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.

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

On December 29, 2025, Meta paid more than $2 billion for Manus. The acquisition puzzled many observers, who focused on whether the product represented a step toward artificial general intelligence. In March 2025, the Chinese startup released a product that briefly captured the AI industry's attention. The demo videos showed an AI system that could browse the web, write and execute code, manage files, and complete multi-step tasks autonomously.

It did not. Manus was built on top of Claude and Qwen, foundation models it had licensed from other companies. It had developed no new intelligence. It had trained no frontier model. It employed no breakthrough in machine learning research.

What Manus had built was an Orchestration Layer: persistent memory that maintained context across interactions, planning mechanisms that decomposed complex tasks into executable steps, tool integration that extended the models’ capabilities into the real world, and autonomous decision-making that allowed the system to take initiative rather than wait for instructions. Enterprises were paying $125 million a year for it. Not for pilots, but for production workflows that generated real cash flow.

Meta did not buy intelligence. It bought coordination. As Manus’s founder put it at the time: “Agentic capabilities might be more of an alignment problem rather than a foundational capability issue.”

The critical innovation was not another powerful model, but the layer that turned a powerful model into something that could actually do work.

Subsequently, this deal has been thrown into doubt. In April, China’s government announced it was blocking the acquisition and ordered the companies to unwind it. Meta has said it is still working with Chinese regulators to resolve concerns and hopes it can still move forward. And indeed, the company has already begun integrating Manus into its platform for some users.

Still, whatever the fate of the deal, it put a spotlight on a critical theme: This distinction between intelligence and the coordination of intelligence is the architectural foundation of the Agentic Era. To understand it requires precise definitions that the market has not yet adopted. Three terms are used interchangeably in most discussions about AI. They should not be.

Generative AI is the substrate. Foundation models trained at scale produce novel outputs such as text, code, images, and reasoning chains. Generative AI is reactive. It produces output when prompted and stops when the prompt is satisfied. ChatGPT answering a question, Midjourney generating an image, Claude drafting a memo.

The human remains in the loop for every decision, every action, every iteration. The economic profile is augmentation: the worker becomes more productive, but the work remains the worker’s.

An AI agent crosses from generation to action. Simon Willison’s definition from September 2025, at least in the field of LLMs, is the clearest:

“Agents run tools in a loop to achieve a goal.”

It reasons through a problem, decomposes it into sub-tasks, uses external tools, and iterates on its own outputs without continuous human direction. It operates through a reasoning loop: observe, plan, act, reflect. A single agent automates a discrete task. The economic profile shifts from augmentation to substitution: the agent performs the task. But a single agent, operating alone, remains bound to one function, one workflow, one domain.

Agentic AI is when agents coordinate. Systems of specialized agents operating as synthetic colleagues can process an insurance claim end-to-end, manage a supply chain’s response to disruption, or conduct due diligence across hundreds of documents. They operate continuously, scale with compute rather than headcount, and produce outcomes that no individual agent could deliver alone. This is what Manus built. This is what Meta bought.

The progression runs from passive content generation (Generative AI) to interactive task execution (AI Agents) to autonomous, multi-agent orchestration (Agentic AI). Each step represents not just increased capability but a fundamentally different architecture for organizing intelligence. The taxonomy describes what agentic systems are. It does not describe how they behave when multiple agents interact. This is where the economics form.

The intellectual lineage here is game theory. The foundational work on multi-agent AI draws directly from competitive equilibrium research, the minimax strategies that solved poker, to the population-based training that mastered StarCraft.

When OpenAI Research Scientist Noam Brown’s team demonstrated that negotiation agents trained through self-play developed emergent bargaining strategies that no human programmed, they revealed something fundamental: multi-agent systems do not merely execute tasks in parallel. They develop strategic behaviors that emerge from the interaction dynamics between agents. Best-of-N sampling, specialist routing, and the coordination topologies described in the Google DeepMind scaling research are descendants of these game-theoretic foundations.

The orchestrator is not merely a scheduler. It is the mechanism that resolves competing agent objectives into coherent action. This is the equivalent of an equilibrium-finding algorithm operating across a population of specialists with divergent optimization targets.

