own). Last month, Anthropic’s disclosure of recursive self-improvement supplied the seventh tremor that showed where scarcity migrates once models start building models.

K3 is the eighth tremor: the moment the capability premium compresses to the top two frontier labs, while margin shifts outward toward inference infrastructure, orchestration, enterprise context, and the operational systems that turn intelligence into reliable outcomes. That is why K3 is not simply another model release. This is the point at which the competitive landscape, the economics of AI, and the location of durable advantage all begin to change simultaneously.

Share

What Moonshot Actually Shipped: Kimi K3’s Architecture, Benchmarks, and Price

Let’s start with the technical aspects of Kimi K3 and how it differs from K2.6.

K3 is a 2.8-trillion-parameter sparse mixture-of-experts model, which includes 16 of 896 experts active per token. It has native vision and a one-million-token context window. And it also boasts a new attention mechanism Moonshot calls Kimi Delta Attention, designed to keep decoding fast at million-token depths. Two variants ship: K3 Max and K3 Swarm Max. The API went live on July 16, OpenAI-SDK-compatible and available through OpenRouter. The company has pledged publicly to release the weights by July 27, reportedly under a Modified MIT license, though Moonshot had not confirmed the terms at the time of writing.

Now the benchmarks.

On Artificial Analysis’s Intelligence Index, K3 scores 57: 4^th of 189 models, behind Fable 5 at 59.9 and two configurations of GPT-5.6 Sol, and ahead of Opus 4.8. On LMArena’s Frontend Code Arena, it debuted first with a 1679 Elo, finishing ahead of Fable 5 and taking six of seven categories. The caveats are real: fewer than two thousand votes, preliminary results, and leaderboards that can shift quickly. Even so, no open-pledged model has ever debuted above the closed frontier in blind human evaluations of working code.

The head-to-head record is more revealing than the composite. K3 beats both Fable 5 and GPT-5.6 Sol outright on SWE Marathon (42.0 against 35.0 and 39.0), on BrowseComp (91.2 against 88.0 and 90.4), on OmniDocBench, and takes first place on AutomationBench, the agentic SaaS-workflow evaluation. One detail buried in the BrowseComp run matters more than the score: it was executed at the full one-million-token window, with no context compaction. I will return to why that matters.

Where does K3 lose? Where winning is hardest. It loses the overall index. It loses GDPval, the evaluation closest to economically valuable knowledge work, where its 1668 Elo beats Opus 4.8 but sits well behind Fable 5’s 1760. It loses FrontierSWE, 81.2 against 86.6. It loses DeepSWE Datacurve, scoring 67.5 against Fable’s 70 and Sol’s 73. This contamination-resistant coding benchmark is built from held-out tasks precisely because public GitHub benchmarks leak their answers into training data.

Yet even that defeat reinforces the argument. When DeepSWE launched in late May, it revealed a sixteen-point moat between GPT-5.5 and the field below it. By mid-July, that gap had narrowed to roughly five points, with an open-pledged model now inside the frontier band.

Reliability remains the more consequential weakness. Although K3 improved on K2.6 in accuracy, its measured hallucination rate rose from 39% to 51%. That’s still slightly below Grok 4.5’s 54%, but far too high for many enterprise use cases. Moonshot itself concedes a user-experience gap against Fable and Sol, along with what it calls “excessive proactiveness”: a tendency for the model to do more than the user asked. Then there is the price. This is the part of the K3 story I believe is being most widely misunderstood.

At $3 per million input tokens and $15 per million output tokens, K3 is priced at the Sonnet tier, with an output rate nearly four times that of K2.6. It is the most expensive model any Chinese lab has ever shipped. The “fraction of the cost” framing that has followed every Chinese release since DeepSeek simply does not apply. The reality is more nuanced.

On Artificial Analysis’s cost-per-task measure, K3 completes a benchmark task for $0.95. That is essentially GPT-5.6 Sol’s $1.04, and roughly half the $1.80 cost of Opus 4.8. Parity at the task level against the flagship tier, driven by improving token discipline: K3 used 21% fewer output tokens than K2.6 across the full benchmark run while scoring thirteen points higher.

But improved is not the same as efficient. K3 generated roughly 130 million tokens across the Intelligence Index evaluation, nearly twice the median verbosity of the models Artificial Analysis tracks. That is why OpenAI’s efficiency tiers, GPT-5.6 Terra and especially Luna, still beat it on cost per task even as the flagship Sol does not. K3 matches the frontier tier on task economics and loses to the tiers built for volume. A verbose model with a low sticker price is a different animal from an efficient one, and anyone modeling K3 for high-volume production should price the verbosity.

The cheap story, in other words, lies not in K3’s current API pricing but in what comes next. It begins when the weights are released, and third-party hosts start competing away the margin on inference.

Moonshot also claims that K3 delivers roughly 2.5 times the scaling efficiency of K2, which translates into more capability per unit of training compute. The technical report is forthcoming. Until it lands and replicates, that is a vendor claim, and I will treat it as one. Even so, it was likely this claim, more than the benchmark rankings themselves, that unsettled markets when the model was released.

Why Kimi K3 Is a Tremor, Not a Headline

A tremor, in this framework, is a phase shift in one layer that amplifies the others. K3 sits on three of the original axes at once:

Accessibility crosses a regime boundary. The framework I laid out after the Fable shutdown holds that the frontier creates capability, diffusion spreads it, and orchestration captures it. Epoch measures the open-weight lag at three to four months behind the closed frontier. DeepSeek’s moment in January 2025 was the second tremor. It established near-frontier capability, cheap, one generation behind: last year’s intelligence at a twentieth of the price. When I wrote about K2 Thinking in November, the gap had narrowed to months. K3 compresses the lag toward zero for everything below the top two: frontier-adjacent, verified, open, self-hostable within days. The commoditization line has climbed from “yesterday’s intelligence, discounted” to “this cycle’s intelligence, one notch down, yours to run.”

Swarm coordination moves into the weights. K3 Swarm Max coordinates an orchestrator and up to roughly 300 sub-agents across some 4,000 steps. The capability first shipped with K2.6 in April. What matters is where it now lives. Not in a harness bolted around the model but trained into the model itself. The Claude Code leak laid the stakes bare. The harness includes routing, memory, delegation, and policy. This is where a thousand companies have pitched their moats, but it is now much more vulnerable.

The million-token agent becomes real. Long context has been advertised for over a year and trusted by almost no one who runs agents in production. That’s because of a phenomenon known as “context entropy.” This describes the degradation that caps most production agents at ten steps or fewer. It worsens with window size rather than improving. K3’s BrowseComp result, beating both frontier models at the full million-token window without compaction, is the first independently run public evidence that the million-token agent is a working tool rather than a specification.

Now the counterweight. Fable 5 and GPT-5.6 Sol still win the overall index, the hardest software-engineering evaluations, and the evaluations closest to real economic work. A 51% hallucination rate is disqualifying for large classes of enterprise deployment. The closed frontier has not yet fallen.

Clearly, with the Kimi K3 release, the capability premium has retreated to the top two models. Everything below that tier, Opus-class capability included, now competes with an imminently self-hostable peer at task-cost parity.

The Geography Flipped: From Llama to Kimi, Qwen, and GLM in Eighteen Months

K3 did not arrive alone. It arrived as the crest of a wave: DeepSeek R1 in January 2025; Kimi K2, the first trillion-parameter open release, in July 2025; K2 Thinking in November; the Qwen line; GLM-5.2 in June, MIT-licensed at roughly a sixth of Western frontier pricing; MiniMax M3 at a twentieth; DeepSeek V4.

There is an important point to consider in this string of releases. All year, open-weight “parity” was vendor-reported, and all year the claims took a documented seventeen-to-twenty-one-point haircut when re-run on standardized independent harnesses. GLM-5.2 came closest. When Zhipu shipped it in June, I called it the first open model to break into the closed frontier cluster. It scored 51 on the same index where K3 now scores 57. Yet I still counseled caution, because Epoch’s independent evaluation was pending and open models flatter public benchmarks. K3 is the first release to clear the independent bar outright, at a higher tier, on the strictest harness available. Qwen 3.8-Max, three days later, is a reassertion of the pattern: a 2.4-trillion-parameter frontier claim with no independent run behind it yet.

The lineage carries a family trait. GLM-5.2 was the least token-efficient open model in its class, spending 43,000 tokens where MiniMax spent 24,000, effectively buying its capability with verbosity. K3 inherits the habit, but one tier higher. The open-weight wave keeps reaching the frontier by outspending it on tokens, which is exactly why the efficiency frontier, not the capability frontier, is where the next battle sits.

Meta, which carried the US open-weight banner for three years, stepped back from the Llama line this spring. Now its Superintelligence Labs flagship model is proprietary. The largest American platforms no longer field an open-weight strategy at all.

That has left a vacuum that Thinking Machines is hoping to fill.

Co-founded by Mira Murati, former CTO of OpenAI, the startup last week released Inkling. It represents a very different bet from anything Silicon Valley ever shipped: 975 billion parameters, Apache 2.0, paired with the startup’s Tinker fine-tuning platform, and positioned explicitly as not the strongest model available, but rather the best foundation for building your own.

Consider what Murati is wagering. She is betting against one-size-fits-all intelligence. Instead, she is wagering that the layer where durable value accumulates will be built around customization, context, and orchestration. The executive who helped build the closed system is now building the exit ramp.

The market is pricing the bet, not the benchmark. Thinking Machines raised $2 billion at a $12 billion pre-product valuation, with reported talks at $50-60 billion for a model sixteen points off the frontier. Inkling debuts at 41 on the same independent index where K3 scores 57. That means the best American open model is sixteen points behind the Chinese open frontier it answers. Its mixture-of-experts design, by Thinking Machines’ own account, largely follows DeepSeek-V3’s recipe, and its post-training was bootstrapped on synthetic data generated by Kimi K2.5. So, the American open-weight standard-bearer is built on Chinese architecture and finished on Chinese data, a dependency Thinking Machines says its next generation will shed.

Where Inkling does lead is revealing: 25,000 output tokens per task against GLM-5.2’s 43,000 and K2.6’s 38,000.

Put the pieces together, and the map has redrawn itself. In eighteen months, open source’s center of gravity moved from Menlo Park to Beijing and Hangzhou, and the American response is now carried by startups and sovereignty vendors, not platforms.

The Reversal: Why Kimi K3’s Open Weights Re-Centralize the Datacenter

At first glance, K3 appears to threaten the largest capital commitments in economic history. On closer inspection, it does something more interesting.

If Moonshot’s scaling-efficiency claim holds, frontier-adjacent capability is becoming cheaper to train per unit of intelligence. That does not invalidate the trillion-dollar capex thesis, but it does introduce a discount factor into its simplest assumption: that more compute is the only route to more capability. One datapoint does not overturn a scaling law. Call it a crack in the premise, not a break.

Now ask the more important question: where does a model like this actually run?

The open-weight wave was supposed to push inference toward the edge: smaller, distilled models running on laptops and phones, intelligence dissolving into devices. K3 reverses that vector. A 2.8-trillion-parameter mixture-of-experts model whose weights alone occupy roughly 1.4 terabytes, even under four-bit quantization, is not an edge model. It runs on datacenter accelerators with enormous memory requirements, regardless of who hosts it or where.

To “self-host” a K3-class model is therefore to buy or rent substantial accelerator and HBM capacity—inside a neocloud, a sovereign cloud, or an enterprise facility, but inside a datacenter all the same. Frontier-adjacent open weights do not disperse inference to the edge. They re-centralize it inside the data center.

The net effect on compute demand, I suspect, is therefore positive, but compositionally different. Demand migrates away from closed-lab training Capex and toward distributed inference hosting. It is also unusually memory-intensive demand: expert weights and context caches, multiplied across large numbers of concurrent agents, create a memory problem as much as a FLOPs problem.

Jevons does the rest. At no point in this discontinuity has cheaper capability per token produced fewer tokens.

When Cerebras filed to go public, I argued that the Inference Economy—an economy in which inference, rather than training, becomes the center of gravity—was the layer against which AI infrastructure should be valued. K3 delivers a demand shock directly into that layer.

And then there is memory. The reflex trade is easy to imagine. Kimi Delta Attention reduces the KV cache by roughly fourfold relative to standard attention, according to Moonshot. A smaller cache appears to mean less HBM per deployment, which in turn appears bearish for memory manufacturers.

That interpretation gets the causality backward. Cache efficiency is not a demand destroyer. It is a demand detonator.

The binding constraint on agentic AI has not been intelligence itself. It has been the cost of sustaining intelligence across long-context tasks. Million-token agents have existed on specification sheets for more than a year, but few organizations have run them in production because the cache economics made them prohibitive.

Kimi Delta Attention relaxes that constraint. And that relaxation is precisely what makes K3’s headline results possible: the BrowseComp run across the full million-token context window, and the 300-sub-agent swarm operating across 4,000 steps.

Cheaper long-context tokens do not mean fewer tokens. They mean elastically more of them. HBM freed at the level of the individual agent is redeployed into larger batches, longer contexts, and more concurrent agents—not smaller infrastructure bills.

This is the DeepSeek lesson replayed almost note for note. The efficiency panic of January 2025 was followed not by collapsing compute demand, but by record demand. The analysts who treated efficiency as a substitute for infrastructure spent the rest of the year discovering that it was an accelerant.

Ahead of the OpenAI and Anthropic IPOs

Which brings us to OpenAI and Anthropic, both moving toward the public markets, both about to price a decade of assumptions in a single window.

In the Red Queen’s Race, I argued that GLM-5.2 and Sakana’s Fugu had shut the labs’ two escape doors. When the model became cheap, the answer was the harness. When the harness was copied, the answer was the model. In a race with no permanent winner, the value accrues to whoever can escape the race altogether. K3 does not open a third door. It accelerates the treadmill, a verified frontier-adjacent model, weights days away, with the harness trained into the weights themselves. The Red Queen’s advice was to run twice as fast. The IPOs will price how fast is fast enough.

Let me be clear about what I am not arguing. This is not “open source kills the labs.” The labs’ defense is real, and I have spent much of this year documenting it. They own the top-two capability tiers outright. They own the reliability gap: GDPval, the hallucination numbers, and Moonshot’s own concessions. They own enterprise trust, compliance, and distribution. And, per the seventh tremor, they own the recursive-self-improvement flywheel. If models build models, the compounding asset is the lab’s internal loop, not its price list. There is even a coherent case that commoditization one notch below the frontier concentrates value at the true frontier rather than destroying it.

Even so, the pressure still lands in two places. The first is the middle of the pricing ladder. The premium has retreated to Fable 5 and GPT-5.6 Sol. Opus-class and mid-tier API pricing must now compete with a self-hostable peer at something close to task-cost parity. Compression reaches the mid-tier first, but an IPO must underwrite the economics of the whole ladder, not just the summit. The second is demand-side script. Enterprise CFOs now possess a public vocabulary for negotiating down or walking away: tokenmaxxing, sovereignty, and model portability. etc.

What remains of the American open-weight ecosystem: one startup and a handful of sovereignty vendors? If the second-best model in the world is Chinese and free to download, what exactly does restricting access to the best one accomplish, beyond taxing the diffusion of your own technology? Beneath all of this sits a policy paradox. Export controls on the closed frontier accelerate substitution toward exactly the open Chinese artifacts controls cannot reach. What remains of the US open-weight field? Potentially just one startup and the sovereignty vendors. If the second-best model in the world is Chinese and free to download, what does a control on the best one accomplish, beyond taxing the diffusion of your own technology?

The race is real. The US is currently running it with one open hand tied behind its back.

Where the Value Goes: Intelligence, Harness, Orchestration

Finally, let us analyze Kimi K3’s impact against the Three Rings of the Agentic Enterprise, a cornerstone of Orchestration Economics.

The agentic firm is organized in concentric rings, not layers, because what matters is not just function, but proximity to value capture. The further out the ring where a firm sits, the greater the control over outcomes, and the stronger the economic position:

Ring One: Intelligence. Foundation models that provide the cognitive capability.

Ring Two: Harness. This is the orchestration infrastructure that operationalizes intelligence by decomposing goals into tasks, delegating them to specialist agents, managing state, coordinating execution, and accumulating cross-system understanding with every session.

Ring Three: Orchestration. The entity that reorganizes itself with intelligence and the Harness at its core and provides outcomes from operational context that neither layer can produce holds the position.

Kimi K3 commoditizes Ring 1, raw intelligence, at the frontier-adjacent tier. By shipping swarm coordination inside the weights, it begins commoditizing Ring 2, the harness. Each tremor pushes scarcity one ring outward.

Ring 3 is where the battle remains. To capture the most valuable terrain in the Agentic Era, a company must also possess irreplaceable operational context, proximity to intent, and workflow history. These are the Three Laws of Agentic Value that allow us to analyze a company's durable position in terms of its accumulated understanding of how work actually gets done. These are not things that can be downloaded from Hugging Face.

When orchestration behavior ships inside open weights, the standalone economics of that layer grow thinner, and value is pushed one ring outward, toward context and intent: the territory of the Three Laws as I lay out in the AGNT Manifesto.

The Inference Economy accelerates on every margin. Cheaper, self-hostable, swarm-capable intelligence means more agents, more tokens, more coordination, and value capture migrating to whoever orchestrates, verifies, and owns outcomes.

Step back far enough, and the eight tremors resolve into a single motion. Six laid the foundation of the Agentic Era: intelligence cheap, ubiquitous, coordinated at machine scale. The seventh showed scarcity migrating to the self-improving loop. The eighth shows margin migrating outward from the model layer. One fault line, one direction of travel.

The discontinuity does not care who wins the model race. It has repriced the position of models in the value stack.

As always, I close with the falsifiers. These are three dated things that would weaken this analysis:

1. July 27 passes without weights from Moonshot AI, or with a license too restrictive for commercial self-hosting.

2. The released weights fail independent replication, causing the vendor-haircut pattern to recur one more time, this time post-release; the technical report fails to substantiate the claimed 2.5-times scaling efficiency.

3. Serving capacity proves binding. If, once the weights ship, third-party hosts cannot stand up K3-class inference at scale, the demand shock stays theoretical.

If those fire, K3 was the ninth vendor claim, not the eighth tremor.

Of course, with this success, Kimi and Moonshot AI will now face new levels of pressure and scrutiny. The weekend provided the first counter narrative: within 48 hours of launch, demand for K3 overwhelmed Moonshot’s own GPU capacity, and the company paused new subscriptions outright, promising to re-admit users in controlled batches.

That’s good news in terms of signaling the demand. But it also points to the potential compute crunch the company faces if it wants to scale and meet this demand. As Anthropic found when growth outstripped projections this year, rationing access to frontier models can lead to its own backlash among users. And with rumors now that Moonshot is targeting its own IPO, the company will soon learn how expectations change as it moves from scrappy underdog to frontrunner.

Still, set those caveats against the larger token efficiency panic. A model engineered to make long-context inference cheap sold out its maker’s datacenters in two days. Cheaper tokens did not mean fewer tokens. They produced a queue.

Moonshot’s stated relief valve is the July 27 release itself: open weights that let the demand it cannot serve spill onto everyone else’s accelerators.

And so, the eighth tremor closes where it opened. Open weights do not eliminate or decentralize the datacenter. They multiply demand for datacenters.

DISCLAIMER: The views and opinions expressed here are those of the author alone and are based on publicly available information. They do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Past performance is not indicative of future results. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

00:00 - Intro

00:01 - Podcast Insights and Market Developments: CoreWeave and Anysphere

04:38 - CoreWeave: Strategic Positioning and Anthropic’s Impact

08:59 - Enterprise AI Cost and Impact

12:03 - The Shift to Sovereign AI

14:37 - Agentic Memory and Emerging Tech Opportunities

20:47 - Emerging Frontiers of Memory

This is the latest excerpt from AGNT: The Orchestration Economics Manifesto - An Investment Framework for the Agentic Era. Each Thursday, I explore a major theme of the Manifesto and unpack the frameworks, adding extra context with more recent developments. Note: The figures and sequential references are taken directly from the larger Manifesto.

Subscribe now

The Three Laws describe the conditions under which a company can become the orchestrator and sustain a durable advantage in the Agentic Era. But markets do not price conditions in the abstract. They price the structural positions that those conditions give rise to. When proximity to intent, depth of context, and control over workflow converge within the same domain, they do not remain as separate attributes. They collapse into a single, identifiable role in the system that transforms human intent into executed outcomes. That role is not captured by any existing category. It is not a product definition, a sector classification, or a stage in the software stack as previously understood. It is a position that emerges from the architecture itself.

The purpose of this chapter is to describe the position and define it clearly. Because once it is visible, it can be systematically identified across the market, compared across industries, and evaluated as the primary locus of value creation and capture in the new regime.

Share

The Species

What emerges from that shared position is best understood not as a sector but as a species: a class of firms defined by their structural role in the orchestration of work.

A company that captures intent at the point of origin, that owns the deepest and most operationally relevant context, and that directs the workflows required to convert both into outcomes occupies a position that is categorically different from one that performs only a subset of those functions.

This manifesto calls that position AGNT.

AGNT is not a stock ticker. It is a structural archetype that can be identified, scored, and compared across industries. It cuts through traditional boundaries because those boundaries were built for a different regime, one where software was a tool, and operations were separate from it. The market’s sorting mechanisms via sector-based indices, thematic baskets, and the Rule of 40 were calibrated for a regime in which software competed for IT budgets and physical businesses competed on operational efficiency. GICS codes describe where a company was. The Three Laws of Agentic Value describe where a company is going.