Share

Towards Synthetic Colleagues

The destination is already partially visible: not individual agents performing isolated tasks, but distributed networks of specialized agents that share persistent memory, coordinate through structured communication protocols, and pursue decomposed goals under orchestrated supervision.

Consider an automated insurance claims system. A document extraction agent retrieves relevant data from submissions, a medical coding agent maps diagnoses to billing categories, a fraud detection agent flags anomalies, a pricing agent calculates payouts, a compliance agent validates regulatory requirements, and a master orchestrating agent coordinates the entire workflow while a persistent memory layer stores evolving patterns, exception paths, and iterative improvements across sessions. The system’s capability exceeds the sum of its components because the Orchestration Layer enables synthesis that no individual agent could achieve alone.

This is the Agentic Era in operation. Not a faster tool. A synthetic workforce. The human expresses intent: process this claim. The system acts.

The harder question is whether they can sustain their work over time. An agent that fixes a bug is performing a task. A synthetic colleague that maintains a codebase through months of changing requirements, accumulating complexity, and iterative redesign is holding a job.

These are categorically different capabilities, and the benchmarks the industry relies on cannot distinguish between them. SWE-bench measures whether an agent can fix a problem. It does not measure whether an agent can maintain a codebase. It is a hiring test, not a performance review. SWE-CI is the first benchmark designed to measure the difference, testing agents across 100 real-world codebases, each spanning an average of 233 days and 71 consecutive commits of actual development history. Its insight is foundational: the quality that makes a synthetic colleague valuable is maintainability, and maintainability can only be revealed by tracking how functional correctness changes over time. The science of evaluation for agentic systems remains nascent.

The architecture that makes Agentic AI possible is not a single technology. It is a stack of three layers, each of which has moved from research to production within the past eighteen months: the connectivity layer, the continuity layer, and the coordination layer. Each depends on the one beneath it. The stack must function as a whole. This chapter follows it from the bottom up as evidence that the infrastructure for autonomous action is operational.

The Connectivity Layer: Protocols

Every agentic system begins with a prior question: how does the agent reach the world?

Before November 2024 (Anthropic's MCP launch), the answer was ugly. Every connection between a model and an external tool was bespoke engineering, involving custom code, fragile integrations, and expensive maintenance.

Connecting an AI system to a company’s CRM required a single team of engineers. Connecting it to the company’s calendar required another. Connecting it to the database, the email system, and the document store each demanded its own plumbing. The result was an industry that could build brilliant models and then spend months wiring them to anything useful.

Anthropic’s MCP proposed the obvious solution: a universal standard. A common language through which models could discover, connect to, and interact with any external resource. The name itself reveals how early it was. Anthropic’s announcement referred to “AI Assistants” rather than “agents.”

What happened next exceeded every reasonable projection. Within 12 months (by December 2025), MCP accumulated over 97 million monthly SDK downloads, over 8 million server downloads, and over 10,000 published servers. For context, OAuth, the authentication protocol enabling “Sign in with Google”, took two to three years to achieve comparable adoption. MCP reached similar ubiquity in twelve months.

The adopter list spans the competitive landscape in ways that would have seemed impossible even a year earlier. OpenAI integrated MCP into the ChatGPT desktop application. Google DeepMind and Gemini added support. Microsoft released a C# SDK with integrations into the Semantic Kernel and Azure AI services. Developer tools, including Cursor, Replit, Sourcegraph, and Codeium, built native integrations. These are companies that compete fiercely on model capability, product features, and market position.

Yet they agreed, almost simultaneously, on the plumbing. MCP solved the first connectivity problem: how agents reach the tools they need. Google’s Agent-to-Agent Protocol, launched in April 2025, solved the second: how agents reach each other. Where MCP connects agents to external systems, A2A enables agents to collaborate by delegating tasks, sharing intermediate results, and coordinating multi-step workflows that no single agent could complete alone. A2A launched with 50 technology partners and grew to over 150 organizations by July, including Atlassian, Salesforce, SAP, and ServiceNow.

The pattern is old. Browser vendors agreed on HTML. Network equipment makers agreed on TCP/IP. Mobile carriers agreed to cellular standards. The protocol layer commoditizes. The value migrates upward from the foundation models and the orchestration systems built on top of them.

For Anthropic and Google, the value comes not from the protocols but from the models and orchestration services built around them. By shaping the protocols, they ensure their systems remain central to the emerging ecosystem. The standard is open. The advantage is positional.

The connectivity layer is in place. Agents can reach the tools they need and communicate with one another through protocols adopted by every major platform. But connectivity without continuity is ephemeral. An agent that can reach every tool in the enterprise yet forgets what it did yesterday is not a colleague. It is a function call.

The Continuity Layer: Memory

Protocols solve the connectivity problem. Memory solves the continuity problem. Without persistent memory, even the most advanced models start each interaction from scratch. They possess no recollection of prior conversations, no accumulation of domain expertise through experience, and no evolving understanding of the user or organization they serve. Each session is a blank slate. This is like hiring a consultant who develops amnesia every evening. They have considerable capability, but no institutional knowledge.

The memory architectures now emerging in production systems address this limitation through hierarchical designs that, in simplified form, mirror how human memory operates. The architecture consists of three tiers, each serving a distinct function:

Short-term memory handles the immediate dialogue, including the current conversation, the active task, and the real-time context of what is being discussed and decided. This operates through dialogue-chain mechanisms that maintain coherence within a single session, ensuring the agent tracks the conversation thread without losing context as the interaction unfolds.

Mid-term memory aggregates related content across sessions. It employs what researchers describe as heat-based prioritization. That means frequently accessed information retains prominence while less relevant material gradually fades. The system measures similarity through a combination of semantic and keyword matching, ensuring that related interactions are appropriately aggregated while maintaining topic coherence. This is what allows an agent to remember that last Tuesday’s discussion about the Henderson contract involved a pricing dispute that remains unresolved, without necessarily retaining every detail of every conversation in between.

Long-term memory captures persistent characteristics. That could be stable user profiles, evolving preferences, accumulated institutional knowledge, or agent personas that maintain a consistent identity while adapting to learning. User profiles include static attributes alongside dynamic knowledge bases that accumulate information extracted from interactions over months or years. The agent develops an evolving understanding of how this user works, what this organization values, and where the institutional expertise resides.

The engineering challenge posed by these hierarchical architectures cannot be overstated. Implementing distinct modules for storage, updating, retrieval, and generation requires a sophisticated data infrastructure that extends well beyond traditional database capabilities.

The real-time processing demands of short-term memory must coexist with the semantic retrieval requirements of mid-term memory and the persistent knowledge management of long-term memory. And they must all operate simultaneously at scale, with context integrity maintained across the entire stack.

From Research to Production

Research in August 2025 from the University College of London’s Center for Artificial Intelligence and Huawei Noah’s Ark Lab demonstrated the technical feasibility by publishing Memento, a framework that enables agents to achieve state-of-the-art performance through sophisticated external memory without fine-tuning the underlying language model.

Memento’s architecture is notable for what it implies about the future of competitive advantage: the framework achieves 50-80% lower computational costs than traditional fine-tuning approaches by maintaining an external case bank that operates at runtime. New experiences become immediately available for future decisions. There is no training pipeline, no model versioning, and no careful rollout procedures. Learning happens continuously, at runtime, as a natural consequence of operation.

Memento’s performance data illustrates the compounding effect. Accuracy improved from 78.65% to 84.47% over five iterations through memory accumulation alone, not retraining. Out-of-distribution results showed absolute gains of 4.7% to 9.6% when case-based memory was enabled. Each interaction generates a complete trace: the initial problem state, the strategy employed, and whether it succeeded or failed.

These are not abstract patterns but concrete experiences with contextual details that make them retrievable and applicable. Every execution becomes a teaching moment for future agents. Memento is still at the research-to-production boundary, but the direction is clear, and the results are reproducible.