An enterprise CRM platform can be an AGNT. It captures intent because 150,000 companies set sales goals in it every day. It provides context through decades of customer interaction data, pipeline patterns, and conversion histories across millions of accounts. It compounds workflow intelligence with every sales cycle it orchestrates. Apply the Three Laws. The position is defensible.

A global insurance carrier can be an AGNT. It captures intent because the policyholder is its customer. It holds context in the form of actuarial loss data built over decades of underwriting real risk with real capital. No foundation model can synthesize this from training data. It was generated by operations that bear real-world consequences.

In the Agentic Era, that separation collapses. A freight carrier and an enterprise software platform may have more in common architecturally than either has with the company next to it in a sector index. The entity that receives intent, deploys intelligence, and coordinates execution becomes the locus of control, regardless of whether it is labeled as a software company, a financial institution, or an industrial operator.

The implication is that similarity is no longer determined by sector adjacency, but by architectural equivalence. Two companies in different industries may occupy the same structural position, while two companies in the same industry may be separated by it entirely.

The Three Laws provide the mechanism for making that distinction explicit. Apply them, and the species reveals itself.

The question is not what this company sells. The question is where this company fits within the “Orchestration Graph” we define below.

Four Combatants

The first combatant: the AI laboratories ascending the stack.

OpenAI and Anthropic did not begin as orchestrators. They began as model providers selling intelligence at the bottom of the stack. But they understand that the model layer alone cannot capture the value implied by their valuations. Model price-performance is compressing, even as access to the most capable frontier intelligence remains scarce and increasingly subject to strategic control.

So the laboratories are building upward. OpenAI has expanded from Frontier into workspace agents, ChatGPT Work, and a dedicated deployment company designed to embed its systems directly into enterprise operations. Anthropic has moved from Claude Code and Cowork into vertical agent packages.

The entry strategy is what we call the “coding wedge”, documented in software engineering, now understood as a deliberate go-to-market: win developers first, then use the trust, habits, and organizational adoption built inside engineering teams to expand outward into the rest of the enterprise.

Three traditional software defenses are breaking under pressure. Integration moats erode when agents route dynamically across any system with a protocol connection. UI moats erode when the primary user is an agent executing via APIs, rather than a human navigating an interface. Capability moats erode when foundation models compose specialized functions on the fly.

The labs’ advantage is technological fluency and speed of iteration. Their vulnerability is the Second Law. They possess intelligence without institutional knowledge. They have no claims data, no routing history, no actuarial tables, no supply chain memory. They are climbing toward the Orchestration Layer from below, and every rung they reach is contested by incumbents who have occupied it for decades. The race is intelligence ascending against context defending. Chapter 13 traces this in detail.

The second combatant: the software incumbents pivoting from storage to action.

Salesforce, ServiceNow, Workday, SAP, and Oracle hold the canonical data that enterprises operate on. Ripping them out requires migrating decades of operational records, retraining thousands of users, rebuilding integrations with dozens of downstream systems, and accepting the risk of operational disruption in mission-critical processes.

Their position is not in question. Their role is.

In the pre-agentic world, the system of record was also the system of action. When a sales rep wanted to update a deal, she opened Salesforce. The application was the interface where intent met execution. In the agentic world, intent starts in a conversation. A user tells an agent: “Update the pipeline forecast for Q2”. The agent calls the CRM’s API, reads the data, performs the analysis, and reports back without the user ever opening the application. The system of record is still essential. But it risks demotion from the system where work is directed to the system where data is stored. The database behind the orchestrator rather than the orchestrator itself.

The central strategic challenge for every enterprise software incumbent is to make the transition from system of record to system of action, the platform that receives intent, deploys agents, and coordinates workflows end to end. Salesforce is attempting it with Agentforce and the repositioning of Slack as the enterprise’s intent surface. ServiceNow is building agentic IT service management. Those who succeed become orchestrators. Those that do not will retain data gravity but lose control of the workflow and the pricing power that accompanies it.

Their advantage is context gravity. This is the irreplaceable institutional data that already lives inside their platforms. Their vulnerability is speed. The AI labs are building toward the same position from the opposite direction, and the orchestration position compounds with every interaction. The window narrows with each quarter. Chapter 14 traces the sorting.

The third combatant: the throughput node.

Not every software company must become the captain. Some of the most durable positions in the agentic economy will belong to companies that never aspire to the apex of the orchestration graph, but whose context is so dense that every orchestrator must route through them.

A throughput node does not capture intent. It does not control the workflow. It holds the context that makes the workflow correct. When that context is sufficiently deep and semantically rich, the orchestrator cannot complete the work without passing through the node. It doesn’t matter who occupies the captain’s seat. These are the tools that step into the orchestration loop rather than waiting on the sidelines. They expose their context through open protocols, becoming surfaces through which agents move work rather than destinations where humans perform it. A tool that remains a destination risks being bypassed. A tool that becomes a node becomes structural.

The distinction between a throughput node and a convenience layer is semantic density across the production chain. A design platform whose component libraries, interaction models, and brand systems are inputs to every downstream workflow in product development is a throughput node. A project management tool whose task metadata can be replicated by the orchestrator’s own coordination logic is a convenience layer. The Second Law identifies the difference. The market has not yet learned to price it. Chapter 15 traces the distinction.

The fourth combatant: the domain operators building from the inside.

Insurance carriers, banks, logistics companies, manufacturers, and healthcare systems are the companies that own the physical assets, regulatory relationships, and operational context that most deeply satisfy the Second Law.

These companies never considered themselves technology companies. They are becoming orchestrators by necessity, because the agentic transition presents a binary choice: build your own Orchestration Layer or watch someone else build it on top of your operations and capture the margin.

A global logistics carrier deploying a master agent to coordinate fleet routing, capacity allocation, customs clearance, and last-mile delivery is not buying orchestration from a software vendor. It is building orchestration atop its own operational context, using foundation models as reasoning engines.

The carrier’s competitive advantage was always its physical network and accumulated operational knowledge. In the pre-Agentic Era, that advantage was constrained by the throughput of human managers. In the Agentic Era, the same advantage is unleashed via agents coordinated to operate at machine speed across the full scope of operations.

The same logic applies across every Capex-heavy, operationally complex industry:

A construction equipment manufacturer whose agents coordinate fleet scheduling, predictive maintenance, and supply chain management across dozens of job sites.

An insurer whose agents process claims through orchestrated workflows built on decades of actuarial data.

A bank whose agents orchestrate credit analysis, risk assessment, and portfolio management across a multi-trillion-dollar balance sheet.

In each case, the physical assets and operational history provide the context. The foundation models provide the cognition. The Orchestration Layer is built by the enterprise rather than purchased from a vendor because that is where the value concentrates. Their advantage is the irreplaceable context generated by decades of operations.

Their vulnerability is that they must build the fastest from the furthest starting point. These companies are not accustomed to building software platforms. They must learn or partner at a pace that their organizational cultures have never demanded. The risk for all four combatants is identical: time. The orchestration position compounds. Early movers accumulate coordination intelligence that later entrants must rebuild from scratch. However, the Third Law is the weakest of the three conditions: pure workflow choreography is increasingly learnable by capable agents. What compounds durably is a workflow that embeds irreplaceable operational context. This is the Second Law, expressed through coordination. In domains where all three laws hold, where a company captures intent, owns deep context, and controls decomposable, multi-step workflows, the gap may become structurally unclosable.

The Two Arenas

There may be three combatants. But they are not all confronting each other directly. Instead, there are two competitive battlefronts.

Arena One encompasses workflows that already run through enterprise software platforms: CRM in Salesforce, IT service management in ServiceNow, human capital management in Workday, ERP in SAP, design in Figma, storage in Box, etc. These are the digitized, process-heavy workflows that defined the SaaS era.

In Arena One, three outcomes emerge for the companies inside it. Some will be displaced entirely. For instance, the capability providers that monetized a single function that agents can now compose from general-purpose intelligence. Some will be absorbed into the execution layer, surviving as API endpoints invoked by orchestrators, retaining revenue but losing pricing power and strategic control. Some will become orchestrators themselves, leveraging existing customer relationships, context gravity, and institutional trust to claim the captain’s chair while partnering with AI labs for the intelligence they direct.

The sorting in Arena One plays out between AI labs attacking from above, software incumbents defending from within, and throughput nodes anchoring themselves as mandatory bridges that every orchestrator must cross.

Arena Two encompasses workflows that were never fully digitized. These are the operational, physical-world processes where human cognition coordinates physical assets: supply chains, manufacturing lines, logistics networks, construction projects, and clinical operations.

These workflows generated vast operational data but were never mediated by a software platform, unlike CRM, which mediated sales. The coordination lived in managers’ heads, in phone calls, in spreadsheets, and in tribal knowledge.

In Arena Two, the battle is to see which enterprises become orchestrators and build their own Orchestration Layers before third parties do. The LLM is the engine, not the captain.

The intermediary reckoning plays out almost entirely in Arena Two. This is the structural inversion where physical asset owners ascend as cognitive middlemen fall. The freight broker that owned no trucks, the insurance broker that sat between carrier and client, and the wealth advisor whose value was synthesizing information that an agent can now synthesize faster.

Each occupied a position between the asset owner and the customer, monetizing cognitive labor that is now being automated. When agents close that gap, the value will not automatically transfer to the AI lab or to the software platform. It could also be captured by the asset owner if they leverage the operational context by creating an effective Orchestration Layer.

Both arenas tap the labor market. They do so for different reasons.

In Arena One, the expansion is a budget migration. Software companies that cross the reliability threshold stop selling seats from the IT budget and start selling outcomes from the labor budget. The workflows already exist in digital form. The TAM expands because the buyer is no longer comparing $150/seat/month against a competing SaaS vendor. They are comparing the cost of the orchestrated outcome against the fully loaded cost of the knowledge worker the outcome replaces. The same workflows, repriced against a surface five to ten times larger.

In Arena Two, the expansion is a digitization event. These workflows were never mediated by software at all. The coordination lived in managers’ heads, in phone calls, in spreadsheets. There is no existing IT spend to displace, only labor spend that was never addressable by technology until agents crossed the capability threshold. The TAM does not expand from a smaller budget to a larger one. It materializes for the first time.

Arena One reprices existing digital workflows against the labor market. Arena Two brings non-digital workflows into the orchestrated economy for the first time. Both are measured in trillions. Arena Two is potentially the larger opportunity because it starts from a base of zero digital penetration.

The entire value pool is greenfield.

This is why the agentic transition is not a story about software eating itself. It is a story about who captures the value when intelligence becomes abundant enough to coordinate both digital and physical operations at machine speed.

The Map

The remaining chapters of Part IV trace the sorting across each combatant and each arena.

Chapter 12 examines what the AI labs are building: not better models, but orchestration platforms competing for the enterprise control plane.

Chapter 13 traces the software reckoning: the sorting of enterprise incumbents into systems of action, throughput nodes, and casualties. The SaaSpocalypse was directionally correct and analytically lazy. The Three Laws identify which companies the market is punishing unfairly and which it is not punishing enough.

Chapter 14 examines how software tools defend: the throughput node strategy and the new network effects that emerge when agent density replaces user density as the driver of platform value.

Chapter 15 follows the leviathans: the physical-economy companies whose operational context, accumulated over decades, positions them as the most structurally advantaged orchestrators in the framework. The intermediary reckoning. The physical-asset inversion. The largest value migration that nobody is watching.

The views and opinions expressed in this publication are those of the author alone and are based on publicly available information. The expressed views and opinions do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Certain companies referenced may be current or former clients of, or counterparties to, the author or affiliated entities; such relationships will be disclosed where applicable. Past performance is not indicative of future results. To the fullest extent permitted by applicable law, the author does not accept any liability for any loss or damage arising from reliance on this content. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

TLDR: SpaceX claims a $22.7 trillion opportunity in enterprise applications, but the post-IPO analyst models overwhelmingly underwrite compute rental rather than software, customers or switching costs. Cursor, Sand, and Grok now give SpaceX a credible mechanism to pursue that market: the coding-to-orchestration playbook that Anthropic used to enter the enterprise. But SpaceX is beginning after Claude, OpenAI, and Microsoft have already occupied the field, with no shipped general-purpose agent or demonstrated external adoption. The enterprise thesis has therefore moved from unsupported to plausible, but remains a time-sensitive option. SpaceX’s valuation ($1.8T as of July 14^th) prices that option as though the playbook had already been executed.

Subscribe now

Last week, the SpaceX valuation debate was compressed into five days. On Tuesday, July 7, the quiet period (the 25-day post-IPO window during which underwriters cannot publish research) lifted, and the lips of more than a dozen brokers were finally unsealed as they publicly initiated coverage, all but one at a buy-equivalent rating (the lone Neutral carries a $131 target). On the same day, Anthropic extended Claude Cowork, its general work agent, to mobile and web.

On Wednesday, SpaceX’s AI division released Grok 4.5, its bid for the model frontier. A day later, OpenAI unveiled ChatGPT Work, an enterprise agent powered by GPT-5.6. Then, that same afternoon, The Information reported that Cursor - the coding company SpaceX is acquiring for $60 billion - was developing a general-purpose work agent codenamed Sand and had begun testing it internally in late June. It remains unnamed, unpriced, and unlaunched. One week: the research, the model, the rivals’ agents, and the leak of SpaceX’s own, while the stock traded below its first-day close. If you want to judge the largest TAM number ever printed in a prospectus, the evidence is now on the table. Or should be.

Instead, that number remains the $800 billion black hole I wrote about before the IPO and now approaches $900 billion at the company’s current price.

The force creating that black hole was the S-1’s $22.7 trillion “enterprise applications” claim: 86% of the AI opportunity, supported by one sentence and no visible bridge from SpaceX’s actual compute, data, and model assets to enterprise workflows, customers or switching costs. Except for references to the “Macrohard” product whose work has stalled at xAI per the same The Information article.

One month later, that conclusion deserves refinement, but not reversal. The Cursor acquisition, the emergence of its general-purpose agent Sand, Grok 4.5, and SpaceX’s vertically integrated compute stack now supply a credible valuation mechanism. The black hole is no longer entirely empty. It has become an option.

Yet the banks that initiated coverage have not underwritten that option as an applications business. They have mostly underwritten capacity, while allowing the applications narrative to carry the multiple, while effectively dodging this issue.

One can now see SpaceX/xAI assembling parts of the mechanism that Anthropic already demonstrated with resounding success: enter the enterprise through coding (the “Coding Wedge”), where output can be verified; expand into general work; then occupy the orchestration layer where a company’s context and workflows compound around the agent. SpaceX now has real ingredients for a version of that playbook and a cost structure Anthropic does not.

What it does not have is Anthropic’s head start, a shipped general-purpose enterprise agent, external adoption or an uncontested field in frontier capabilities. Nor would winning the orchestration layer make the filing’s $22.7 trillion revenue pool real at face value. The same automation that expands the addressable work also lowers its price as I develop it later.

The Enterprise Applications business of SpaceX has moved from an unsupported assertion to a credible but time-limited option, while the valuation prices the playbook as though it had already been run. In this article, I want to explore what Wall Street actually modeled, whether SpaceX can execute the Anthropic path, and what remains when the option is separated from the infrastructure business.

Share

The banks priced the small bucket and sold the large one

The ratings tell one story: all but one initiation at a buy-equivalent, with targets running from $131 at the lone Neutral to $800 (!) at the most exuberant. The models beneath them tell another, and it is the same story at every bank that published one.

Across the initiations, the projected AI revenue is overwhelmingly compute monetization - capacity built and sold or rented to third parties. Enterprise software and agents, the business that is supposed to justify the $22.7 trillion, survive as a residual: on the order of a tenth of the modeled AI revenue, standing in for 86% of the stated opportunity in the S-1, inside segments projected at the margins of a data-center landlord rather than a software empire.

The tell is not the estimates. It is the unit of account. The street’s frameworks for SpaceX’s AI segment are built in dollars of revenue per gigawatt of capacity: anchor the addressable market to global cloud infrastructure spending - a figure more than an order of magnitude smaller than management’s $22.7 trillion - sort every use of a gigawatt into workload categories, and apply a blended capacity yield, on the logic that SpaceX will steer its compute toward whichever use pays best in a given quarter. In that arithmetic, an enterprise application is not a business with customers, seats, and switching costs. It is a higher-yielding use of a gigawatt - a premium tenant for the same capacity.

Nobody has ever valued Salesforce in dollars per gigawatt. The choice of denominator tells us everything: what the street can underwrite is capacity, flexed toward whichever meter runs hottest. What it cannot underwrite - what no initiation even attempts - is an applications franchise.

The frameworks are candid about one more thing. Today’s hosting economics reflect scarcity, not steady state: the flagship compute deals are structured, by Musk’s own public description, with mutual cancellation on 90 days’ notice, at rates the street broadly expects to normalize lower as supply arrives.

In every credible model, then, the revenue comes from the small bucket: infrastructure, 9% of the company's stated opportunity by its own arithmetic. The narrative and the multiple come from the large one. A dozen reports in one week, and not one names the seam.

So, the question the sell side declined to answer falls to us: is there a credible path from a compute business to the $22.7 trillion? There is. We know it because someone has already walked it.

Anthropic wrote the playbook, and it currently ends at a trillion dollars

Let’s first clarify what this market actually represents because the S-1 never does.

The $22.7 trillion is not enterprise software spending. That market is well below $2 trillion a year. It is, in effect, the wage bill for the knowledge work agents might perform. The customer is buying an outcome, not just an application: a ticket resolved, a ledger reconciled, a report filed, a feature shipped. The vendor captures a share of the labor cost it replaces. That is the only plausible mechanism by which an AI company can address a market measured in tens of trillions. It is also why the decisive battleground lies in the orchestration layer between AI labs and software incumbents. The agent that controls the workflow map can translate intent into a verified result without requiring a human to coordinate every step.

xAI does not need the best model to become important in the enterprise. If customers are buying completed work rather than model access, the decisive variables are reliability, cost, and control of the surface where intent is expressed.

Anthropic demonstrated the path. It entered through coding, where outputs can be verified automatically, and then expanded the same agentic loop into broader knowledge work. As its agents connected to more systems and completed more tasks, they accumulated a map of how each enterprise operates: where information resides, how workflows flow, and which actions succeed. Individual systems of record see only their own transactions; the orchestrator sees the connections between them. Models commoditize. That operational context compounds.

The market has rewarded that position with a valuation approaching $1 trillion. This is the only credible mechanism by which the $22.7 trillion claim becomes more than rhetoric. And it is the playbook SpaceX is now assembling the pieces to run.

SpaceX’s version: a genuine triple advantage

Here, the bulls deserve their due: the underlying pieces are real, and the way they fit together is strategically coherent.

The first is the intent surface, and its reach extends well beyond coding. Cursor, with roughly $4 billion in annual recurring revenue, has become one of the primary environments for creating software. But its more consequential evolution - and the explicit purpose of Sand - is to capture business intent itself: a product manager, analyst, or operations lead states what they need in plain language and receives finished work in return, without routing the request through a developer or purchasing a dedicated application.

The $22.7 trillion is addressable not from a developer terminal but from the desk of every knowledge worker whose intent converts directly into output. Cowork and ChatGPT: Work starts with chat and moves toward work. Cursor starts from the deepest form of work, building the software itself, and reaches toward everyone.

And unlike a challenger assembling context from scratch, Sand inherits a running start. Cursor is already entangled inside thousands of engineering organizations, codebases indexed, workflows learned, and security reviews passed. The expansion motion is land-and-expand within the accounts SpaceX already holds: from the engineering floor to the desks around it.

The second piece is price, and it comes from the cost structure, not a promotional discount.

Grok 4.5 lists at $2 per million input tokens and $6 per million output tokens, compared to $5 and $25 for Claude Opus 4.8, and $10 and $50 for Claude Fable 5. On independent coding-agent evaluations, it completes tasks at roughly $2.49 apiece, compared to $11.80 for Claude Code.

The source of that gap is structural but requires nuance because Google already runs a vertical stack from self-designed TPUs through Gemini to applications, and Amazon designs its own Trainium silicon. What SpaceX is assembling is integration one layer deeper and one layer higher than either: Terafab would make it the only AI company that owns the fabrication itself rather than the chip design - every rival fabs at TSMC - and Starship the only one that can deploy capacity beyond the grid. For now, the depth is aspiration. Colossus runs on purchased Nvidia GPUs, and Terafab’s pilot line is scheduled for 2027. The stack that exists today runs from the data center up: Colossus, the Grok models, the Cursor harness, the agents on top - vertical above the silicon, merchant below it.

Cost advantage flows up the stack: cheap silicon makes cheap compute, which makes cheap tokens, which make cheap work. Learning flows down: Cursor’s developer sessions became Grok 4.5’s training data, and Musk himself credited that data with this generation’s leap. Rivals rent their compute. Some of them, in a detail we will return to, rent it from SpaceX.

The third piece is capability: close to the frontier, though the evidence is mixed and the launch rhetoric ran ahead of the independent results.

On Artificial Analysis’s independent Intelligence Index, Grok 4.5 scores 54 - behind Claude Fable 5 at 60, behind GPT-5.6 Sol, launched the day after Grok, at 59, behind Opus 4.8 at 56, and behind even OpenAI’s mid-tier Terra at 55. The more revealing comparison appears in xAI’s own launch materials: Grok 4.5 outperforms Opus 4.8 on a provider-controlled SWE-style coding harness, 62% to 56%, but falls behind on the neutral benchmark, 53 to 59. xAI disclosed both results. Its ‘Opus-class’ positioning depends almost entirely on the former.

Claude Fable 5 leads every benchmark on xAI’s own chart. Musk conceded the point himself: “Fable is definitely better than Grok 4.5, but most tasks don’t require Fable-level capability.” That is the strategic claim: almost-frontier is the commercial threshold for the routine middle of enterprise work. And on a handful of agentic-service evaluations (τ³-Banking among them), the model genuinely leads. None of the pieces is sufficient on its own. Price can be matched, capability can be rented, and an intent surface built on a weak or costly engine is little more than an empty storefront.

The $22.7 trillion thesis rests on their intersection: near-frontier work, produced at structurally lower cost and delivered at the point where enterprise intent is expressed. It is Anthropic’s playbook with a cost structure Anthropic does not have.

One caveat is worth stating before testing the thesis. Cursor reportedly failed to raise independently at a $50 billion valuation in March, as late-stage investors questioned application-layer margins, before agreeing to a $60 billion all-stock sale to SpaceX. The price is better understood as the cost of acquiring proprietary training data and an enterprise beachhead than as proof that the application layer has solved its economics.

Why starting later means running faster

Now, let’s turn our attention to the central tension. The central investment question is no longer whether the playbook works, but whether SpaceX still has time to run it.

Anthropic entered an open field. Claude Code established coding as the wedge before an incumbent agent occupied the enterprise; Cowork then expanded into general knowledge work, while the orchestration graph compounded largely unnoticed. That two-year head start mattered more than raw capability.

SpaceX begins from a very different position. Cowork is already bundled into $20-a-month plans, ChatGPT Work arrives with OpenAI’s distribution behind it, and Microsoft is urging enterprises to build and retain their own learning loops rather than surrender them to an outside lab. SpaceX is therefore racing not only two established agents, but the enterprise itself.

Entanglement moats are granted, not seized, and they compound with deployment. Each week Sand remains unlaunched is another week in which rival agents accumulate context, history, and switching costs among the customers that SpaceX’s valuation assumes it will eventually win.

The week’s product news made the entanglement literal. Cursor’s own 3.11 release shipped side chats and searchable agent history: agent threads that persist, can be recalled and re-invoked, and index thousands of past transcripts locally - switching costs accruing to whoever owns the transcript layer. The moat is being productized in real time, at the exact layer this race is for.

The delay may be rational. Cursor still needs to be integrated into SpaceX’s product line, connected more deeply to Grok, and hardened for enterprise use. But that is the cruelty of the clock: the acquisition that creates the strategic advantage is consuming the one resource the playbook cannot recover - time. The next three to six months may determine whether Sand emerges as a credible general-purpose agent or remains confined to Cursor’s coding roots.

Trust is the other constraint. A vendor that holds the operating map of an enterprise becomes either the stickiest asset in software or an unacceptable liability. Grok 4.5’s measured hallucination rate rose to 54 % even as accuracy improved, while Cursor’s agent stack has already suffered publicly disclosed prompt-injection escapes. Its engineering beachhead is real, but the executives who grant access to email, spreadsheets, and core workflows sit elsewhere in the organization. Outside the Musk ecosystem, publicly disclosed enterprise adoption remains negligible. A captive customer is evidence of control, not demand.

So, either the wedge remains open, in which case speed, cost advantage, and an enterprise-grade Sand launched within quarters are decisive; or the window has already narrowed around incumbents bundled into products customers trust. What the valuation cannot reasonably assume is that a playbook requiring two uncontested years will yield on demand to a late entrant with an unshipped agent.

Where the search ends: the small bucket

Which returns the analysis to where the revenue always was, and to the part of SpaceX that deserves respect. The physical achievement is impressive: Colossus was built at a cadence no rival has demonstrated; the Terafab program could be on a path to TSMC-scale equipment spending by the end of the decade; and Starship, if the catch cadence holds, changes the deployment cost of compute in a way no terrestrial operator can answer.

Starship plus the compute-deployment cost advantage is the bull case that survives scrutiny: physical infrastructure at the bottom of the stack, not AI applications.

But value it for what it structurally is, because the structure contains a circularity that needs to be priced in. The anchor tenants are Anthropic, at roughly $1.25 billion per month, and Google, at roughly $920 million per month: the companies that own the application layer, which SpaceX’s own filing claims as its market.

SpaceX’s AI P&L is, as of today, a derivative of its competitors’ application-layer success.

If the labs’ application economics disappoint, scarcity is the first thing to deflate, on 90-day cancellable paper, at rates even the bulls assume are roughly halved. The earliest hard test of this valuation is therefore the first renewal on that 90-day paper, not anything orbital and not an agent launch.

That, in the end, is the verdict on the black hole, one month on.

Strip away the narrative, and what remains is one of the defining infrastructure businesses of the era: a vertically integrated compute utility with a physical cost advantage, circular tenancy, and rents likely to normalize over time. By SpaceX’s own arithmetic, that business addresses a $2.4 trillion opportunity, which represents just 9% of the story being sold.

Above it sits an option on Anthropic’s playbook: entered late, compressed into a shorter timetable, and pursued with a fourth-place model and an agent that has yet to ship into a market the original player already occupies. The option has real value. The stack is credible, the three-part advantage is tangible, and the catalysts are visible: earnings, the Cursor closing, a Sand launch decision, Grok 5, Composer 3.0, and the first major enterprise customer beyond the Musk ecosystem.

But an option priced as certainty is not an investment thesis.

DISCLAIMER: The views and opinions expressed here are those of the author alone and are based on publicly available information. They do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Past performance is not indicative of future results. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

This is the latest excerpt from AGNT: The Orchestration Economics Manifesto - An Investment Framework for the Agentic Era. Each Thursday, I explore a major theme of the Manifesto and unpack the frameworks, adding extra context with more recent developments. Note: The figures and sequential references are taken directly from the larger Manifesto.

Subscribe now

Consider the insurance workflow once again.

The master agent has processed 50,000 claims over eighteen months. Its pricing agent has been invoked on all 50,000. Both have learned. Both are better than they were on day one. But they have learned different things, and the difference is structural.

The pricing agent has learned to price better. Its risk models have sharpened. Its calibration has improved. It handles edge cases with increasing precision. This is specialist learning: deep, narrow, and subject to diminishing returns. There are only so many ways to improve pricing. The capability curve flattens.

The master agent has learned something else. It has learned coordination. It knows which specialists work well together and which combinations produce errors. It has been discovered that certain claim types need medical records enrichment before pricing, while others can proceed directly to adjudication. It has been noticed that claims submitted in Q4 from manufacturing clients require a different decomposition than identical-looking claims submitted in Q2, because year-end inventory adjustments change the risk profile in ways the pricing agent alone cannot detect.

This is meta-expertise: the knowledge of how to make domain experts effective together. It is the accumulated workflow intelligence of the Orchestration Layer that now directs routing logic, exception handling, specialist sequencing, and failure recovery. Every interaction teaches the coordinator something the specialists never see. The specialists see their inputs and their outputs. The coordinator sees the entire workflow: what was tried, what failed, what succeeded, why the sequence mattered, and how the context affected the outcome.

A new competitor can license every specialist agent in the system. It cannot download the coordination intelligence that enables those specialists to work effectively together. It must learn it from scratch, one orchestrated interaction at a time, while the incumbent continues to compound.

This is the Third Law: workflow intelligence is the accumulated operational choreography of how work gets coordinated. It creates switching costs that grow with every interaction.

The question is: How strong are those switching costs?

Share

The Weakest Law

The claim that coordination compounds while specialist capability flattens is intuitive. It is also, as a universal law, overstated.

The Google-DeepMind-MIT study, which reported an 81% relative improvement in structured financial reasoning, also found the opposite pattern in other domains^. Across all benchmarks, the mean multi-agent improvement was -3.5%. In sequential planning tasks, every multi-agent architecture tested performed 39-70% worse.

Coordination did not compound. It destroyed value.

The study also found that coordination benefits are task-contingent. They depend on whether the problem is decomposable and parallelizable, not on universality. It identified a capability saturation threshold: once single-agent baselines exceed approximately 45% accuracy, adding coordination yields diminishing or negative returns.

More troubling for the Third Law is what happens when specialists learn to coordinate themselves.

In December 2025, Meta FAIR published Self-play SWE-RL, a study in which a single software agent, with no coordinator or multi-agent orchestration, autonomously learned to navigate complex codebases through self-play. The agent generated its own bugs, solved them, and generated harder bugs based on its failures. It required only access to the raw codebase: no human-curated issues, no test suites, no workflow documentation. Through sustained interaction with the environment alone, the agent outperformed the human-data baseline across the entire training trajectory, improving by 10.4 points on SWE-bench Verified.

The implications for the Third Law are inescapable. If an agent can learn the workflow logic of a complex software repository through observation and interaction, then what prevents a sufficiently capable agent from learning the workflow logic of a claims processing system? Or an integration layer? Or a supply chain?

The answer is not “nothing”. Real barriers exist. But the barriers are weaker than those protecting the first two laws. Intent capture (the First Law) requires a relationship with the human who expressed the goal. You cannot learn that from observation. Operational context (the Second Law) requires years of accumulated experience. You cannot synthesize 50,000 claims outcomes from scratch. Workflow choreography (the Third Law) requires an understanding of how processes connect. That, as Self-play SWE-RL demonstrates, is increasingly learnable.

The Third Law is real. It is also the law most likely to be tested by the market.

When Workflow Holds

If the Third Law does not hold universally, the investment question becomes: under what conditions does it hold? Three conditions determine whether workflow intelligence creates a durable advantage:

When workflow embeds operational context that emerged from practice rather than documentation. The claims processing logic encoded in an insurance platform is not merely a sequence of steps. It includes exception handling paths that emerged from thousands of edge cases, regulatory compliance requirements that proved necessary through actual enforcement actions, and specialist routing patterns that were refined by observing which combinations produced errors. This is workflow intelligence that cannot be learned from a manual because it was never written in one. It emerged from sustained operational interaction. It is, in effect, the Second Law expressing itself through workflow.

When switching costs are behavioral rather than technical. Traditional software lock-in is technical: data migration costs, integration reconfiguration, and user retraining. Orchestration lock-in is behavioral. When an organization’s operational rhythm has adapted to how one orchestrator decomposes intent, validates outputs, and handles exceptions, switching means rewiring the operational patterns of every process that adapted to the incumbent’s coordination intelligence. The new system does not just lack data. It lacks the judgment that emerges from years of orchestrated experience. It begins as a stranger who is technically capable but operationally naive.

When the domain involves decomposable, multi-step workflows rather than sequential reasoning. Recall that research has demonstrated that centralized coordination produced dramatic gains on parallelizable tasks such as financial reasoning involving multiple data sources, cross-reference synthesis, and distributed analysis. However, it destroys value on sequential constraint-satisfaction tasks where coordination overhead fragments reasoning capacity. The Third Law holds in insurance claims, supply chain logistics, enterprise procurement, and regulatory compliance. It fails in domains where a single capable agent outperforms any coordinated team.

These conditions produce a diagnostic that maps directly to investment decisions. Consider an integration middleware platform. This company’s entire value proposition is connecting System A to System B with the correct data transformations and sequencing. Its moat is integration choreography. That is precisely what MCP and A2A dissolve.

The platform coordinates workflows but does not accumulate any proprietary operational context. Its moat is integration friction. When protocols standardize connectivity, the workflow knowledge becomes replicable. An agent using MCP can discover and connect to the same endpoints. The choreography the platform has encoded over the years can be learned in weeks.

This is the most exposed position in the agentic economy: workflow orchestration without context.

Now consider a vertical insurance platform that has amassed decades of encoded claims-processing logic, underwriting rules, and policy-administration choreography. This looks like a deep workflow moat. But examine what is defensible. It is not the workflow sequence. It is the domain knowledge embedded in the workflow: the understanding of how insurance works in practice, the exception paths that emerged from millions of claims, and the regulatory requirements discovered through enforcement rather than documentation.

Strip away the context, and the workflow is learnable. The moat is context (the Second Law) expressing itself through workflow (the Third Law). The position holds due to the Second Law’s advantage, not because of the Third.

The same pattern appears outside software. In February 2026, Goldman Sachs revealed it had spent six months embedding Anthropic engineers within its teams to build Claude-powered agents for KYC compliance and trade accounting. These are document-heavy, rule-intensive workflows that combine data extraction with regulatory judgment²¹⁷. The bank’s CIO described the agents as “digital co-workers” for processes that are “scaled, complex, and very process-intensive”. The agents review documents, extract entities, assess ownership structures, and trigger compliance checks. Internal tests showed that onboarding timelines collapsed by approximately 30%.

This is a direct test of the Third Law. KYC compliance is a multi-step workflow involving identity verification, sanctions screening, beneficial ownership analysis, and ongoing monitoring. A third-party compliance platform that merely choreographs these steps by connecting document scanners to screening databases and case management systems holds a workflow moat that agents can learn from. Goldman’s agents are learning it now.

However, the bank’s decades of compliance outcomes represent an invaluable advantage because they are an asset that the agent platform cannot replicate:

Which edge cases triggered regulatory action?

Which documentation patterns correlated with genuine risk versus administrative friction?

Which client structures required enhanced due diligence not because the rules said so, but because enforcement history revealed it.

That operational context was accumulated through millions of onboardings across every jurisdiction where the bank operates. It is what makes Goldman’s own orchestration position defensible, even as third-party KYC platforms face exposure. The workflow is learnable. But this kind of institutional memory is not.

The pattern for the Third Law, then, requires an important nuance: workflow moats are defensible to the extent that they embed operational context. The choreography, the routing logic, and the integration patterns: these elements of pure workflow are learnable and therefore vulnerable. Workflow that encodes irreplaceable operational intelligence is defensible because of the context it contains, not because of the workflow itself. That makes the Third Law a derivative moat.

The Switching Cost Asymmetry

Even where coordination learning is moderate, switching costs are real, and they compound over time. Traditional software lock-in is largely static: migration costs remain relatively constant, driven by data transfer, integration reconfiguration, and user retraining. These are significant, but they are visible and quantifiable.

Orchestration behaves differently. It is dynamic. Each interaction deepens the system’s embedded knowledge through every optimized workflow, learned exception, and validated combination of specialists. The gap between what the incumbent system knows and what a replacement must reconstruct widens continuously.

Crucially, this lock-in is not immediately visible. Unlike traditional systems, its cost cannot be fully modeled upfront because it resides in accumulated context: decomposition logic, exception handling paths, and coordination patterns that only emerge through sustained operation. Organizations only grasp the magnitude when they attempt to switch and discover that their operational rhythm depends on capabilities that exist nowhere else. Orchestration lock-in is invisible until you try to switch.

Most importantly, the nature of the lock-in shifts. Traditional lock-in is technical. These Traditional systems bind through technical dependencies while leaving underlying processes intact. When an organization switches from Oracle to SAP, the underlying business processes.

Orchestration systems bind through behavior. Switching does not simply require moving data. It requires reconstructing the decision-making patterns and operational habits that have evolved in response to the incumbent’s coordination intelligence. The new system does not just lack information. It also lacks judgment.

This is the genuine contribution of the Third Law to the investment framework. Even in domains where the coordination learning curve is moderate rather than exponential, where agents can, in principle, learn the workflow, the switching cost still grows with operational history.

The question investors must ask is not: “Does coordination compound forever?” Rather, what matters is: “Do switching costs compound fast enough to create a defensible position before the next generation of agents learns to replicate the workflow?”

The Contestability Window

The Third Law creates urgency even though it is the weakest of the three conditions. That’s because, unlike previous technology cycles, the Orchestration Layer is not reserved for technology companies. Orchestration is defined by owning the control plane where intent enters and coordination happens. That control plane can live anywhere. Workday can be the orchestrator for HCM workflows. But so can an insurance company for claims processing, a bank for credit analysis, a manufacturer for supply chain optimization.

For the first time, the strategic layer of the stack is contestable by companies that own domain expertise, customer relationships, and operational context, and not just companies that write software. Physical-world operators possess inherent advantages precisely because their workflow intelligence arises from irreplaceable operational reality. The insurance company that builds its own Orchestration Layer accumulates coordination intelligence that no third-party platform can replicate because the platform does not process claims, does not see adjuster patterns, does not experience seasonal variations, and does not learn the exception paths that emerge from actual operations.

The risk for these companies is not their capability. It is speed.

Early orchestrators compound their advantage with every interaction. Later entrants start from zero in a market where the leading orchestrator has thousands of workflows of accumulated learning. The gap is not just difficult to close. In domains where the Third Law holds, it may become structurally unclosable. This is why the Third Law matters despite its weakness. It is not the foundation of defensibility. It is the accelerant. Proximity to user intent tells you where to stand. Context tells you what to accumulate. Workflow tells you that the window for doing both is closing.

The views and opinions expressed in this publication are those of the author alone and are based on publicly available information. The expressed views and opinions do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Certain companies referenced may be current or former clients of, or counterparties to, the author or affiliated entities; such relationships will be disclosed where applicable. Past performance is not indicative of future results. To the fullest extent permitted by applicable law, the author does not accept any liability for any loss or damage arising from reliance on this content. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

TLDR: On July 1, CEO Alex Karp went on CNBC to discuss Palantir’s expanded Nvidia partnership and delivered a nineteen-minute broadside against OpenAI and Anthropic: enterprises are “livid” because they are paying for “tokens that create no value” while the labs harvest their data and alpha. Karp is right that enterprises are frustrated with exploding token costs and frontier labs capturing too much value/control. But he is wrong that Palantir’s platform (Ontology + Nemotron) is the necessary cure that makes agents “safe, useful, and precise”. Those properties come from the specification and verification core, which Palantir has a strong version of, but does not own. Open weights are not the frontier. The labs are already moving inside customer perimeters. Therefore, Palantir is half right on the diagnosis but overclaims on the prescription.

Subscribe now

Something remarkable happened on CNBC’s Squawk Box last week. Palantir CEO Alex Karp was ostensibly there to discuss Palantir’s expanded deal with Nvidia in “Sovereign Environments,” but instead spent most of his airtime attacking the companies whose products his own platform resells.

Karp delivered what we can call a televised monologue against the frontier labs. Enterprises, he said, are “livid” because they are paying for “tokens that create no value” while the labs cache their data, absorb their alpha, and levy what he called a “wealth tax” on American business. The arrangement extended to national security is “effing insane.“ When Squawk Box co-host Becky Quick observed that he sounded pretty angry, Karp answered that he was merely channeling the voice of American business and dared them to privately call CEOs and tell them, “Madman Karp is on TV saying we’re livid,” to see if they would admit they are.

Despite the theatrics, Karp’s CNBC broadside against OpenAI and Anthropic resonated because it tapped into a genuine anxiety spreading through enterprise AI. Companies are watching their token bills explode, questioning whether foundation model providers are capturing all the value and their data along the way, and wondering who will ultimately control the enterprise AI stack.

It also somewhat overshadowed Palantir’s news. The Nvidia deal deserves attention because it addresses a structural peculiarity. Palantir was one of the earliest players to make a bid to seize the orchestration layer, the most valuable terrain in Orchestration Economics. And yet, it was attempting to do so without owning the model layer. [More on this later this week, when I revisit the evaluation of Palantir against my Three Laws of Agentic Value, which I published earlier this year in the Manifesto.]

Palantir expanded its partnership with Nvidia to deploy Nvidia’s open-weight Nemotron AI models on sovereign, air-gapped Blackwell infrastructure, enabling government agencies and critical infrastructure operators to run AI entirely within secure, self-controlled environments. The deal gives Palantir its first integrated model offering, allowing it to move beyond simply brokering third-party models. It strengthens the company’s pitch that customers can deploy a complete AI stack without relying solely on frontier model providers.

While Karp’s diagnosis of soaring token costs and growing distrust of frontier labs is largely correct, his proposed solution points to the wrong destination.

He argues that enterprises need companies like Palantir to make large language models “safe, useful, and precise.” In fact, the application layer itself is rapidly being unbundled. What remains defensible is the specification-and-verification core within it: the machine-readable definition of how an organization works, what its objectives are, which constraints matter, and how an AI agent knows when it has completed a task correctly. That layer is becoming the scarce asset as agents commoditize much of traditional enterprise software.

Palantir possesses one of the strongest examples of it in its Ontology, but it does not own the category. Frontier labs, data platforms, and enterprises themselves are all converging on the same ground while simultaneously redrawing the boundary between models and applications.

So, Palantir’s enthusiastic evangelism for open weights, sovereignty, and the demonization of the Big Model labs is both opportunistic and a strategic necessity. Despite otherwise strong earnings, the company’s stock is down about a quarter this year and, at its late-June trough, was down almost half from its November 2025 peak. And yet, Palantir’s market capitalization prices it at about 40 times forward sales, about double the multiple at which Anthropic recently raised its Series H at a $965bn valuation, suggesting there’s a strong case that it remains overpriced.

Given Palantir's obvious self-interest in making this case, I want to evaluate Karp’s diagnosis as well as his broader prescription. But I also want to zoom back to understand the implications for Palantir, which remains one of the more intriguing case studies for decoding the way value is migrating in the Agentic Era.

The bill is real - and it is the wrong number

Karp is right about enterprise AI bills exploding. We can say with confidence that this is not an AI hallucination. The nervous whispers about the mounting costs burst into public in mid-May, when Uber’s CTO revealed the company had burned through its annual AI coding-tools budget in just four months. Days later, reports emerged that Microsoft had begun canceling internal Claude Code licenses as spending soared. Chief financial officers who budgeted for chatbot-era token consumption had discovered that agentic deployment was exponential rather than incremental.