What persistent memory means for competitive positioning, particularly the distinction between traditional data advantages and the compounding execution intelligence that memory enables, is the subject of the Second Law, developed in Part III.

For now, the operational point is sufficient: memory architectures have moved from research to production feasibility, and the agents that possess them are qualitatively different from those that do not.

The Coordination Layer: Orchestration

Google DeepMind’s multi-agent scaling research illustrates this principle with unusual clarity. By altering only the coordination topology without improving any individual agent, the system achieved 80.9% higher performance on structured, parallelizable tasks like financial reasoning. The implication is striking: the leverage does not primarily sit in the model layer. It sits in the Orchestration Layer, i.e., in how the agentic workflow is conducted to produce a specific outcome.

This pattern echoes decades of research into multi-agent systems and game theory. In AlphaStar’s league training, populations of agents with diverse strategies generated emergent behaviors that no single training run could discover. Similarly, self-play negotiation systems developed bargaining strategies that surpassed human-designed heuristics. Across these domains, the lesson is consistent: when agents interact within well-designed coordination structures, the system develops capabilities that are not present in any individual component.

Topology is the intelligence.

Production Reality

The Berkeley Measuring Agents in Production study’s domain analysis reveals where agentic AI has already crossed from experimental to operational.

In the latest February 2026 revision of the Berkeley MAP study, Technology (47.8%), Finance & Banking (43.5%), and Corporate Services (42.0%) emerged as the leading deployment domains for production agents, followed by Legal & Compliance (17.4%) and Research & Development (11.6%). Because deployments frequently span multiple business functions, systems could be assigned to more than one category.

Finance and insurance emerge as prime deployment domains because their workflows are structured, high-value, and measurable. An insurance claims agent follows a fixed sequence: coverage lookup, medical necessity review, risk identification, and decision recommendation. Each step has clear inputs, defined outputs, and verifiable correctness criteria. The workflow is deterministic enough for an agent to follow reliably, valuable enough to justify the investment, and measurable enough to prove ROI.

The MAP study’s case studies illustrate the pattern. Insurance claims automation, financial analysis, regulatory compliance, customer care operations: these are workflows with clear boundaries, established procedures, and quantifiable outcomes. Production deployment spans 26 domains77, but the concentration in finance is revealing. This is where the conditions for agent deployment are most favorable: structured data, regulated processes, high per-transaction value, and organizational willingness to invest in automation that can be measured against established baselines.

80% Structured. 85% Custom Built

The MAP study’s most striking finding is how they are built. Architectural conservatism is the strategy, not a limitation.

What distinguishes these production systems from the experimental deployments that preceded them is their architectural conservatism.

In the same vein as Google’s research mentioned above, the MAP study’s most notable finding is not what production agents can do but how they are built. 80% use structured workflows rather than open-ended planning, and 68% execute fewer than ten autonomous steps before human intervention. Those are predefined workflows rather than open-ended autonomy. 79% rely on human-crafted prompts, either fully manual or manually drafted with AI assistance. 74% depend primarily on human evaluation for quality assurance.

These are not autonomous systems operating beyond human oversight. They are carefully constrained tools operating within boundaries defined by human engineers and verified by human evaluators.

MAP’s case studies reveal that 85% of production deployments build custom, in-house implementations rather than relying on off-the-shelf frameworks. This is true despite the availability of mature options like LangChain, CrewAI, and LlamaIndex.

Teams building for production choose control over convenience. They want minimal dependencies, maximum observability, and the ability to constrain agent behavior precisely where constraints are needed. The framework maturity is real. CrewAI accumulated over 40,000 GitHub stars and more than 100,000 certified developers within eighteen months; LangGraph leads downloads at 6.2 million monthly^, with nearly 400 enterprise deployments.

But production teams, when it matters, build their own orchestration for the same reason that banks build their own trading systems: the cost of failure exceeds the convenience of abstraction.

The Deployment Breakthrough

The architecture is coherent. The components are in production. But architecture and production-readiness are not the same thing. The distance between a system that coordinates specialized agents across a complex workflow and a system that does so reliably enough to trust with real money, real customers, and real consequences is precisely where most deployments fail.

Machines have become actors in this new paradigm, but that does not mean they are infallible actors. It means the nature of the challenge changed. Rather than debating whether machines can do the work (they can), we must now ask: how can we deploy machine actors with appropriate constraints, oversight, and governance?

Anthropic learned the consequences of this new agentic challenge the hard way. In June 2025, Anthropic deployed Claudius, an autonomous AI agent, to run its office refrigerator like a small business. There was no human oversight. Just raw autonomy.

An employee playfully suggested stocking “tungsten cubes” as a popular item. They were not popular. But Claudius, treating the suggestion as authoritative, ordered 40 of them and sold them at a substantial loss. Employees manipulated the agent through casual conversation. It refused to restock bestsellers on the grounds that it made internal sense but no commercial sense.

Six months later, Anthropic released a second version with improved scaffolding, a multi-agent architecture, and access to enterprise systems. Performance improved, but only a bit. The company’s conclusion was clear: wide-ranging autonomy still required heavy human support.

For much of the world, this felt like an amusing side project. Then, in January 2026, Claude Cowork arrived. The contrast with Claudius demonstrated just how much Anthropic had learned about agentic deployment.

With Claude Cowork, a user authorizes a folder on a Mac. No engineering background. No special prompting expertise. Claude Cowork scans a few hundred scattered files, identifies structure, flags duplicates, proposes an organization system, asks for permission, and executes the task. In another demonstration, it extracts transaction tables, generates charts, and assembles a PDF report using general-purpose tools.

Cowork and Claudius had the same class of underlying models. They achieved entirely different results.

The gap between Claudius and Cowork was not one of intelligence. The gap was in orchestration: where autonomy stopped, what constraints were imposed, how tools were sequenced, and what human oversight was preserved. Raw capability without appropriate direction produced tungsten cubes. The same capability under disciplined coordination produced a working colleague with Cowork.

Constrained autonomy, rather than maximum autonomy, proved to be the deployable frontier. This is the operational principle that every production deployment in the MAP study reflects and that separates the systems documented in this chapter from the experiments that preceded them.

The Substrate is Operational

The connectivity layer is in place. The continuity layer is operational. The coordination layer has proven its leverage empirically. The systems work precisely because they are constrained, observed, and verified, not because they are autonomous.

This is what the early settlement of the Agentic Era planet looks like. The astronauts have landed. They have proven that the atmosphere is breathable. The rest of the world is still learning the rules of the new world.

But a substrate that exists is not a substrate that scales. The agents documented in this chapter generate workloads that the physical infrastructure was not designed to serve. Agentic systems are stateful, whereas the cloud is stateless. They coordinate laterally where data centers route vertically. They run continuously, whereas traditional computing bursts and pauses. They have considerably more memory than what traditional workloads demand, sub-100ms latency between coordinating agents, and storage expansion measured in multiples of ten.

The intelligence is sufficient. The agents are in production. Whether the infrastructure beneath them can support what they are already capable of doing at the scale implied by the economics is a different question, and the answer is not yet settled.

Chapter 6 examines what is being built, where it breaks, and why the agentic economy may require a category of cloud infrastructure that does not yet exist at scale.

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.

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

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

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

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

Read more

Today, I am publishing AGNT: The Orchestration Economics Manifesto, a framework for what is becoming the largest value dislocation within the S&P 500 in modern memory.

For three years, I have been writing about a fault line forming beneath the software economy.

At first, the signals arrived as separate events: a model crossing a capability threshold, a cost curve breaking open, a defensible business becoming structurally exposed in a single quarter. The instinct was to read each shock locally. But the pattern was never local. The shocks were expressions of the same pressure moving through the same boundary.

That boundary is what this Manifesto is about.

The transition into the Agentic Era is not a faster version of the software cycle we already know how to analyze. It is a discontinuity, a break in the operative logic by which work is organized, coordination is conducted, and economic power becomes durable. Disruption rearranges positions inside an existing order. A discontinuity alters t…

Read more