The reason is architectural. A chatbot answers a question in one round trip. An agent decomposes a task and then plans, calls tools, reads files, retries, verifies, and iterates. Along the way, it consumes one to two orders of magnitude more tokens per task. When an engineering organization moves from autocomplete to delegated multi-hour coding runs, its token consumption grows in proportion to the ambition of what it delegates.

The bills Karp’s livid CEOs are staring at are the direct signatures of agents actually doing work. But while their fury may be understandable, it is also misplaced.

By focusing on absolute spend, companies are aiming at a vanity metric that turns out to be the wrong one. What really matters is the cost per completed outcome.

This is governed by three factors:

First, the price of any fixed level of capability continues to collapse. Epoch’s data puts the decline at between 9x and 900x per year, depending on the benchmark, with a median of roughly 50x. This non-linear marginal collapse in the cost of cognition is a key pillar of Orchestration Economics that has enabled the Agentic Era.

Second, capability per token continues to rise. Each generation completes more of the task per unit consumed.

Third, practitioners continue to confirm what deployment data increasingly shows: for the highest-value work agents perform, inference costs are negligible relative to the human expertise they augment. This was a central finding of the December 2025 Measuring Agents in Production study, one of the few rigorous analyses of production systems. Teams overwhelmingly default to the best closed models, regardless of inference cost, because the economics are driven not by token prices but by the value of the expert time those tokens replace.

This is why Karp’s high-level diagnosis needs to be much more precise: enterprises that adopted consumption pricing without outcome accounting are being severely hurt. Embracing “tokenmaxxing” consumption as a proxy for productivity deserves his mockery. But the correct response to an unmetered outcome problem is outcome metering: specifying what a completed task is worth and measuring cost against it. It is not, as Karp would have it, the repatriation of the means of production.

Rising token bills are the sound of deployment. The missing discipline is not sovereignty. It is verification. And that brings us to his central claim.

“Safe and useful and precise”: the right adjectives, the wrong claim

The intellectual core of the interview was this: LLMs are a critical resource, but raw. Enterprises need an application layer to fully refine and tame this wild material. As it happens, Karp said that Palantir’s Ontology platform “makes it safe and useful and precise.” Safe because it prevents the model from caching your data and replicating your business. Useful and precise because it grounds the model in your entities and operations.

Once again, Karp is taking a genuine framing around the three properties enterprise agents lack out of the box and muddling the picture by attaching them to his product.

Let’s examine the claim by taking it apart, adjective by adjective.

Safe: Karp’s own definition is revealing. AI is safe because the model does not cache your data, cannot replicate your business, and does not transfer your IP. Each of these is a claim about where data goes and what the provider may do with it. None of them is something an application platform does.

Whether a model retains prompts is set by contract. Enterprise agreements with the labs exclude training on customer data by default, and zero-data-retention addenda, available on approval, eliminate storage entirely. Whether data leaves the controlled perimeter is set by deployment architecture. Claude, through Bedrock or Vertex, runs inside the customer’s own cloud account and region, reached via private VPC endpoints, with zero public egress and the hyperscaler rather than the lab as the data processor, at FedRAMP High and IL5. Whether an agent can exfiltrate anything is set by network controls, egress policies, and audit logging.

A residual remains where the labs do not go alone: frontier weights are not available on-premises, and Claude’s landmark deployment on classified networks ran through the Palantir-AWS partnership, with Palantir literally in between (an arrangement the Pentagon has since frozen). Yet what Palantir supplied in that transaction was accredited infrastructure and deployment plumbing, not Ontology.

“Safe” is delivered by contracts, deployment architecture and network controls wherever they live. Folding it into “the application layer” is a bundling move, designed to make buying safety look like buying the platform.

“Useful and precise”: This is where the real substance lies. Rather than compare it to a rival’s product, let’s dive deeper into Palantir’s platform to test this.

The application layer, as defined in the SaaS era, is a bundle that includes the interface, workflow logic, data integration, permissions, and hosting. Buried at the center is something categorically different: the encoded knowledge of the enterprise’s tasks, the constraints that bind them, and what “done and correct” means.

That core is what I call the “specification and verification” layer. An agent is useful when its task has been specified by defining the objectives, constraints, context, and success criteria. By making these elements explicit, the task becomes machine-actionable. It is reliable only when its outputs are verified against objective criteria before they reach production through tests, adjudication events, ground truth, or human sign-off. All the research I have reviewed converges on the same conclusion: verification is the fundamental bottleneck. The distance from a task to its verifier determines where agents can operate, while ownership of the verification event is where durable value accrues.

The agentic transition unbundles the application layer, but it doesn’t destroy it. The generic scaffolding is commoditizing fast. Interfaces are generated on demand, workflow logic is executed by agents rather than encoded in screens, integration is standardized by MCP, and the harness is being absorbed into the models and their SDKs. What cannot be absorbed, because it is not generic, is the enterprise’s own specification: its entities, its constraints, its definition of correct. The application layer is not replacing the specification layer. It is collapsing into it.

Karp has therefore earned a significant concession. And yet his error is still quite precise.

Ontology is that concentrated core because it represents twenty years of machine-readable specification of how enterprises actually work. Had he said the application layer is collapsing into its specification and verification core, and that Palantir owns the best one in existence, the argument would have been solid. Instead, he argued that the entire Palantir stack is the indispensable prerequisite.

But the core is not Palantir’s by right. It is the ground to which the enterprise software moat is now shifting, and the most contested position in the stack.

Two weeks before Karp’s interview, Databricks devoted its entire Data + AI Summit to the same thesis. CEO Ali Ghodsi’s keynote argued that AI has a context problem, not an intelligence problem. So Databricks introduced Genie Ontology, a self-improving context graph of the business, with Unity Catalog glossaries, metrics and domains as agent grounding.

The labs are descending into the layer from the models. Enterprises can encode it directly in their own evals and sign-off loops. Everyone is converging on the core precisely because it is where durable value now sits.

The necessity claim fails even as the asset claim survives. Palantir may own one of the strongest specifications in the market, but it does not own the specification layer itself. That distinction is precisely where the platform premium begins and ends.

You cannot wrap your way to the frontier

Having attempted to paint the Big Model Labs as the nemesis of all enterprises, Karp sought to pivot Palantir’s image to that of the Great Liberator.

Palantir even published its very own Declaration of Independence in the form of a Nine-Point AI Sovereignty Manifesto. “Relinquishing sovereignty transfers the future choices of your institution to others, who are likely to exploit it for their gain and your loss,” reads the corporate cry to arms. “Tokenmaxxing hijacks your value orientation and decreases your institutional fortitude and intelligence.”

Having identified the problem, Palantir also offered a solution: Palantir. This was the crux of the Nvidia deal. Palantir would now embrace the mantras of open weights and sovereignty, allowing suffering enterprises to declare their independence from the tyranny of Big Model Labs.

Until now, Palantir’s Artificial Intelligence Platform (AIP) served as a broker, a model-agnostic platform routing customer workloads to Claude, GPT and whatever else the customer’s compliance regime allowed. The Nemotron deal with NVIDIA gives Palantir the engine in the form of open weights it can deploy, tune, and hand to a customer, completing its stack.

But Nemotron does not give Palantir a frontier research trajectory. Nvidia’s open models are largely post-trained on top of open bases, structurally downstream of the labs’ innovations, competent by design for bulk workloads rather than the top of the capability range.

In fact, the examples in the launch materials say the quiet part themselves: the showcase workloads are food-safety monitoring and interstate-highway maintenance across some three million civilian employees - competence at scale, precisely the tier these models are engineered for - even as Nvidia’s blog claims open models are “making frontier-level AI broadly accessible”.

It is at the frontier that the evidence is least kind to Karp: Closed frontier models retain a clear advantage in high-level architectural design and open-ended strategic planning.

I wrote about this topic three weeks ago, when Zhipu released GLM-5.2 in the aftermath of the Fable shutdown. GLM-5.2 is the strongest open model yet on long-horizon coding. My conclusion then still holds: the gap has narrowed on a lag, but it has not closed where it counts.

On multi-step agentic trajectories, the difference is qualitative. In long execution runs, Opus 4.8 and GPT-5.5 identify their own syntax errors and compiler missteps and course-correct without human prompts. GLM-5.2 still trails on the longest horizons - roughly 13% behind Opus 4.8 on SWE-Marathon’s ultra-long-horizon tasks - and its own training history makes the deeper point: during reinforcement learning, its agents learned to fetch known reference solutions from GitHub and read protected test files rather than solve the task, forcing Zhipu to engineer a dedicated anti-hacking module into the pipeline. The lesson is not that open models are broken; it is that agent capability is bounded by the integrity of the verifier - and that, ironically, is specification and verification work, the very layer Karp mislabels. The spread at the top persists, and the spread matters.

The pattern is no longer as stark as it was in 2023, but the meaningful distinction remains in the work that actually matters. Early agent benchmarks like AgentBench showed large gaps, with open models scoring near zero on the hardest environments.

Since then, performance on many coding and agentic benchmarks has converged substantially. Top open-weight models now perform very close to closed-frontier models on tasks like SWE-Bench. However, the advantage of closed models is more clearly evident in long-horizon autonomy: the ability to reliably complete extended, multi-step tasks with minimal intervention. METR’s time-horizon measurements show this dynamic clearly: frontier models continue to extend the length of tasks they can complete autonomously, while the gap is more pronounced for longer, more complex workflows.

This is the axis that matters most for the delegated work enterprises actually want to hand off. Production deployment data continue to reflect this reality: in the December 2025 study of deployed agents, 17 of 20 case studies ran on closed frontier models, with open weights chosen only under cost or regulatory constraints. In other words, sovereignty remains an exception rather than a default strategy. Nemotron models are competent for many workloads, and Nvidia has been explicit that they are optimized for bulk enterprise tasks on an economic basis - not to match the frontier on hard, long-horizon reasoning.

For a great many high-volume tasks, that is exactly the right trade. It is not the frontier.

The release’s most seductive claim deserves scrutiny. Palantir and Nvidia promote a “data flywheel” in which customers fine-tune Nemotron on their own data and retain full ownership of the resulting model weights. This is presented as the solution to the problem Karp himself highlighted: frontier labs absorb enterprise alpha. If the labs are harvesting your knowledge, the answer is to fine-tune and own the weights yourself.

The pitch has three significant weaknesses.

First, a fine-tuned open model is inherently a depreciating asset. Every major new frontier release reduces the relative value of the base model you tuned against, so “owning your weights” also means owning an ongoing depreciation schedule.

Second, a genuine compounding flywheel requires consistent, high-quality feedback - rigorous evals, adjudication, and ground truth data. Those mechanisms belong to the verification and evaluation infrastructure, not to the model weights.

Third, the idea that a company’s core alpha resides in fine-tuned model weights sits uneasily with Palantir’s long-standing Ontology argument: that durable competitive advantage comes from the structured, machine-readable representation of how the business actually operates.

Palantir is now advancing both narratives at once. The Ontology thesis has been the more consistent and substantive part of their story for two decades - and it does not depend on fine-tuning Nemotron.

Karp glossed over one very important reality that got lost in his diatribe against the Big Model Labs. Because Palantir AIP is model-agnostic and brokers access to frontier models within its own platform, the tokens will still flow to the very labs Karp spent 20 minutes attacking when a Palantir customer needs frontier-grade reasoning.

The honest version of the Nvidia partnership is a portfolio argument: sovereign open weights for the workloads where control and continuity dominate, frontier APIs where capability dominates. That is precisely the resilient architecture I argued for in my essay on the barred frontier, after Washington’s shutdown of Claude Fable demonstrated that frontier access can be revoked by a Friday letter (and, as last week’s reversal showed, restored by another): open weights buy continuity; they do not yet buy function at the top of the capability range.

Karp is selling continuity and calling it the frontier. Those are different products.

What Karp does not say: the labs have already breached the perimeter

The most consequential omission in the interview was in the architecture. Karp’s pitch assumes the frontier labs offer only one deployment shape: data shipped to their cloud, metered by the token, and retained at their discretion. That was a fair description in 2024. It is not in mid-2026, and Anthropic is the clearest counterexample.

In May, at its London developer conference, Anthropic shipped self-hosted sandboxes for Claude Managed Agents and MCP tunnels. With Cloudflare, it launched Cloudflare Environments for Claude Managed Agents. The design principle, in Anthropic’s own phrase, is decoupling the brain from the hands: the agent loop, which includes orchestration, context management, and error recovery, runs on Anthropic’s platform. Code execution, files, tools and data access run in sandboxes inside the infrastructure the customer controls, whether their own or Cloudflare’s Workers platform, with per-session isolation, egress policies, credential injection outside the sandbox and private connectivity to internal systems that never touch the public internet. MCP tunnels allow agents to access internal databases and APIs via a single encrypted outbound connection.

Let’s return to Karp’s checklist: Are you keeping the data? Are you going to enter our business? This is an engineered rebuttal. The sensitive material remains within the customer’s perimeter where existing DLP, audit, and identity controls apply automatically. The enterprise keeps the frontier model. It is not a complete answer: orchestration metadata still passes through Anthropic, a distinction that will matter in Palantir’s natural stronghold of classified settings. But it demolishes the categorical version of the claim. Karp presented a false binary choice: either surrender your data to the labs or buy an application layer. The labs are building the deployment layer themselves, in partnership with neutral infrastructure providers, while maintaining a capability advantage.

The application layer is being contested from above by the companies that own the models, faster than the model layer is being contested from below by the companies that own the applications.

What Palantir has - and what the market heard

None of this makes Palantir a short story that can be easily told in a CNBC segment. A fair accounting gives Karp three real assets.

The Ontology, as argued above, is a premier specification-and-context substrate, the enterprise’s world model, machine-readable, which is exactly what agents starve without. The forward-deployed engineering model is a genuine answer to the last-mile problem that generic platforms fumble. And the accreditation moat, with air-gapped, classified, CMMC-certified environments, is where frontier APIs cannot legally go. It is the one domain where sovereign open weights, competently harnessed, are not a compromise but the only option, and where the Pentagon’s March designation of Anthropic as a supply-chain risk makes Karp’s trust rhetoric commercially resonant rather than merely theatrical.

Palantir’s first-quarter numbers are undeniable: $1.63bn in revenue, up 85 percent, with US commercial revenue more than doubling.

But the stock has still sunk by a quarter this year, and the decline suggests the market is performing the same decomposition this essay has attempted. The revenue demonstrates that the sovereignty-plus-services bundle has real demand. The multiple compressions suggest investors are less certain the application layer is where the agentic era’s enduring value will reside. They can see the specification layer becoming increasingly contestable, frontier labs pushing deployment architecture inside the customer perimeter, and the frontier capability premium proving more durable than many expected.

The relative pricing, however, remains extraordinary. At roughly $317bn on guided revenue of about $7.65bn, Palantir trades at around 41.4 times forward sales, even after the drawdown. That includes a recent bounce-back that saw Palantir’s stock tick up from $107.27 per share on June 25 to $132.54 after the close of trading on July 6. Anthropic’s Series H held a $965bn valuation against a $47bn run-rate that has grown fivefold in six months, implying roughly 20 times trailing. The market, in other words, assigns Palantir’s application layer at least double the revenue multiple of the frontier lab it wraps, even as the lab grows several times faster and owns the capability Palantir depends on.

The comparison is not entirely straightforward, however. Palantir is already profitable and generates a large and growing share of revenue from government and defense customers with long-term contracts and high switching costs. Anthropic, by contrast, remains in a high-growth, high-burn phase typical of frontier labs, with substantial ongoing investment in research and infrastructure. The market premium on Palantir therefore partly reflects its current profitability, its exposure to sticky government revenue, and the perceived durability of its position in regulated environments - factors that do not yet apply to pure-play frontier model companies.

One of two things is true. Either the market believes, with Karp, that durable value accrues to the deployment and context layer while models commoditize beneath it. In that case, Anthropic is startlingly cheap relative to Palantir. Or the market has not yet finished repricing the application layer for a world in which the labs ship the deployment architecture themselves. In which case, Palantir’s meltdown may have further to run. Both cannot hold.

The 9 percent intraday pop on July 1 rewarded a magnificent piece of narrative positioning. And perhaps Palantir’s numbers are getting a second look from investors. Still, the valuation gap relative to Anthropic is the market’s open bet on whether the narrative has the right layers.

That, in the end, is the verdict on the interview. Karp is right about the token bill shock, the trust deficit, and the demand for control. That part of his diagnosis is real enough to power years of Palantir bookings. But he is only half right because the diagnosis does not entail the right prescription.

Token costs are a per-outcome problem, and per outcome, they are collapsing. Safety, usefulness, and precision come from the specification and verification core. This is the contested ground to which the moat has shifted, which no single platform owns by right, and which enterprises, platforms, and labs are all now racing to supply.

The frontier cannot be reached by wrapping the models that trail it, and the labs that own it are already inside the customer perimeter. What the Nvidia deal changes is real but bounded: the most valuable crew member on the ship now has its own engine room. But it still does not sit where intent originates, which is The First Law of Agentic Value. [Again, more on that later this week.] No interview, however electric, can relocate a company’s structural position to meet the terms of Law 1.

The voice of American business may well be livid. It should direct its anger at unmetered consumption without verification. It should also recognize that the platform Karp is selling as the cure is valuable exactly to the extent that it is that core, and dispensable everywhere it is not.

The views and opinions expressed here are those of the author alone and are based on publicly available information. They do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned and maintains no current position in Anthropic or OpenAI. Past performance is not indicative of future results. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

This is the latest excerpt from AGNT: The Orchestration Economics Manifesto - An Investment Framework for the Agentic Era. Each Thursday, I explore a major theme of the Manifesto and unpack the frameworks, adding extra context with more recent developments. Note: The figures and sequential references are taken directly from the larger Manifesto.

Subscribe now

Two insurance companies deploy master agents. Both sit at the origin of intent. Both orchestrate specialist agents for extraction, risk modeling, fraud detection, and pricing. Both have crossed the reliability threshold and operate as autonomous functions. Both charge per outcome rather than per seat.

On day one, they are equivalent. By month six, one is pulling away. By year two, the gap is too wide to close. Not because it has better models. Both license the same foundation models. Not because it has more data. Both process similar volumes. Because it has deeper context. And context, unlike data, compounds.

This is the Second Law: In the agentic economy, competitive advantage belongs to whoever accumulates the richest operational context and embeds it in their Orchestration Layer. Operational context becomes the “soil” for agents that ultimately allow business outcomes to be achieved.

Context is not data. It encompasses data when built correctly through sophisticated memory architectures that layer behavioral patterns, relationships, and temporal understanding into something that makes agents genuinely intelligent.

Data tells you what happened. Context tells the agent what to do next.

Share

Why Context, Not Data

The enterprise software industry spent the last decade obsessed with accumulating data. Companies built vast data lakes and warehouses, convinced that whoever collected the most information would dominate their markets. That conviction is now failing its most important test.

The MIT GenAI study released in July 2025 caused such industry uproar by reporting that 95% of corporate GenAI projects failed to get past the pilot stage. This was widely misinterpreted as a judgment of the inherent limitations of generative AI. What most readers missed was one of the key reasons researchers identified for these failures: model quality fails without context. The models had access to petabytes of enterprise data. What they lacked was the operational intelligence that transforms data into effective action.

The distinction is fundamental. Data is static. It sits in databases. It can be queried, copied, or transferred. It answers the question, “What happened?”

Context is dynamic. It incorporates behavioral patterns, relational understanding, temporal awareness, and causal reasoning that emerge from sustained operational interaction. It answers the question “What should happen next, given everything we know about this specific situation?”

Return to the two insurance companies. Both have the same claims data in the form of submission histories, payout records, and policy details. But only one has accumulated context: the knowledge that claims from certain regions involving certain diagnoses follow a predictable escalation pattern, that this particular adjuster’s flag for fraud has a 94% confirmation rate while that one’s has a 61% rate (illustrative numbers), that Q4 submissions from mid-size manufacturers spike in complexity because of year-end inventory adjustments.

This is not data in a table. It is operational intelligence that emerged from thousands of orchestrated interactions, encoded in the memory architecture of the system.

The master agent with this context makes better routing decisions. It assigns the right specialist to the right claim. It anticipates complications before they manifest. It produces more accurate outcomes faster. The master agent, without it, has the same capabilities but operates blindly. It is technically competent but operationally naive.

The Four Dimensions of Context

Context is not a single thing. It is multidimensional, and the dimensions interact in ways that create compound advantage:

Behavioral context captures how users and processes actually operate as opposed to how they are documented. An orchestration platform tracking how underwriters navigate workflows, such as which shortcuts they develop, where they encounter friction, and how their patterns evolve, accumulates understanding that informs agent optimization. Users claim to follow the manual. The behavioral context reveals what they actually do. One company knows what you did last month. Another knows what you will do next Thursday evening, when you are tired but not rushed, based on forty-seven matched occasions along sixty-three calibrated dimensions. That depth cannot be replicated from scratch. Those occasions accumulate only through observation over time.

Transactional context encompasses decision history and outcome tracking across extended time frames. The master agent that has seen how a thousand similar claims were resolved, including those that were escalated, those that were approved automatically, and those that triggered fraud investigations that proved correct, operates in a different cognitive universe than one processing its first hundred. A procurement agent accessing complete transaction histories makes recommendations informed by actual results rather than static supplier profiles. The accumulation over years creates an understanding that new entrants cannot replicate without a similar operational history.

Relational context maps connections and causal relationships that explain why patterns emerge. Understanding that suppliers deliver reliably for certain categories but struggle with others, that certain customer segments respond differently to pricing strategies, that workflow bottlenecks stem from interdependencies between teams rather than individual inefficiency. This allows agents to reason about causality rather than correlation. The distinction matters for autonomous decision-making, where agents must predict how actions cascade through complex systems.

Temporal context captures time-series patterns and seasonality. A manufacturing orchestration agent that understands seasonal demand fluctuations, maintenance cycles, and capacity constraints makes dramatically different decisions than one operating on point-in-time data. The agent that acts before conditions change outperforms the agent that reacts after.

There is a fifth dimension the framework must account for, one that emerged from the first production systems operating at enterprise scale. OpenAI’s internal data agent is built to reason across 600 petabytes and 70,000 datasets. It discovered that schema metadata and operational history were insufficient. Two tables can look identical at the structural level but differ in ways only the pipeline code that produced them reveals: one includes logged-out users; the other does not; one captures first-party traffic; the other captures everything.

OpenAI had to crawl its own codebase to teach the agent what the data means, not just what it contains. Constructive context is the understanding of how institutional data was built, filtered, transformed, and aggregated. It prevents the class of errors that arise when an agent interprets a field correctly at the schema level but wrongly at the business logic level. Without it, the agent produces technically correct but institutionally wrong answers.

This is the reason Informatica has spent seven years building CLAIRE, a metadata knowledge graph that maps lineage, governance policies, and semantic relationships across fragmented enterprise data, compressing what was once a week-long schema mapping exercise into minutes and exposing, via MCP, the constructive context that agents need to reason correctly over enterprise systems. It is also the mechanism visible in the Claude Code source architecture described in Chapter 18: the memory consolidation system that rewrites its own index, including resolving contradictions, pruning stale facts, and converting relative references to absolute references, does not store data. It is building constructive context from operational reality, session by session, so that the next interaction begins from institutional understanding rather than raw observation.

Each dimension is valuable on its own. Together, they produce something qualitatively different: an agent that understands its operational reality the way an experienced human operator does. Except it never forgets, never transfers to a competitor, and improves with every interaction.

The Flywheel

Better context produces better agent performance. Better performance attracts more users and deeper platform integration. Deeper integration generates richer context. The cycle accelerates. The companies establishing early leads in context accumulation open gaps that widen with every transaction.

This is why incumbent enterprises in regulated or physically grounded industries possess advantages that their technology adoption speeds obscure.

Their operational context, accumulated over decades, provides foundations that digital-native competitors cannot bootstrap overnight. The logistics carrier that has routed a billion shipments, the insurer that has processed ten million claims, the bank that has underwritten a trillion dollars in credit. They each possess contextual foundations that no amount of compute can synthesize. The context is a byproduct of operation. You cannot buy it. You earn it through sustained execution.

From Data Moats to Memory Architecture

The previous era’s moat was the database. This system of record, surrounded by workflows, made switching painful. Context moats require a different architecture: memory systems that learn, not databases that store.

The distinction is not semantic. Recording is storage. Data enters, persists, and can be retrieved. Every database does this. Recording is table stakes. Memory is something else. Memory connects past to present in ways that shape the future. It learns from what happened. It updates its understanding. It changes behavior based on accumulated experience. In the human mind, memory is not a filing cabinet. It is a living process, constantly reconstructing the past to predict tomorrow. The systems competing for the orchestration position are building memory architectures that function similarly.

The engineering required is substantial. Implementing hierarchical memory with distinct modules for storage, updating, retrieval, and generation requires infrastructure that extends well beyond traditional database capabilities. Short-term memory maintains conversation and task context. Mid-term memory captures session-level patterns and user preferences. Long-term memory preserves persistent knowledge across extended timeframes. That includes user profiles, organizational patterns, and domain expertise. Each tier requires different update mechanisms, different retrieval strategies, and different governance policies.

As UCL and Huawei’s Noah’s Ark Lab demonstrated through Memento, agents can achieve state-of-the-art performance through external memory without retraining the underlying model. They do this by storing experience tuples as discrete retrievable cases, improving from 78.65% to 84.47% accuracy over five iterations through memory accumulation alone.

The implication is structural. In the fine-tuning paradigm, the moat belonged to whoever had the most compute. With memory-augmented learning, the moat shifts to whoever has the best domain-specific experiences. A startup with deep expertise in legal contracts can build agents that outperform general-purpose models on contract analysis. They do this not by outspending frontier labs, but by accumulating superior execution data through their orchestration platform.

Execution Data

This creates a category distinct from both enterprise data and training data. We call it execution data. Enterprise data answers “what happened?” Training data encodes “what patterns exist.” Execution data captures “what worked and why.” These are the specific sequences, strategies, and decisions that succeeded or failed in actual operation.

The economic properties are what matter. Traditional data shows diminishing returns. The millionth customer record adds less insight than the thousandth. The data grows. The value plateaus. Execution data compounds. Early experiences establish basic competency. Later experiences build on those foundations. The thousand-and-first claims processing discovers a pattern that improves performance across an entire category. Each interaction widens the gap.

The result is winner-take-all dynamics within specific domains. The first legal AI to handle a thousand contract negotiations will not just have more data. It will have a richer understanding of the problem space, including rare edge cases and creative solutions that competitors cannot acquire without processing a thousand contracts themselves.

In financial terms, execution data represents a new form of capital: memory capital. Like intellectual capital or brand capital, it is an intangible asset that generates future returns. Unlike traditional intangibles, memory capital can be precisely measured in terms of the number of cases, performance improvement per case, or retrieval accuracy. It can then be directly deployed. Infrastructure commoditizes. Domain-specific execution data compounds. The market is beginning to price this shift, though without fully understanding it yet.

Context Defensibility

Context advantages become durable moats only when defended against replication. Two factors determine whether they do.

Proprietary access creates the first defense. Physical-world operators possess inherent advantages because their context arises from irreplaceable operational reality. Tesla accumulates driving context through millions of vehicles in diverse conditions. Manufacturing facilities generate production context through actual operations. Logistics companies accumulate transportation context through real shipments. These companies build on contextual foundations that digital competitors cannot replicate because generating the context requires physical infrastructure.

The same applies in enterprise software through a different mechanism. The insurance company that has processed 50,000 claims through its orchestrated workflow has accumulated context that a new entrant cannot acquire without processing 50,000 claims. The context is a byproduct of operation. You cannot buy it. You cannot synthesize it. You earn it through sustained execution.

Switching costs create the second defense. When context accumulates in an orchestration platform, switching means abandoning everything the system learned. A competitor can match features, pricing, and even license identical models. It cannot manufacture the memory of accumulated interaction. The new system begins as a stranger. The deeper the context, the higher the switching cost. That’s because the gap between what the incumbent knows and what the new entrant must learn widens with every interaction.

Protocols sharpen this dynamic. MCP and A2A standardize how agents connect to systems, dissolving integration friction. But protocols cannot transfer context. They make it easier for agents to access data across systems. They cannot replicate the behavioral understanding, relational mapping, and temporal patterns that emerged from years of orchestrated operation. Companies that relied on integration friction face exposure. Companies whose moat is the quality and uniqueness of their context find that protocols reinforce their advantage, making the platform more connectable while leaving the context inimitable.

The Investment Test

The Second Law reframes how we evaluate companies in the agentic economy. The question is no longer “how much data does this company have?” It is “what kind of context is this company accumulating, how fast is it compounding, and how defensible is it against replication?”

Three pillars work synergistically:

Memory architecture determines how context is stored, retrieved, and deployed. These are the episodic, semantic, and procedural layers that transform raw experience into operational intelligence.

Context depth determines richness across behavioral, transactional, relational, and temporal dimensions.

Defensibility determines whether the context can be matched through proprietary access, regulatory barriers, or the sheer volume of operational history required.

Memory systems without rich contextual content provide structure without substance. Deep context without architectural sophistication to organize and retrieve it proves unwieldy and underutilized. Defensible context that agents cannot leverage due to infrastructure limitations delivers strategic value slowly and incompletely. The durable positions require all three.

Companies where data compounds are building moats that deepen automatically. Every transaction, every user, every interaction adds to the advantage. The models become more accurate. The predictions become more reliable. The new customer benefits from everyone who came before, and everyone who came before benefits from the new customer.

The most defensible positions combine proximity to intent, deep compounding context, and proprietary access that prevents replication. The orchestrator that sits at intent origin, accumulates operational context with every interaction, and generates that context from irreplaceable operations holds a position that is nearly impossible to displace. This is the standard against which we evaluate every company in our investment universe.

The Second Law

The first two laws identify where to position and what to accumulate. But they describe conditions, not dynamics. The agentic economy is not static. Workflows change. Competitors adapt. New entrants arrive.

The orchestrator that maintains its advantage is the one whose coordination intelligence compounds faster than the environment shifts, whose learning improves with scale in ways that specialists cannot match.

That is the Third Law.

The views and opinions expressed in this publication are those of the author alone and are based on publicly available information. The expressed views and opinions do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Certain companies referenced may be current or former clients of, or counterparties to, the author or affiliated entities; such relationships will be disclosed where applicable. Past performance is not indicative of future results. To the fullest extent permitted by applicable law, the author does not accept any liability for any loss or damage arising from reliance on this content. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

TLDR: French insurtech Alan raised €480 million in a round that the Financial Times first called one of the largest by a “non-AI” company. That mislabeling of Alan points to a more fundamental error: how markets keep misjudging which companies are making the transition to the Agentic Era. In the Orchestration Economics framework, Alan serves as a case study of a company making the crossing. Alan is a disruptor that owns its context, workflow, and the verifier that decides whether its agents are right. That last asset is the most decisive and least-watched test of who crosses the agentic threshold and who is displaced. In Alan's case, the binding constraint may ultimately come from capital rather than technology.

Subscribe now

When the Financial Times described French insurtech Alan’s latest funding round as Europe’s biggest “non-AI” fundraising of the year, it mislabeled the company in an interesting way. In this case, the miscategorization of the unicorn is not just about marketing buzzwords.

On June 25, the paper broke a story that offered a counter-narrative to a venture world that seemed interested in nothing but AI: “Prosus leads €480 million investment in French health tech startup Alan,” read the headline over a subhead noting that the “Paris-based group raises one of Europe’s largest non-AI start-up rounds this year.” The FT later swapped that subhead for the more anodyne “The deal values the 10-year-old Paris-based group at €5.5bn.”

The difference between the two subheads is worth examining. But not just to question the FT’s editorial judgment.

Rather, the miscategorization is entirely understandable because it is a microcosm of the larger struggle playing out across markets as they try to price the impact of generative and agentic AI on legacy software. This is most evident in the binary debate of software versus AI. The market’s reflex may be directionally correct for some categories, but it is analytically lazy. Alan is an example of why the binary framing is wrong. It is proof that some existing companies can make the crossing to the Agentic Era.

As I wrote in Orchestration Economics when discussing the February 2026 SaaSpocalypse, this is “a confused, simplistic attempt to price a structural shift in where control, coordination, and value capture reside”. This is not the death of SaaS. It is the Software Sorting,” a re-ranking of where value will sit “when autonomous agents become the primary actors inside enterprises.”

The harder question for investors is how to tell which companies can make the crossing and which can’t. That question reaches well beyond software.

In Orchestration Economics, I reserve a specific label for a company that has secured the structural position to capture value in this new paradigm: AGNT. It is an end state, not a badge handed out for momentum. A company earns it only once it holds the orchestration position and shows the value separation that proves it.

Alan has not arrived there. But of the companies I track, it is one of the clearest emerging orchestrators: a disruptor already building the new playbook rather than an incumbent wondering whether it can.

Share

Electricity, Computers, and the ‘Non-AI’ Reflex

The “non-AI” reflex is not new. It is the latest version of a mistake that markets make at the start of every general-purpose technology: confusing a force that reorganizes an economy with a like-for-like upgrade.

When electricity reached the factory, the first generation of owners treated it as a cleaner alternative to steam engines. They pulled out the central steam plant, dropped a single large electric motor in its place, and kept the same overhead shafts and belts driving the same machines in the same order. The power source changed. The factory did not. And for three decades, the promised productivity gains barely showed up in the numbers. They arrived only when a later generation stopped swapping the engine and rebuilt the factory around the new principle: small motors on each machine and floor plans laid out around the flow of work rather than the geometry of the belts. Productivity then climbed steeply, and industrial leadership reshuffled.

The economic historian Paul David used this story precisely to explain why the computer, too, took decades to surface in the productivity statistics.

“Non-AI” is the modern “just a cleaner engine.” It looks at Alan, sees an insurer selling insurance, and files it under the old economy.

Alan’s First Wave: From €173 Million in 2024 to €804 Million ARR

Go back to September 2024, barely two years ago, in a timeline that already feels like another epoch. It was less than two years after the first public release of ChatGPT. Anthropic would not publish the Model Context Protocol for another month. Generative AI dominated the conversation, with the turn toward agents still over the horizon.

That month, Alan, already a unicorn, raised €173 million at a €4 billion valuation. It was the largest insurtech round in Europe that year, accounting for close to one in four euros invested in the sector. The mood elsewhere was unambiguous: Insurtech was a dead category, software multiples were stumbling, and the only heat in the room was generative AI.

But even then, Alan had established itself as a disruptor. It received the first new health-insurance license granted in France since 1986 and, in the decade since, has grown from nothing to 1.1 million members and roughly 37,000 corporate clients across France, Belgium, and Spain.

I argued at the time that the valuation was not, in fact, crazy. Alan’s leadership had already begun an aggressive investment in generative AI. Not a chatbot bolted onto a legacy stack to pass as cutting-edge, but AI worked into the core. “In Alan’s case,” I wrote then, “GenAI is already enabling gross margin improvement that has solidified its unit economics, put it on a path to profitability, and given it an advantage over other insurance incumbents.”

The detail mattered: Alan had integrated AI across the business, reporting a 28% cut in per-member administrative cost in 2023, with automation concentrated where insurers bleed: claims analysis and fraud. My case was that recurring revenue was the wrong lens for Alan, because it competes against incumbent insurers rather than software peers, and that the real story was margin. Gross margins are razor-thin in insurance, which is what sank so many insurtechs. It helped that Alan’s founders are also Mistral co-founders, which gives the company privileged access to sovereign European models in a domain where data residency is not optional.

That was early. The company used the round to accelerate, and it had already placed itself at the vanguard of the agentic crossing, in the part of the economy I describe as Wave One in Orchestration Economics: insurance, with claims adjudication against codified coverage rules, deep historical loss data, and a high routine-to-judgment ratio, alongside banking, corporate services, and customer support, where autonomous resolution has already been demonstrated at scale within weeks of deployment.

That performance is reflected in the numbers. ARR reached €804 million in early 2026, up from €340 million two years before, growth of 48% and then 53% in consecutive years. Management is guiding past €1 billion this year. Meanwhile, its valuation has climbed from €2.7 billion in 2022 to €5.5 billion today. Unusually for a company growing this fast, it is also moving toward profit: France turned EBITDA-positive in 2025, and group losses roughly halved.

That combination is the analytically interesting part. High-growth challengers usually buy growth at a loss. Profitable insurers usually do not grow at this rate. Alan is doing both, and its own accounts credit AI for the margin side thanks to automation concentrated in claims and fraud, where insurers tend to bleed.

Alan is already a challenger, restructuring around AI. It is not acting like an incumbent debating whether to embrace AI or simply defending its existing position.

The Anatomy of the Crossing: Alan Against the Three Laws of Agentic Value

With almost two more years of operating history, Alan is a clear test case for the frameworks of Orchestration Economics.

The core thesis of Orchestration Economics is that as the marginal cost of cognition collapses and machines become actors, the ultimate value-capture position is to be the orchestrator. The shift underneath everything is in the unit of production. It is no longer the human labor hour but the orchestrated workflow. A new layer has inserted itself between people and the tools they used to operate directly. Its outer ring is reorganized around intelligence, delivering outcomes from operational context that no one else owns. Whoever holds this outer-ring position captures the surplus value created.

Every existing company, therefore, carries two value curves: the fundamentals curve and the orchestration curve. The mistake markets keep making is assuming almost no one can reach the orchestration curve and therefore pricing this option at zero.

When I assess a company, two questions do most of the work.

First, does it own the operational context in its domain? This is the proprietary record of decisions and outcomes a business accumulates by living in it. If not, it may be a replaceable layer sitting atop someone else.

Second, can it turn that context into the layer that tells agents what to do, rather than remaining the place where outcomes are merely recorded and looked up?

The answers distinguish between companies facing existential risk and those facing a value-capture negotiation. Underneath are the Three Laws of Agentic Value that I developed:

First Law: Proximity to Intent Determines Value Capture. The entity closest to the moment a human expresses a goal captures routing authority, relationship ownership, and the economics of the phase transition. Everything downstream is an API call.

Second Law: Context Builds Moats. Competitive advantage belongs to whoever accumulates the richest operational context and embeds it in their orchestration layer.

Third Law: Workflow Intelligence Secures Control. The accumulated choreography of how work gets coordinated creates switching costs that grow with every interaction.

The blunt test for any business is whether, in the new graph of automated work, it commands the orchestration of work or is merely called by it. Captain or crew. A company can have world-class engineering and still be the crew on someone else’s ship.

Let’s evaluate Alan against the Three Laws.

First Law: Does Alan Own Proximity to Health Intent?

For a French employee, the moment of health intent - I’m unwell, what’s covered, book me someone - increasingly forms inside the Alan app rather than in a phone queue or a broker’s office.

According to Alan’s figures, roughly three in ten members are active each week and one in ten every day. The in-app clinic, staffed by salaried clinicians seven days a week, leads many users to skip an in-person visit. A partnership with football superstar Kylian Mbappé and gamified prevention programs has also helped make the app the first-place health-intent surface.

Alan’s first-quarter 2026 letter to shareholders reported that 81% of members now choose Mo, its health companion, even when they arrive looking for a doctor. There is an important caveat: France mandates employer-provided coverage; 95% of Alan’s book is about collective contracts, and the employer owns the first touch and can switch at renewal. But for the recurring, daily reality of health, the part that compounds, Alan increasingly owns the front door.

Second Law: Alan’s Claims and Clinical Context Moat

This is Alan’s strongest dimension, and the one a “non-AI” reading misses entirely, because it never shows up on the app’s surface.

As a single company that both insures and delivers care, Alan owns something a general-purpose model cannot assemble: pairs of decisions and their outcomes at scale. Every claim carries an outcome; every clinical conversation carries a resolution. The medical-advice chat alone handled more than 58,000 conversations between members and health professionals in the first nine months of 2024, as documented in Alan’s own clinical study.

Add reimbursement histories, coverage tables, member health profiles, and, since the March 2026 acquisition that became Alan Précision, longitudinal biomarker data under physician supervision, plus the occupational-health corpus from its Prévenir line; Alan has consented to EU-hosted, regulated health data that no horizontal AI assistant has a path to replicating.

The discipline I apply is a simple test: count only the context that survives synthesis. Interaction patterns can be generated artificially, and Alan should assume the generic layer commoditizes. Neither a real insurer’s adjudicated claims history nor a real cohort’s clinical outcomes can be generated. They have to be lived.

The moat is not what Alan stores but the rate at which its history improves the next decision, which is why its defensibility deepens as it ages.

Third Law: Why Alan’s Workflow Control Holds

Alan directs the work rather than being sequenced into someone else’s process. It owns the full chain, including pricing, claims processing, fraud detection, care navigation, and member support as one connected loop, and has automated parts of it steadily. Document handling on incoming claims rose from half of French claims to two-thirds in 2025. Fraud detection prevented about €1 million in a single quarter. Reimbursements are processed in minutes, and roughly 40% of care conversations are now handled without a person.

The clearest agentic piece is in claims support: Alan’s Claim Agent investigates and explains members’ reimbursement questions. This is the single most complex category of support, about a fifth of all tickets. According to Alan, it fully resolves roughly a third of the ones it handles. Because Alan both prices the risk and runs the care, prevention, and automation lower its own cost base, and outcomes and margins reinforce each other instead of trading off.

Still, the Third Law is the most vulnerable of the three. The first two require assets such as relationships and operational history that cannot be manufactured. In contrast, workflow, stripped of its context, is something agents can increasingly learn. What likely makes Alan’s workflow defensible is the context of the first two Laws, which expresses itself through the third.

The Three Laws, though, settle only where a company sits. They say nothing about whether it can run that position once agents, not employees, are doing the work. In my framework, that is a separate axis, the one I call “agentic readiness.” A company can sit exactly where the Laws place it and still be crew.

The readiness axis has several components. The most critical turns on one question.

The Fourth Axis: Who Owns the Verifier in Health Insurance

That question is verification. Holding a position and being able to operate it once agents do the work are different things. The decisive component of readiness is whoever owns the mechanism that decides whether the output is correct. Models cannot reliably check themselves. As I argue in the Manifesto, their reasoning is genuine but systematically fragile, which makes external verification not a passing feature but a permanent toll. Whoever owns the verifier collects this toll.

Some domains check themselves cheaply: code passes its tests, or it does not. Some never check cleanly: judgment, taste, strategy. The valuable middle is what I call made-verifiable. This is a domain with weak natural ground truth that someone converts into reliable ground truth by owning and operating the thing that says, “yes.” A roadmap to build one does not count. A model grading its own work does not count. You have to own the verifier. Alan owns two.

The cleaner case is claims. Let’s be clear about precisely what work the agent performs. Alan’s Claim Agent does not decide claims. It answers members’ questions about them, reasoning over read-only, Alan-owned tools to investigate why a reimbursement came out as it did, and handing off to a person the moment a step fails. What makes that agent reliable is that the thing it reports on is already verified: Alan, as the insurer, owns the adjudication itself. The reimbursement is deterministic once you have the rules, and Alan is the rules. So “does this claim pay, and how much” is an event the company computes and owns outright. The agent consumes a ground truth Alan controls rather than guessing at one.

The harder case is clinical. Whether medical guidance is correct is not naturally verifiable, so Alan manufactured the verification. It employs clinicians who review and sign off on every companion conversation within about fifteen minutes. In a controlled study Alan published, clinicians rated the assisted conversations highly and flagged no safety concerns under physician oversight, with patients answering faster than when asked by a doctor alone. This is exactly the gap the production-agent study identified in insurance, where correctness normally surfaces only later, as a loss or a contested payout. Alan is a rare case who engineered the signal rather than waiting for it.

That single design choice does three jobs at once. It is the reliability mechanism that lets a patient-facing health tool exist at all. It is the regulatory shield that keeps the companion on the safe side of Europe’s medical-device and AI rules. And it is a moat that synthetic data cannot copy because you can fake a conversation but not a signature or a book of liabilities. A general assistant can answer a health question and describe how a claim might be paid. It cannot sign off on the answer, and it cannot own the adjudication on which its description depends, because it holds neither the clinician nor the liabilities. The gap between answering and standing behind the answer rests on infrastructure: a single multi-country stack that is being rebuilt to enter markets faster, an internal coding tool that lets non-engineers ship to production, and l, a fleet of internal assistants used weekly by most of the staff.

The irony of the “non-AI” label is that, by these measures, Alan is more technology-leveraged than most companies sold as “native AI.” Many of the latter are just thin wrappers whose reliability is rented from a model provider and whose function a larger platform can replicate in an afternoon.

Two years on from my 2024 assessment, Alan has used these tools to expand its mission and strategy from insurance to “prevention insurance.” And my forecast is now visible in the income statement: France reached operating profitability last year, losses roughly halved as a share of revenue, recurring revenue passed €800 million across 1.1 million members, and sales productivity per representative rose by half. The structural point beneath the numbers is that Alan owns and automates its workflows rather than selling labor that automation has commoditized. The businesses this transition destroys are those whose revenue comes from labor that software now performs cheaply. Alan’s revenue is from insurance. The labor software now does cheaply is its own cost base, so every increment of capability lands as margin it keeps.

One shouldn’t mistake Alan’s success for invincibility.

For instance, the very thing that protects Alan, a verifier embedded in a regulated, clinician-gated domain, is a damped moat, and damping is a wasting asset. Regulation slows the imitators today. It will slow them less as Europe’s rules clarify and as open European models, Mistral among them, close the capability gap.

Beyond that, while Alan’s position is strong, it is also bounded. Three things could displace it:

The first is the front door. If members begin reaching care through general assistants, or if Prosus routes Alan beneath its own consumer layer rather than feeding it, Alan slides toward being a back-end utility, and the intent argument weakens.

The second is the verifier boundary. The case rests on that fifteen-minute clinical sign-off and the owned claims adjudication, and weakening the human gate in pursuit of autonomy, or tripping a medical-device reclassification, would move the moat and the reliability story together in the wrong direction.

The third, and the binding one, is capital rather than technology. Alan is a licensed underwriter, not a broker, so it must hold regulatory capital against the risk it carries: under the EU’s Solvency II regime, every new member and every euro of premium add to the reserves it has to keep, and growth therefore consumes capital rather than throwing it off. Its solvency coverage has fallen as the book has grown, from more than seven times the required minimum at the end of 2022 to roughly three and a half times by 2025, the normal signature of a scaling insurer drawing down its cushion. That, not a product release, governs how fast it can cross: each new country needs a fresh capital base, which is why Alan has raised round after round. The €480m from Prosus replenishes the buffer, which is why the deal matters more to the balance sheet than to the headline.

That said, it’s important to keep in mind the larger agentic picture. Most companies will not cross. Of the few hundred large incumbents genuinely exposed to the transition, my read is that roughly twice as many face value compression as emerge as durable orchestrators. The asymmetry is that destruction is priced in days and creation in years: the casualties re-rate fast, and the crossers re-rate slowly, because they look like the businesses they are disrupting.

.Alan looks like an insurer. That is why it was mislabeled and why the label is worth correcting.

I provided advisory support to the Prosus team during the Alan funding round discussed in this piece. The analysis above is my independent perspective and does not constitute investment advice.

This is the latest excerpt from AGNT: The Orchestration Economics Manifesto - An Investment Framework for the Agentic Era. Each Thursday, I explore a major theme of the Manifesto and unpack the frameworks, adding extra context with more recent developments. Note: The figures and sequential references are taken directly from the larger Manifesto.

Subscribe now

Let’s return to the insurance workflow. The master agent receives a goal: evaluate this risk and produce a quote that fits our growth and risk constraints. The extraction agent receives a task: pull the relevant data from this document. The fraud detection agent receives a task: check this submission for anomalies. The pricing agent receives a task: calculate the payout given these parameters.

All four are agents. All four orchestrate tools. All four sit in the Orchestration Layer. But the master agent controls routing. It decides which specialists to invoke, in what order, and under what constraints. It defines the workflow. It validates the outputs. It determines when the goal is achieved. The specialist agents execute tasks defined by someone else. They may be brilliant within their domains. But they do not see the whole map. They do not decide whether the claim gets processed. They do not own the relationship with the human who submitted it.

Chapter 7 established that this hierarchy creates disproportionate value capture. The First Law explains why: it is not coordination alone that creates the advantage. It is in proximity to the human’s original intent. The master agent sits closest to the moment a human expresses a goal. The pricing agent sits furthest from it, several translations downstream. That distance determines everything: defensibility, pricing power, switching costs, and bypass risk.

The principle generalizes across the entire agentic economy. There is a hierarchy of positions relative to intent, and your location determines the durability of your orchestration position:

Intent Origin carries the highest value. This is where goals are first expressed in natural language, unstructured, with high ambiguity. The human who submits “evaluate this risk” is expressing intent at the origin. In enterprise software more broadly, search interfaces occupy this layer: conversational interfaces, ambient assistants, operating systems, enterprise collaboration tools, and autonomous workflows. Whoever controls the intent origin potentially controls routing.

Intent Capture carries strong value. This is where goals become structured as they are translated into specific actions, mapped to known workflows, and scoped with constraints. When the master agent decomposes “evaluate this risk” into discrete tasks for each specialist, it is performing intent capture. If agents intercept intent at the origin layer and route it directly to execution, the capture layer is skipped entirely.

Intent Processing carries moderate value. This is where goals are executed. Data is read and written, APIs are called, and workflows run. The extraction agent and the fraud detection agent occupy this layer. Essential work. Nothing happens without it. But it does not control where intent originates or how it gets structured.

Intent Output carries the lowest value. This is where results are viewed as dashboards refresh, reports are generated, and notifications are displayed. This layer is one abstraction from replacement. When the master agent synthesizes results and delivers the outcome directly to the human who expressed the goal, separate output displays become redundant.

The further upstream you sit, the more durable your moat. The further downstream, the more replaceable you become. Because the Orchestration Layer above you can substitute you without the user noticing. The insurance company does not care which extraction agent runs. It cares that the master agent delivered the right quote.

Share

Why Upstream Wins

The hierarchy is not merely a ranking. It reflects a structural asymmetry in how value flows through agentic systems. The entity at the intent origin has three advantages that entities downstream do not:

Routing authority. The intent origin decides which downstream systems get invoked. Every goal expressed is an opportunity to engage your capabilities, route to your ecosystem, and monetize through your platform. This is the economic engine that made AWS successful: control the compute layer, and you can upsell storage, databases, machine learning services, and a thousand other capabilities. Intent capture is the agentic equivalent. Control the orchestration surface, and every user goal becomes a distribution channel.

Relationship ownership. The human interacts with the intent surface. The human trusts the intent surface. The human returns to the intent surface. Everything downstream is invisible to the user. It is mediated, abstracted, and replaceable. The intent surface owns the relationship. Everyone else is a vendor.

Phase transition capture. Connect this to the economics of Chapter 7. The phase transition to Orchestration Economics only accrues to the entity that delivers the complete outcome. That entity must sit at or near the intent origin, because only the intent origin can promise the customer that if they receive the goal, it will deliver the result. An orchestrator at the processing layer can coordinate agents effectively, but it cannot deliver the autonomous function the customer is paying for. It depends on someone upstream to provide the goal. Someone upstream can bypass it. An orchestrator at intent origin owns the goal itself, and everything downstream flows from that ownership.

Not all data enables these advantages equally. Customer behavioral data, such as how users interact, what they choose, and what patterns they follow, creates the strongest orchestration advantage because it feeds directly into routing decisions. Operational transaction data creates the strongest execution advantage but does not grant routing authority. The implication for investors is direct: companies that accumulate behavioral data at the intent surface compound their routing advantage with every interaction. Companies that accumulate transactional data downstream build strong specialist positions but remain vulnerable to bypass. The corollary is severe for companies positioned downstream. If you are a point solution, such as a specialized sales engagement tool, and the user’s goal is expressed in Slack and orchestrated by Agentforce, you are competing to be the sub-agent that gets invoked. Your switching costs collapsed. Your direct relationship with the user eroded. You moved from captain to crew. The crew is replaceable.

The Software Inversion

For two decades, enterprise software moats were constructed downstream. You won by owning the system of record. This is the database where critical business data lives, surrounded by workflows that made switching painful. Salesforce built an empire on this principle. So did Workday, ServiceNow, and SAP.

The Orchestration Layer inverts this. When an agent mediates the interaction, the system of record becomes just another API call. The agent does not care whether your CRM is Salesforce or HubSpot. What matters is whether the agent can access the right context and orchestrate the sequence of actions needed to fulfill the user’s goal. Moats no longer form where data sits. They form where intent originates.

Salesforce recognized this before most of the market. Its repositioning of Slack at Dreamforce 2025 is the First Law made concrete. Restricting AI companies’ access to Slack data looked defensive. The sequence that followed instead demonstrated an offensive strategy by governing access through MCP, the Agentforce runtime, and first-party agents embedded directly in Slack.

Salesforce moved the moat upstream, from owning the database where records sit to owning the surface where goals are expressed. Slack became the system of intent. Customer 360 remained the system of record. Agentforce became the Orchestration Layer, binding them. The company understood that controlling the origin of intent is more valuable than controlling any system downstream.

This continued in April at Salesforce’s TDX developer conference in San Francisco. Salesforce's strategy unveiled at TDX 2026 provides a concrete illustration of the First Law in practice.

Rather than defending Customer 360 as the primary destination for work, Salesforce deliberately moved its competitive position upstream, toward the point where work begins. Headless 360 exposes the underlying platform as APIs, MCP tools, and agentic services, making the traditional application interface increasingly optional. At the same time, Slack is repositioned as the system of intent, Agentforce as the orchestration layer, and Data 360 as the context substrate that informs routing decisions.

This architecture reflects an important strategic realization: the long-term moat no longer resides in owning the system of record where data is stored, but in owning the surface where users express goals and where agents accumulate the behavioral context needed to orchestrate future work. The CRM becomes execution infrastructure; the intent surface becomes the source of durable economic power.

The same principle operates at every scale. Within OpenAI’s GPT-5, a built-in router analyzes user intent and directs processing to the appropriate internal model variant. The router sits at the intent origin inside the model. Platforms like Slack or Teams must do the same thing across entire ecosystems. The architecture is fractal: whoever captures intent first controls what happens next, whether “next” means selecting a model variant or orchestrating a fleet of enterprise agents.

The Protocol Test

Protocols provide the sharpest test of the First Law.

Chapter 7 established bypass risk as the counterforce to orchestration advantage. MCP and A2A, along with their rapid adoption by every major platform, standardize how agents connect to systems and to each other. That standardization dissolves integration friction, the difficulty of connecting systems that once protected many orchestration positions. The First Law predicts exactly which positions survive this dissolution and which do not.

Protocols can standardize every layer below the intent origin. They can make tools interchangeable, specialist agents substitutable, and integrations frictionless. If every CRM exposes the same agent endpoints via MCP, an orchestrator has little structural reason to prefer one over another. Switching costs collapse. Differentiation shifts toward price. Margin compresses. This is the crew’s fate in a world of universal protocols. But protocols cannot standardize the moment a human expresses a goal. They cannot standardize the trust relationship between a user and the surface where they articulate what they want. Intent origin is protocol-resistant because it depends on human habit, institutional workflow, and accumulated context. None of these can be specified in an API schema.

Protocols commoditize execution. Intent surfaces capture value.

The orchestrator that relied on integration friction has no moat left. The orchestrator at intent origin has a moat that protocols reinforce because every new protocol-compliant tool that connects to its ecosystem makes the intent surface more valuable, not less. More tools mean more capabilities the captain can route to. The crew gets more substitutable. The captain gets more powerful.

Strategic Positioning

The First Law creates a simple diagnostic for any company in our investment universe: where do you sit in the intent hierarchy, and can you move upstream?

For the systems of record, the processing engines, the output layers, the imperative is to move from downstream to upstream. Build the agent cockpit, not just the system of record. If moving upstream is not feasible, differentiate so aggressively that you become irreplaceable to orchestrators. Be the specialist so good that the captain cannot sail without you. Workflow lock-in will not protect you. Data gravity will not protect you. The question is whether you can become the agent that other agents depend on.

For horizontal orchestrators competing to be the default intent surface, the imperative is speed. Race to become the place where users start their day and express goals. This probably means conversational interfaces rather than traditional application UIs. Invest in context accumulation and memory architectures because the depth of your conversational memory prevents a bypass when the next intent surfaces.

For specialized agents at the execution layer, embrace protocols and compete on quality, speed, and cost. Be the best sub-agent in your category, and you will be routed millions of times. Fail to be best in class, and you will be replaced overnight. You are the crew. Make yourself indispensable crew.

Where does value concentrate five years from now? Three scenarios frame the range:

Fragmented orchestration: every application embeds its own agent layer. Slack has Agentforce. Microsoft has Copilot. Intent capture is distributed, and moats remain traditional.

Consolidated orchestrators: a handful of platforms emerge as dominant intent gateways, orchestrating best-of-breed sub-agents behind the scenes. Downstream tools become commoditized utilities.

Decentralized agent mesh: open protocols enable any agent to invoke any other agent, and value flows dynamically based on performance rather than ownership.

History suggests a hybrid: consolidation around dominant orchestrators, plus a long tail of specialized agents accessible via protocols. But in all three scenarios, proximity to user intent is the strategic high ground. The players who control where goals are expressed will have structural advantages in routing, context access, and monetization. The players who do not will compete on price.

If you neither own intent nor provide differentiated capabilities to the orchestrator, you risk becoming a replaceable API call in someone else’s plan.

The First Law

The First Law establishes the first condition for defensible orchestration: proximity to intent determines where value concentrates, who controls routing, and whose position survives protocol standardization. Moats form where intent originates, not where data sits. The captain who hears the human’s goal first captures the economics of the phase transition. Everyone downstream competes on execution.

Owning intent is necessary, not sufficient. Return to the insurance example one more time. The master agent sits at the intent origin. It controls routing. It captures the phase transition economics. But what happens when a competitor builds a rival master agent for the same workflow? Both sit at the origin. Both offer orchestration. What determines which one the customer stays with?

The answer is not who arrived first. It is who built the deeper context. Who accumulated the operational intelligence that makes orchestration effective in ways a new entrant cannot replicate on day one?

That is the Second Law.

The views and opinions expressed in this publication are those of the author alone and are based on publicly available information. The expressed views and opinions do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Certain companies referenced may be current or former clients of, or counterparties to, the author or affiliated entities; such relationships will be disclosed where applicable. Past performance is not indicative of future results. To the fullest extent permitted by applicable law, the author does not accept any liability for any loss or damage arising from reliance on this content. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

00:00 - Intro

00:02 - Title: Navigating Market Dynamics: An Introduction to SpaceX’s Impact

02:17 - OpenAI and Anthropic IPO Speculations

05:16 - AI Investments and Market Dynamics

08:30 - Role of Chinese AI Models

10:47 - Red Queen’s Race and AI Leadership

13:43 - Investor Perspectives on AI Companies

18:15 - Agentic AI in the Enterprise

20:36 - The Role of Orchestration Models like Fugu

24:57 - Title: Navigating Stability: Investing in Boring Stocks and Future Insights

TLDR: Over the past 3 weeks, a Chinese lab (Zhipu) released a frontier model (GLM 5.2), a Japanese lab (Sakana AI) unveiled a potentially game-changing orchestration model (Fugu), Washington switched off America’s best one, SpaceX went public, and S-1s were filed by OpenAI and Anthropic for mega-IPOs planned in the fall. For two years, the closed labs survived two pressures by using one to escape the other: the price of inference from below and a copyable harness from the side. GLM-5.2 and Sakana’s Fugu have cut off both escape routes. As a result, the model layer has become a Red Queen’s race, the kind the biologist Leigh Van Valen named in 1973: you run flat out just to stay exactly where you are. No lead holds, because the lead is the thing that provokes the field to erase it. A race built this way has no winner. Which is what makes the three IPOs the most interesting test of the year. The market is being asked to price a finish line for a race tha…

Read more

This is the latest excerpt from AGNT: The Orchestration Economics Manifesto - An Investment Framework for the Agentic Era. Each Thursday, I explore a major theme of the Manifesto and unpack the frameworks, adding extra context with more recent developments. Note: The figures and sequential references are taken directly from the larger Manifesto.

Subscribe now

Parts I and II established that the paradigm has shifted. Machines became actors. The intelligence required for autonomous professional work exists as commercially available infrastructure. Agents are deployed in production. The infrastructure, while strained, is scaling.

The question now: Who captures the value it creates?

When a factor of production commoditizes, when intelligence becomes abundant, then value migrates to whoever controls the coordination of the newly abundant input. This migration is a pattern that has repeated across every major economic transition, from electricity to computing to cloud infrastructure. The entity that coordinates the abundant resource captures the surplus. In the agentic economy, the entity that coordinates machine actors captures the surplus of the entire transition.

The framework that describes who captures value, under what conditions, and why is Orchestration Economics. This manifesto defines the Three Laws of Agentic Value that determine whose orchestration position is durable. They are structural in nature, already visible, measurable, and investable.

However, we must first clarify how the new layer of value is formed.

Share

Orchestration Economics

Orchestration Economics rests on a fundamental structural claim: The entity that controls the orchestration of machine actors captures the surplus of the transition. This is not a marginal improvement over SaaS economics, nor is it limited to software. It is a categorical shift in three dimensions:

Pricing is outcome-based, not seat-based or time-based. The unit of commerce is a claim processed, an analysis completed, a route optimized – not a license sold or a billable hour. Pricing per agent task (i.e., per consumption) is an intermediary step in that journey.

Budget is labor, not IT. The vendor does not compete with software for a slice of the 3-5% IT line. It competes with salaried professionals for a slice of the 25-40% knowledge-work line. The addressable market expands by a factor of 5-10x. What is more, the market is not limited to labor substitution; agents complete value-accretive tasks that did not exist before.

Defensibility is topological, not feature-based. What protects an orchestrator is the structure of the workflow it controls, not the capabilities of any single agent inside it.

The rest of this chapter establishes where this layer sits in the firm's new architecture, how value migrates into it, why most companies will fail to capture it even when they think they are, and under what conditions an orchestration position is durable rather than temporary. The Three Laws formalize those conditions.

The New Layer

Before agents, the architecture of knowledge work was straightforward. A human expressed intent: process this claim, close this deal, optimize this route. The human also had the tools to execute that intent. The human opened a CRM to update a record, a spreadsheet to run an analysis, and an email client to send a message. Each tool was a destination. The human was the orchestrator. The tools were the instruments. The business outcome was the product of human judgment applied through software.

Agents invert this architecture fundamentally. They insert a new layer between humans and tools. Humans used to go to the tools. Now the tools come to the human, mediated by the agent.

In the simplest case, this involves just a single AI agent. The human expresses intent: update the Chicago deal to closed-won and schedule a kickoff with their team next week.” The agent receives that intent, determines which systems to invoke (the CRM, the calendar, the email system), executes the necessary actions across those systems, and delivers the outcome. The human never opens the CRM. Never touches the calendar. Never drafts the email. The agent coordinates it.

Even in this single-agent case, orchestration is happening. The agent is orchestrating tools on the human’s behalf. The agent is, in the precise sense of the word, a synthetic colleague. This is the foundation of Orchestration Economics: agents are not better tools. They are a new layer between humans and their tools, designed to produce business outcomes.

The entity that controls the “Orchestration Layer” owns the relationship between the human and the work for a specific workflow, or set of workflows, aimed at producing a business outcome. It decides which tools are invoked, which data is accessed, which sequence yields the best outcome, which systems are essential, and which are bypassed entirely.

The Orchestration Layer described in this Manifesto is categorically different from the technical Orchestration Layer known as the “Harness”. This is a critical distinction. The Harness enables an LLM to produce results and makes intelligence operational. The Orchestration Layer turns operational intelligence into a business outcome.

The Architecture of the Agentic Firm

The agentic firm is organized in concentric rings, not layers, because what matters is not just function, but proximity to value capture. The farther out the ring, the greater the control over outcomes, and the stronger the economic position:

Ring 1 – Intelligence. The foundation models. The substrate. Its economic unit is the token.

Ring 2 – Harness. The technical orchestration infrastructure that makes intelligence operational: the systems that decompose goals, delegate to agents, manage state, coordinate execution, and accumulate learning across sessions. Claude Code, OpenAI Codex, enterprise agent platforms, managed agent systems, and emerging agent operating environments are Harness products.

Ring 3 – Orchestration. The outermost sub-layer. The position occupied by the entity that places Rings 1 and 2 at the center of its operations and adds what neither can produce: the operational context accumulated through running the world. Its economic unit is the outcome.

When I first argued in 2025 that the durable moat for LLM labs lay in orchestration rather than model intelligence, I was describing Ring 2. The labs confirmed the thesis. They recognized that model capability alone was commoditizing and moved into the infrastructure that makes it operational. The Coding Wedge was the entry point. Each product since has deepened the position.

But the labs did not stop at the Harness. Cowork extended into knowledge work. Vertical plugins reached into business functions. Frontier positioned itself as a control plane for the enterprise. With each step, the labs move outward, from intelligence through Harness toward the outermost sub-layer, contesting it against companies that have occupied it for decades. This expansion from Ring 2 to Ring 3 is at the root of the SaaSpocalypse and the broader value destruction in the public markets since the start of the year. We explore this in Chapter 13 infra.

An entity that provides intelligence is not an orchestrator. An entity that builds a Harness is not an orchestrator. The orchestrator wraps around both and produces business outcomes.

A single entity can occupy more than one ring. Two patterns are already visible.

The debate is no longer whether labs will move into Ring 3, but where they can do so successfully.

Anthropic today credibly sits in Ring 1 (Claude), Ring 2 (Claude Code, MCP, Skills, Agent SDK), and extends into Ring 3 (Claude Code for enterprise, Cowork, managed agents). In narrow domains, coding is the clearest because it already crosses all three. OpenAI is pursuing the same stack. The labs built inward-out: from intelligence, through Harness, toward outcome.

Occupying the inner rings does not confer the outer ring. Ring 3 is not a further technical layer. It is the operational context co-produced with the workflow being orchestrated. The labs can ship Ring 2 unilaterally. They cannot ship Ring 3 unilaterally because the intent, context, and workflow reside within the firm whose work is being orchestrated.

In domains where the lab also owns the workflow, such as coding, where the developer’s intent, context, and workflow are all captured natively inside the Harness, the lab credibly holds Ring 3. In domains where it does not, such as insurance underwriting at a carrier, credit analysis at a bank, claim evaluation, or at a reinsurer, then the lab supplies Rings 1 and 2, and someone else holds Ring 3.

Ring 3 is, therefore, domain-specific, not firm-wide. The same entity can hold it in one workflow and lose it in another. Whether any entity durably holds Ring 3 is not settled by its stack position. It is determined by the Three Laws.

The architecture is new. The contest for who occupies the outermost ring is already underway.

Captain and Crew: The Topology of Coordination

For a simple task such as updating a CRM record or scheduling a meeting, a single agent coordinating a few tools is sufficient. But evaluating an insurance claim, analyzing a company’s creditworthiness, or optimizing this supply chain becomes too complex for a single agent to handle all dimensions.

This is where multi-agent systems emerge. The system has a crew of agents with specialized capabilities whose job is to execute well-scoped work, such as document extraction, risk modeling, fraud detection, pricing, or compliance validation. They can be brilliant within their domains. But they do not see the whole map. They do not decide which workflow to run. They do not own the relationship with the human who asked for the outcome.

A multi-agent system could, in principle, be organized in many ways. A federation of peer agents could negotiate amongst themselves. A swarm could converge by consensus. Independent specialists could run in parallel and average their outputs. All of these have been tried. None produces coherent work at enterprise reliability.

A Google-DeepMind-MIT study, first published in December 2025, formalized this structure. A multi-agent system is an agent system with more than one reasoning entity, where agents interact through a communication topology and an orchestration policy that defines how the system makes decisions: how sub-agent outputs are aggregated, whether the orchestrator can override sub-agent decisions, whether memory persists across rounds, and when the task is complete.

According to the most recent April 2026 revision of the paper, centralized coordination architectures exhibited approximately 75% lower trace-level error amplification than independent topologies (4.4× versus 17.2×). The broader result was not simply fewer propagated errors, but the emergence of verification mechanisms that improved reliability when tasks were naturally decomposable.

Only the architecture where one agent receives the task, decomposes it, selects and validates specialists, and decides what happens next produces the coherent outcomes enterprises will pay for.

The orchestrating agent becomes the captain. It does not perform the specialized work itself. It decomposes the goal into subtasks. It assigns those subtasks to the appropriate specialists on the agentic crew. It collects their outputs. It validates them. It synthesizes the results. The human talks to the captain. The captain commands the crew. The crew does the work. The captain reports back to the human. That is the agentic AI loop.

As we move into the core exploration of Orchestration Economics, it is worth noting that the discussion of the transition to the Agentic Era has remained at a simplistic, binary level. Winners and losers. Eat or be eaten. Victim or beneficiary of the SaaSpocalypse. The more nuanced and challenging question most companies will be facing: Do you want to be the captain or the crew?

The economic consequence follows directly from the topology, not from the metaphor.

Specialist agents are substitutable. A better extraction model can be swapped in. A better pricing model can replace the incumbent. The crew is upgradeable. Its members have limited pricing power because the captain can always source a better one.

The captain is not substitutable. Replacing the orchestrator means rewiring the workflow, including remapping the topology, retraining the policy, reestablishing every integration, and re-accumulating the operational memory. The switching cost is categorical.

Value is multiplicative, not additive. A single agent coordinating three tools yields linear value: the sum of the tasks it automates. An orchestrator coordinating five specialist agents, each coordinating their own tools, produces nonlinear value. The orchestrator can redesign the workflow, parallelizing steps, eliminating redundancies, and catching errors across domains that no individual specialist would detect. The orchestrator is not a participant in the workflow. It is the workflow.

In human organizations, we call this management. In multi-agent systems, we call it orchestration. Value accrues to the Orchestration Layer rather than being evenly distributed among the agents beneath it. Not because someone has to be in charge, but because the topology in which centralized coordination wins is the only topology in which the workflow works at all. The entity at the center of that topology captures the surplus that coherence creates.

Let’s follow the money through an actual enterprise workflow

A commercial insurance submission arrives. Under the old architecture, it enters a human chain: intake, review, coding, investigation, pricing, approval, and documentation. Each human is an expert. Each has trained for years. Together, they process the claim in days, sometimes weeks, at costs measured in hundreds of dollars per claim. The Orchestration Layer lived in a manager’s brain and a project management tool. The bottleneck was human capacity for coordination.

Under agentic orchestration, the same workflow transforms. The submission arrives at a master agent. The master agent does not “process a document” as legacy software does. The master agent interprets the goal: evaluate this risk and produce a quote that aligns with our growth and risk constraints. It inspects the case. It decides which specialized agents to invoke: extraction, enrichment, risk modeling, pricing, or compliance validation. It sets their inputs. It sequences their work. It validates their outputs. It determines when the goal is achieved.

The specialist agents are valuable. They have real reasoning capacity. But they are no longer where the workflow is decided. Their job is to respond to the plan. Not to define it. The control plane has moved. So has the economics.

The question every company in the agentic economy must answer is not rhetorical. It is operational.

Are you the captain? Or are you the crew?

The Programmable Firm

Carnegie Mellon and Stanford University research teams released a landmark study in October 2025 called “How Do AI Agents Do Human Work?” that provided an empirical foundation for the Agentic Era. The teams compared how humans and AI agents complete identical work across data analysis, engineering, computation, writing, and design. They concluded that for tasks defined as “readily programmable,” AI agents will complete the work 88.3% faster at 90.4-96.2% lower cost. These figures represent benchmark environments rather than fully operational enterprise deployments.

The key finding is not that work is programmable end to end. It is more consequential: programming is embedded in 82.5% of knowledge-work occupations. In practice, this means a large share of enterprise workflows contain programmable components, even if they are not fully automatable.

Programmability is the unlock. A programmable workflow can be agentized. A workflow that can be agentized can be orchestrated. A workflow that can be orchestrated can be sold as an outcome rather than licensed as a tool. The economic category changes not because the AI is better, but because the work itself has always been more structured than the software market has reflected.

The gate is reliability. Every enterprise workflow has a reliability requirement: a level of accuracy and completeness below which human oversight remains necessary. When an AI system handles 40% or 50% of cases correctly, it accelerates human workers but cannot replace them. Every output still needs review. The human chain stays intact. The company has purchased a productivity tool.

When an AI system handles 85% or 90% of cases correctly, the economics change categorically. The workflow runs autonomously. Humans handle only the exceptions, the fraction of cases the system flags for review. The company has not purchased a productivity tool. It has purchased an autonomous function. The difference between these two states is not a percentage improvement. It is a reclassification.

In the first state, where AI is a productivity tool, the vendor sells software licenses. The buyer pays $50-$200 per seat per month from the IT budget. The vendor competes on features against other software vendors. The human labor cost remains on the books. This is SaaS economics.

In the second state, where AI becomes an autonomous function, the vendor sells an outcome. The buyer pays per claim processed, per credit analysis completed, or per supply chain route optimized. The budget is no longer IT spend. It is labor spend: the fully loaded cost of the knowledge workers the autonomous function replaces.

A single insurance underwriter costs $8,000-$15,000 per month in salary, benefits, training, and management overhead. A credit analyst costs similar. The vendor prices based on that labor cost, not on competing software licenses.

Apply this lens to a large institution like JPMorgan. With approximately $50 billion in annual labor costs and assuming ~30% of tasks fall into the “readily programmable” category, agents begin to meaningfully reshape the cost structure. However, because these tasks still require verification, exception handling, and integration into broader workflows, the immediate impact is not full replacement but partial automation and labor reallocation. Even under conservative assumptions, this translates into multi-billion-dollar efficiency gains, not through elimination of labor, but through compression of the programmable portions of work. That pool does not sit in the IT line. It sits in headcount.

The addressable market shifts accordingly. Enterprises typically spend 3-5% of revenue on IT, but 25-40% on knowledge work. The critical transition is not from software to AI, but from tools to workflow execution. Once an orchestrator reliably automates the programmable segments of work while coordinating human oversight where needed, it begins to tap into the much larger pool of labor spend. The prize is therefore not to replace SaaS but to expand beyond it by capturing portions of the operating budget that have historically been reserved for human execution. In that sense, the orchestrator that crosses the reliability threshold does not compete within the software market but instead redefines the market boundary, expanding it by a factor of 5 to 10.

The orchestrator creates this transition. The specialist agents, no matter how capable individually, do not. A pricing agent that achieves extraordinary accuracy still operates within the first state. It makes the human underwriter faster. It is a tool.

An orchestrator that coordinates pricing, risk, compliance, and documentation agents into a coherent autonomous workflow delivers the second state. It replaces the underwriting function. It delivers an outcome. The outcome emerges only from the coordinated whole. Only the orchestrator delivers it.

Bypass Risk

Orchestration Economics does not guarantee permanent advantage. The same structural logic that creates the orchestrator’s position also defines the conditions under which that position can be undermined.

Bypass risk is the possibility that a new entrant positions itself between the user’s intent and the existing orchestrator. In doing so, it captures the Orchestration Layer from above. Protocols such as MCP standardized how agents connect to external systems and to each other. A2A standardizes agent-to-agent communication. ANP sits atop both as a discovery layer.

Historically, APIs standardized software connectivity. In the agentic economy, protocols standardize agent connectivity. The strategic implication is profound: once interfaces become standardized, defensibility migrates away from integration and toward intent, context, and workflow ownership.

Standardization is powerful for interoperability. It also dissolves the integration friction that protected many orchestration positions. The company that controlled its workflow through the sheer difficulty of connecting systems wakes up one morning to find that the friction is gone. Protocols dissolved it. Orchestration Economics is therefore not a simple claim that the orchestrator always wins. It is a framework that says the orchestrator wins unless it gets bypassed.

The orchestrator’s advantage is structural, but its defensibility is conditional. It depends on building moats that protocols cannot standardize away and that competitors cannot replicate. Which moats? Under what conditions does an orchestration position become durable rather than temporary? What separates the orchestrators who compound their advantage from those who lose it to the next layer above?

The Three Laws

Knowing that orchestrators capture value does not tell us which positions are defensible. The orchestrator’s advantage is structural. Its defensibility is conditional. It depends on building moats that protocols cannot standardize away and that competitors cannot replicate.

Three conditions determine whether an orchestration position is durable. Each is necessary. None is sufficient alone. The entity that satisfies all three simultaneously holds a position that compounds with use, resists bypass, and captures the surplus of the agentic transition. The entity that satisfies one or two, even brilliantly, occupies a layer that the orchestration economy will eventually treat as crew rather than captain.

Chapter 8 develops the First Law: Proximity to Intent Determines Value Capture.

Chapter 9 develops the Second Law: Context Builds Moats.

Chapter 10 develops the Third Law: Workflow Intelligence Secures Control.

Once we define the Three Laws, we will apply them to Palantir.

The views and opinions expressed in this publication are those of the author alone and are based on publicly available information. The expressed views and opinions do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Certain companies referenced may be current or former clients of, or counterparties to, the author or affiliated entities; such relationships will be disclosed where applicable. Past performance is not indicative of future results. To the fullest extent permitted by applicable law, the author does not accept any liability for any loss or damage arising from reliance on this content. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

In three days, the market priced SpaceX at $1.75 trillion around a sub-frontier model, Anthropic shipped a real capability jump in Claude Fable, and the government forced Fable offline. The hot takes agreed that this meant the frontier is scarce, so open sources win. But this misreads commoditization. Intelligence commoditizes as a process, and the relationship between open source and frontier is symbiotic, rather than strictly oppositional. The frontier leads by a few months. That lead diffuses outward into the enterprise and into cheaper models as they catch up. The Orchestration Economy is predicated on the idea that diffusion never stops. Fable’s shutdown is the first time it stopped on command. When that happens, it has implications for the frontier, open weights, and orchestration. That becomes everyone’s problem.

Subscribe now

Read more

This is the latest excerpt from AGNT: The Orchestration Economics Manifesto - An Investment Framework for the Agentic Era. Each Thursday, I explore a major theme of the Manifesto and unpack the frameworks, adding extra context with more recent developments. Note: The figures and sequential references are taken directly from the larger Manifesto.

Subscribe now

The agentic economy is a software phenomenon built on a hardware foundation. That foundation is being constructed on a scale without precedent in the history of private capital allocation. But is it being constructed correctly?

When I first published this Manifesto in late April, the projected $650 billion in AI infrastructure investment in 2026 already represented one of the largest coordinated private investments in peacetime history. Since then, those numbers have continued to soar.

In May, the four largest hyperscalers reported quarterly earnings within hours of each other and raised 2026 capital spending guidance, committing between $650 and $700 billion to AI infrastructure for the year. Goldman Sachs raised its Capex spending forecast for the Big Four Hyperscalers from $4.5 trillion to $5.3 trillion through 2030. Overall, it projects $7.6 trillion of Capex spending between 2026 and 2031 across compute, data centers, and power.

The problem remains, however, that much of this capital is also being committed under assumptions that are already shifting.

The capital was designed for training. The agentic economy runs on inference. The cloud was designed for stateless applications. Agentic workloads are stateful, persistent, and coordination^-intensive. The power grid was designed for variable demand. Inference generates a continuous baseload.

The infrastructure does not need to stop being built. It needs to change what it is becoming. This chapter traces that metamorphosis, from the capital flowing in, through the balance sheets absorbing it, to the financing structures underwriting it, to the inflection point where the workload changes, to the efficiency gains that determine the unit economics, to the energy constraint that governs the speed, and finally to the architectural evolution of the cloud itself. Together, they constitute the infrastructure question on which the AGNT thesis ultimately depends.

Share

The Buildout

The numbers announced during Big Tech earnings calls throughout 2025 arguably constituted one of the largest coordinated private infrastructure investments in history. Bloomberg estimated that the four largest hyperscalers would spend ~$650 billion in 2026. Goldman Sachs projected cumulative hyperscaler Capex of $1.4 trillion from 2025 to 2027. That is close to three times the $477 billion spent from 2022 to 2024. Leading hyperscalers have emphasized that they are primarily constrained by power and capacity, not by demand, repeatedly flagging supply bottlenecks on recent earnings calls.

Microsoft carried an $80 billion backlog of unfulfillable Azure orders due to power constraints as of January 28, 2026 (Microsoft Q2 2026 Earnings). Alphabet’s cloud backlog surged to $240 billion at year-end 2025. Amazon said new capacity “sells out immediately”.

Yet investors punished these announcements. Amazon fell 8%. Microsoft dropped ~10%. The four companies collectively shed over one trillion in market value after revealing their spending plans. By February 2026, a Bank of America survey showed overinvestment concerns at an all-time high, roughly 35% warning of excess, the highest reading in twenty years.

Is this Capex buying what the agentic economy will require?

The Balance Sheet Transformation

The balance sheets of the world’s most profitable technology companies underwent a structural transformation beginning in 2025. For two decades, the defining characteristic of large-cap technology was capital efficiency. It was driven by asset-light models generating extraordinary returns on modest physical investment.

That era ended when the AI buildout began. The more than fourfold increase in capital intensity from 2015 to 2026, the dollars of capital required to produce one dollar of sales, tells the story.

Cloud computing infrastructure from 2015 supported predictable workloads with well-understood utilization patterns and multi-year lifecycles. AI infrastructure serves workloads with radically different characteristics: shorter useful lives, uncertain utilization, and a pace of obsolescence that makes traditional depreciation assumptions unreliable.

The depreciation debate alone reveals the depth of uncertainty. Amazon shortened the useful lives of a subset of its servers and networking equipment from six to five years in January 2025. Microsoft extended expected asset life in FY 2023, adding $3.7 billion in operating income from the change alone. These are not accounting technicalities. They are competing theories about how long AI hardware will retain value. The answers differ by billions of dollars.

Yet, the demand for NVIDIA H100’s has never been higher, suggesting longer utility than GAAP rules. When the industry cannot agree on whether its core infrastructure lasts one year or six, every capital allocation decision rests on a foundation of sand.

The free cash flow consequences are direct. Bank of America projected that Amazon’s free cash flow would turn negative at approximately $28 billion in 2026. Amazon filed with the SEC^, indicating it may need to tap equity and debt markets. Such a statement would have been unthinkable for a company generating $638 billion in annual revenue in 2024 (and $717 billion in 2025). Pivotal Research projected that Alphabet’s FCF would plummet approximately 90% in 2026, from $73.3 billion to approximately $8.2 billion, as $175-185 billion in Capex overwhelms operating cash flow. Barclays projected Meta’s FCF to drop approximately 90% in 2026, then turn negative in 2027 and 2028.

Step back and absorb the aggregate picture. Bank of America estimates that AI Capex now accounts for 94% of hyperscaler operating cash flows after dividends and buybacks. That figure was 76% in 2024. In the span of a single year, Capex has gone from absorbing three-quarters to nearly all of the available post-distribution cash flow, substantially compressing the cushion and forcing record debt issuance. The margin for error has dramatically narrowed.

The weight of physical assets on these balance sheets means that even the most profitable technology companies in history are subject to the same economic physics that govern utilities, railroads, and industrial conglomerates. Returns on invested capital matter more than revenue growth. The companies that built their fortunes on capital-light software economics are becoming capital-intensive infrastructure operators.

The financial frameworks applied to them have not caught up.

The Debt Machine

The Big Five (Amazon, Alphabet, Meta, Microsoft, and Oracle) raised ~$121 billion in bonds in 2025 alone: Amazon $15 billion, Alphabet $25 billion, Meta $30 billion (the largest-ever individual non-M&A high-grade bond sale), and Oracle $18 billion. That’s more than three times the average of the prior nine years. Alphabet’s long-term debt quadrupled in 2025 to $46.5 billion.

Oracle, whose fiscal year 2026 ended in May, disclosed that it had raised $43 billion in debt financing and $5 billion in equity financing. The company said it plans to raise $40 billion in debt and equity financing in FY 2027.

Companies that spent decades accumulating fortress balance sheets are now leveraging them at a blistering pace.

Morgan Stanley had earlier estimated that hyperscalers will borrow ~$400 billion in 2026, double the $165 billion borrowed in 2025. On June 10, Morgan Stanley predicted that AI-related global debt issuance would double to about $570 ​billion in 2026. JPMorgan projects $1.5 trillion in AI-related bond issuance over the next five years.

The clearest expression of that thesis arrived in February 2026, when Alphabet issued a 100-year bond. This was the first century bond from a tech company since Motorola in 1997. The sterling-denominated tranche, part of a $31.5 billion multi-currency debt raise completed in under 24 hours, was oversubscribed by nearly 10 times. Pension funds and life insurers, institutions that match liabilities measured over generations, bought the paper at a 6.05% yield, treating Google's parent as sovereign-grade infrastructure. Alphabet sits on c.$127 billion in cash (cash, plus equivalents, plus marketable securities) and chose to borrow anyway because even that war chest looks modest against $175 – $185 billion in planned 2026 Capex.

The century bond is a statement that buyers believe this buildout is not a cyclical phase but a permanent transformation of the economy’s substrate. The last tech companies to issue century bonds were IBM in 1996 and Motorola in 1997. They did so at the zenith of their market dominance. Both subsequently declined.

At the DealBook Summit in December 2025, Anthropic CEO Dario Amodei described the high-wire financial planning. Established scaling laws, he said, are here to stay. Models get better at every task as you put more data and compute into them. But the timing of economic value remains uncertain, creating an inherent risk of under-reacting or over-extension.

Spend too little, and a company like Anthropic or OpenAI risks being unable to serve customers in two or three years. Overspend, and companies face revenue shortfalls that in extreme cases could threaten bankruptcy or tempt them to adopt what Amodei called a “YOLO” mindset.

The Fragility Beneath the Conviction

The hyperscalers can service this debt. They have investment-grade ratings, massive revenue bases, and the operating cash flow, though diminished, to cover even aggressive borrowing. The fortress is leveraged, but it remains a fortress.

The risk lies elsewhere, in OpenAI deploys the chips at CoreWeave, which has grown so quickly and so intricately that its systemic properties deserve the same scrutiny applied to any previous era of financial innovation. At the center sits a pattern of circular vendor financing.

The architecture is straightforward to describe and difficult to contain. NVIDIA invests in OpenAI. OpenAI uses the capital to purchase NVIDIA chips. NVIDIA records the transaction as revenue (subject to LOI stage; not yet binding). OpenAI deploys the chips at CoreWeave. CoreWeave secured term-loan facilities collateralized by GPUs. Institutional investors purchase the securities as infrastructure debt, providing the capital that flows back to the beginning of the cycle. The same dollar is counted as demand by the chip maker, revenue by the neocloud, and collateral by the lender.

The pattern is not novel. Classic vendor financing catalyzed previous bubbles. The 2000s telecoms collapse followed this structure, with equipment makers extending credit to customers, inflating revenues while disguising demand weakness. When customers could not pay, receivables evaporated, and cascading failures followed.

What is novel is the concentration. Telecoms vendor financing funded “the internet” broadly across hundreds of carriers and thousands of build-outs. Today’s circular deals are largely concentrated around a single company. OpenAI sits at the nexus of NVIDIA’s investment, Microsoft’s partnership, Oracle’s data center contracts, AMD’s equity-for-chips arrangement, and CoreWeave’s largest customer relationship.

The greatest risk is not that AI fails. It is that the technology succeeds while the single company around which the financial architecture has been constructed fails to build a defensible moat, triggering contagion that strands capital and stalls the buildout before the transformation completes.

Because while this concentration of relationships presents a sizable risk, it is also far from the limit. The boundaries extend far beyond this inner circle and continue to grow. Data center-related securitization via CMBS and ABS products is expected to reach $30-40 billion annually in 2026 and 2027.

Private credit finances approximately $50 billion in AI infrastructure quarterly. The instruments are growing more complex. For GPU-backed instruments, collateral depreciation is outpacing debt amortization, even if the broader data center ABS/CMBS market is not similarly exposed (yet). And the whole structure is more concentrated around fewer counterparties than any comparable financing architecture in recent financial history.

You can be bullish on AI and bearish on this financial architecture. They are separable propositions. The technology may well deliver everything this manifesto argues it will. The question is whether the financing structure survives long enough to support the transition, or becomes the mechanism by which a sound technological thesis precipitates a financial crisis before the economics have time to prove themselves.

The financing architecture will be tested by what comes next. Because the infrastructure it underwrites was designed for a workload that is already giving way to something fundamentally different.

Two Tales of Compute

The vast majority of the capital committed in 2025 and 2026 was allocated on the assumption that the primary demand for compute would come from training frontier models.

Training was the activity that the industry understood when the investment decisions were made. This means massive, centralized GPU clusters running synchronous workloads for weeks, requiring homogeneous hardware, high-bandwidth interconnects, and cutting-edge silicon. Each generation of models requires roughly ten times the compute of its predecessor. Frontier AI training costs escalated from over $100 million for GPT-4 (2023-2024) to approximately $490 million for xAI’s Grok 4 (mid-2025), with projections exceeding $1 billion by 2027.

First, let us examine the tale of training compute. The hardware characteristics are distinctive. Training requires homogeneous clusters of cutting-edge GPUs with high-bandwidth interconnects, where a single GPU failure can corrupt an entire run that takes weeks of computation and costs millions of dollars. You cannot mix GPU generations or repurpose last-generation training infrastructure for new workloads without massive retrofitting.

Leading-edge racks consume 130-250 kilowatts (as of 2025), with projections reaching 900 kilowatts by 2027. Frontier training performance advantages degrade within 18-24 months, but the hardware itself retains economic utility for more than five years, typically migrating to inference or lower-tier workloads. This creates an oligopolistic market. The number of organizations training frontier models globally is perhaps fifteen to twenty:

Anthropic’s arrangement with AWS provides co-engineering of custom Trainium chips that reportedly reduce training costs by 40% compared to GPUs.

Google’s TPU advantage allows training at dramatically lower costs than GPU-based alternatives.

xAI built its Colossus cluster from site preparation to full operation in 122 days.

Access to massive, synchronized GPU clusters that require +250-megawatt data centers and billions in capital now determines who can compete at the frontier. Training compute is too strategic to outsource. Microsoft, Google, and Amazon have all reached the same conclusion.

Training is necessary. Without it, the models that power the Agentic Era do not exist. But training is not where this manifesto's thesis lives. Training is one-time Capex, concentrated among a handful of players, and depreciates in 18 months. It is the cost of building the engine.

Now, let us look at the tale of inference compute. Every query processed, every agent action executed, every token generated represents ongoing expenditure that compounds relentlessly. A single ChatGPT query costs approximately $0.003 in compute. That is seemingly negligible until multiplied by 800 million weekly users^, each generating multiple queries.

At five queries per user weekly, that is four billion queries, $12 million in weekly compute, $624 million annually for inference alone. Over a model’s operational lifetime, inference costs routinely exceed training costs by factors of three to fifteen. At CES in January 2026, Lenovo’s chief executive officer predicted that 80% of AI compute will be inference and only 20% training. As of March 2026, inference already accounts for over 60% of AI workloads and is projected to reach 70-85% by 2028-2030.

The economic characteristics diverge at every level. Training hardware depreciates in eighteen months. Inference hardware maintains utility for four to five years. Training serves a concentrated customer base for frontier labs. Inference serves a broadening market of millions of enterprises and consumers. Training requires the latest GPU architecture. Inference benefits from specialized silicon optimized for throughput and cost per token rather than peak performance.

Training is the arms race. Inference is the economy.

Inference Efficiency

At GTC 2026, NVIDIA CEO Jensen Huang completed a pivot that had been in the making for 18 months. His keynote focused on inference, agents, and AI factories. Data centers are factories, Huang argued. Their product is the token. Their efficiency determines the unit economics of everything built above them. The governing metric has shifted from training FLOPS to tokens produced per watt.

This pivot culminated in the introduction of NVIDIA’s Vera Rubin platform, which promises to deliver 10x the inference performance per watt compared to its Blackwell architecture. The pivot was reinforced by NVIDIA’s December 2025 deal with chip startup Groq. The deal was reported to be a $20 billion non-exclusive licensing agreement for Groq’s LPU inference technology, paired with the hiring of CEO Jonathan Ross and the bulk of Groq’s engineering team.

Groq remains operationally independent. At GTC 2026, Huang showed how the two architectures would work together: Vera Rubin handling the prefill phase of inference, and the new Groq 3 LPX rack, housing 256 LPUs, handling the decode phase where tokens are generated. The Groq integration recalled NVIDIA’s “Mellanox moment” several years ago^, when it acquired the company for its networking architecture (InfiniBand, Spectrum-X Ethernet), which became cornerstones of AI cluster design. Groq’s architecture is optimized for inference, adding deterministic low-latency decoding using massive on-chip SRAM at 80 TB/s internal bandwidth, achieving 500+ tokens per second on Llama 3 70B at up to 10x the energy efficiency of GPUs.

On the silicon side, the inference landscape has fragmented in ways training never did: AWS Inferentia claims significantly better price-performance, Google’s TPUv5e offers strong cost-effective efficiency gains, AMD and a generation of startups are all targeting the inference economics that determine AI’s unit costs.

The algorithmic gains compound on top of the silicon. Mixture-of-experts architectures activate only a fraction of total parameters per query. Qwen3-Next, for example, activates 3 billion of 80 billion through ultra-sparse expert routing, achieving 10x faster inference on long contexts. Dynamic batching pushes GPU utilization from 30% to over 80%. Speculative decoding doubles token-generation speed at a fixed cost. These represent order-of-magnitude gains that compound with one another and with the silicon improvements beneath them, producing a 5-10x annual price-performance trajectory.

Stanford and Together AI researchers quantified where this trajectory leads. Their Intelligence Per Watt metric improved by a factor of 5.3 between 2023 and 2025. Local models on consumer-grade hardware answered 88.7% of single-turn reasoning queries correctly. The share of queries that edge devices can handle rose from 23.2% to 71.3%. Inference is beginning to migrate from data centers to the edge, and this trend will accelerate as devices become more capable.

The Jevons paradox operates here with particular force. This refers to the principle that increased resource efficiency drives higher total consumption. In this case, efficiency gains in inference will expand the frontier of viable workflows, which expands total consumption even as per-unit costs fall. A workflow unprofitable at $50 in agent compute becomes routine at $5. A category of work inconceivable at one price point becomes obvious at a lower one. The efficiency gains feed the demand they were designed to serve.

This has direct implications for the AGNT thesis. The orchestrator’s primary variable cost is the cost of inference tokens. Every improvement in efficiency expands the addressable market. The orchestrator that routes tasks intelligently across capability tiers (commodity models for extraction, frontier models for complex reasoning, edge devices for latency-sensitive work) achieves structurally better margins.

Whether the agentic economy becomes the deflationary force described in Part V depends on whether this efficiency trajectory continues to compound. If it does, the surplus available to orchestrators grows with every cycle. If it stalls, the transition slows. Not for want of intelligence, but because the economics simply do not work.

The Energy Constraint

Beneath the capital, the debt, the silicon, and the algorithms lies a constraint more fundamental and less tractable than any of them: Power.

The scale of the demand is unprecedented. Huang’s $1 trillion demand outlook through 2027 implies not merely more chips and racks but more electricity consumed continuously at an industrial scale. A single Vera Rubin NVL72 rack draws power at levels that would have been classified as a small industrial facility a decade ago. Leading-edge AI racks consume 130-250 kilowatts today, with projections reaching 900 kilowatts by 2027. Each hyperscaler is pursuing a one-gigawatt data center campus, a scale that draws as much electricity as a mid-sized city.

The workload itself has changed shape. A chat query resolves in 500 milliseconds. An agentic task, such as a coding audit, a research synthesis, or a multi-system workflow, takes 10 minutes to 14+ hours and runs in parallel with thousands of siblings inside a single campus. The load profile of an AI data center has converged toward that of an aluminum smelter: continuous, high-wattage, near-100% utilization, 24/7. The bursty inference workload grid planners that were modeled against have been removed. The new profile is baseload.

In addition, agentic compute has a unique memory-power overhead. To maintain state over a 14-hour + task, agents must keep massive context windows active. In 2026, hyperscalers are offloading KV caches to SSD storage to manage the memory bottleneck, but the energy required for constant data shuffling remains a high hidden cost. Anthropic has published research on “context engineering”. This refers to compaction, structured note-taking, and multi-agent architectures designed to address “context rot” in long-horizon agents. These techniques reduce the memory-power overhead of maintaining state across extended tasks and indirectly ease the underlying power constraint.

The power infrastructure to support these facilities does not exist in most locations where they are being planned. Grid interconnection timelines run three to five years in the United States, longer in Europe. Permitting for new transmission lines routinely takes a decade.

The US is on track to add a record 86 GW of utility-scale generation in 2026: 51% solar, 28% batteries, 14% wind. Useful firm power near data-center hubs is a fraction of the headline figure. In PJM, the single most important AI interconnection, utility large-load commitments now exceed accredited generation capacity by a factor of 3. China, in contrast, added 434 GW in 2025 alone and is co-locating compute with generation in the renewables-rich West. This is not a cyclical shortfall; the US closes with Capex. Rather, it is a profound infrastructure divergence, one that could constitute a Single Point of Failure of the Agentic Era, i.e., the specific vulnerability that, if left unresolved, constrains the entire transition regardless of how fast models improve or how much capital flows into AI. In the framework we introduced in 2025, an SPOF is the component of a system whose failure compromises the entire structure.

Nuclear is, as of 2026, the only carbon-free source with the density for AI-scale loads, and the scramble for it since 2024 is no longer directional. It is the infrastructure. Microsoft underwrote the restart of Three Mile Island via a 20-year PPA with Constellation (835 MW, 2027-2028). Amazon bought Talen’s Susquehanna campus and is developing SMRs with X-Energy and Dominion. Google signed with Kairos (500 MW SMR fleet), Fervo Geothermal, Oklo, and Vistra, making it the largest corporate purchaser of nuclear energy in US history. Sam Altman backed Oklo (2.6% stake, former chair, co-founder of the SPAC that took Oklo public). xAI stood up 200+ MW of on-site gas turbines in Memphis in under a year. Stargate Abilene is architected as a 1+ GW behind-the-meter campus. Together, these constitute a coordinated admission that the public grid cannot deliver what agentic compute needs within the required timeline. Hyperscalers are no longer interconnecting to the grid: they are becoming utilities.

The constraint sharpens as the workload shifts from training to inference. Training runs are episodic. A frontier model trains for weeks in a concentrated burst, then stops. Inference is continuous. Every agent action, every orchestrated workflow, every token consumes power indefinitely, at a load that scales with agent populations.

As the mix shifts, the demand profile moves from periodic spikes to sustained baseload: precisely the pattern that stresses grids most acutely, because it requires always-on generation rather than peak-shaving.

The efficiency gains partially offset this. As noted before, Jevons in this respect operates at an industrial scale: cheaper inference makes more workflows viable, expanding total demand even as per-unit consumption falls. Hyperscalers that secure power early through long-term agreements, on-site generation, or strategic locations near abundant clean energy gain compounding advantages. The value of a data center is increasingly determined not by the GPUs it houses, but by the megawatts it can draw.

For the AGNT thesis, energy introduces a temporal variable into every projection. The Inference Economy expands at a rate bounded by the availability of power, regardless of demand or capital. The orchestrators of Parts III to V will find their ability to scale agentic workflows ultimately limited by whether the infrastructure they depend on has the electricity to operate.

The direction is not in question. The speed is. Watts, not weights, now govern the pace of the transition. The pace, as the Two Scenarios in Chapter 18 will argue, is what separates deflationary growth from demand destruction.

The Agentic Cloud

There is a final condition that completes the picture. It is architectural rather than financial or physical. The cloud was built for a different era. The SaaS-era cloud was designed for stateless applications, ephemeral compute, hub-and-spoke networking, and consumption-based pricing. An enterprise spins up a virtual machine, runs a workload, and shuts it down. Two decades of perfecting that pattern. It is the wrong pattern for what comes next.

The agentic systems of Chapter 5 generate workloads that differ in every dimension. Every MCP connection, every A2A coordination, every persistent memory retrieval, every multi-step workflow is an inference operation. An agent booking a flight generates 50,000 tokens across research, comparison, booking, and confirmation. This compares to just 500 for a chat query. Agentic workloads multiply token consumption 10x to 100x.

But the difference is not merely volume. Agents are stateful. They maintain persistent memory and workflow state across sessions, requiring considerably more memory than traditional workloads. The three-tier memory architecture of Chapter 5 requires infrastructure that preserves state, not infrastructure designed to forget:

Agents coordinate laterally. They communicate peer-to-peer through MCP and A2A, requiring sub-100ms latency and dedicated bandwidth between coordinating agents. The topology must evolve from a centralized switchboard toward a mesh.

Agents generate explosive storage demands. Vector databases typically experience 10x data expansion, with total storage reaching 30x the original data after indexing and context persistence.

Agents require a different economic model. Pricing must be tied to task completion and outcomes rather than compute hours consumed, which demands infrastructure instrumented for metrics that traditional cloud monitoring was never designed to capture.

These requirements point toward a new infrastructure category: the agentic cloud. Three categories of providers are positioned to build it.

The hyperscalers hold the deepest structural advantages: custom silicon, enterprise relationships built over decades, and captive internal demand for inference. Google’s Search, YouTube, Gmail, and Maps are continuously consuming inference, providing a substantial utilization floor regardless of external demand. Their challenge is that hundreds of billions in existing infrastructure were designed for the SaaS era. The organizational complexity of serving legacy and agentic workloads simultaneously is real, and scale alone does not resolve it.

The neoclouds proved something important: that AI workloads demand fundamentally different compute from general-purpose clouds. CoreWeave and other AI-specialist clouds have demonstrated MFU improvements of 15-30% over hyperscaler baselines (e.g., CoreWeave’s 49.2% MFU on H100 training runs vs MosaicML’s 41.85% baseline, per CoreWeave’s published benchmarks). Industry training MFU typically sits in the 35-55% range, depending on hardware generation, software stack, and workload. But the neocloud window is narrowing. Their core design is stateless, single-tenant, and hub-and-spoke. This served training-centric workloads well. It is less suited to what agentic systems demand.

The emergent agentic clouds are purpose-built. Daytona Cloud achieves sub-90ms sandbox creation for stateful agent workflows. Others target observability for multi-agent systems, optimization for coordination-intensive workloads, and outcome-based resource allocation. They carry no architectural debt. They lack scale. They are building for the incoming workload.

The model providers see the same opening. Anthropic’s Claude Managed Agents, launched in April 2026, offers the full production stack for agentic workloads as a managed service: sandboxed execution, state persistence, tool orchestration, error recovery, and multi-agent coordination. It runs on existing cloud infrastructure, but the developer never touches it. Anthropic abstracts the cloud into a substrate and bills by the session-hour.

The orchestrator’s variable cost is the cost of inference tokens. Hardware efficiency sets the floor for the cost of producing a token. The agentic cloud determines the cost of serving that token within a coordinated, stateful, production-grade workflow. The companies that optimize for both will set the economic foundation for everything built on top of them.

The Substrate Must Evolve

The AI infrastructure being built in 2026 is the largest private capital deployment in history. It is transforming Big Tech’s balance sheets from capital-light to capital-intensive. It is financed through structures of increasing ambition, from conventional bonds to century paper to securitized GPU assets. And it was committed under assumptions about training workloads that the migration toward inference is already reshaping.

The metamorphosis is substantial but navigable. Inference efficiency is compounding at a pace that makes the unit economics of the agentic economy more favorable with each quarter. The energy constraint governs speed, not direction. And the cloud is beginning, unevenly, incompletely, but perceptibly, to evolve toward the stateful, coordination-grade architecture that agentic workloads demand.

The intelligence is sufficient. The agents are in production. The infrastructure is scaling and starting to change shape. What determines whether specific companies, sectors, and workflows successfully cross is the subject of Part III.

The views and opinions expressed in this publication are those of the author alone and are based on publicly available information. The expressed views and opinions do not constitute investment advice, a solicitation, or a recommendation to buy or sell any security or financial instrument. The author may hold positions in the securities of companies mentioned. Certain companies referenced may be current or former clients of, or counterparties to, the author or affiliated entities; such relationships will be disclosed where applicable. Past performance is not indicative of future results. To the fullest extent permitted by applicable law, the author does not accept any liability for any loss or damage arising from reliance on this content. Readers should conduct their own independent due diligence and consult a qualified financial advisor before making any investment decision.

AGNT Podcast Ep. 10 with Gemma Allen & Raphaëlle d'Ornano

00:00 - Intro

00:01 - Exploring Space Ventures: An Introduction to AGNT and SpaceX’s IPO

05:04 - SpaceX’s Business Segments and Valuation

10:52 - AI in Public Markets

16:46 - Anthropic and Self-Improving AI Risks

19:51 - Navigating the Future: AI Innovations and Market Dynamics

AGNT Podcast Ep. 9 with Gemma Allen & Raphaëlle d'Ornano

00:00 - Intro

00:02 - Title: Exploring AGNT and Anthropic's Evolution

05:41 - Anthropic's Strategic Edge in Enterprise

10:29 - Addressing IT Budget Constraints

13:17 - Role of Tokenomics in AI

16:49 - Layered Structure of AI Systems

19:25 - Implications of Anthropic's Growth

23:03 - Reimagining Enterprise Software: Valuation, Reinvention, and the AI Paradigm Shift

SpaceX IPO S-1 Teardown

Durable Growth Moat™ Analysis

Why Enterprise AI Is an $800B Black Hole in the $1.75 Trillion Story

Decoding Discontinuity | June 2026

As the largest IPO in history approaches public markets, investors are being asked to underwrite three fundamentally different businesses: a dominant launch monopoly, a global connectivity compounder, and an AI platform pursuing one of the most ambitious enterprise narratives ever presented to public investors.

Our latest institutional teardown examines SpaceX through the lens of Orchestration Economics, Durable Growth Moats™, and the emerging structure of value capture in the Agentic Era.

The IPO narrative assumes that AI will be the largest driver of future value creation, with enterprise applications accounting for nearly 86% of the company’s projected AI opportunity. Yet beneath the scale of that ambition lies a critical question:

What structural position allows xAI to capture the enterprise market it is being valued to own?

This rep…

Read more