Part 3 in the series after The Hyperscaler Reckoning (5 May) and The Hyperscaler Dispersion (4 August).

TL;DR: Alphabet, Amazon, Meta, and Microsoft reported about $203 billion of Q3 2026 capex on company-specific measures. Only Microsoft’s operating cash flow exceeded capex. My central case keeps their combined free cash flow negative through 2029, with recovery in 2030 if capex growth slows. That’s two years later than the chart investors still carry. Why is my projection further out? Cloud growth is accelerating, the core businesses that pay for it are showing the first signs of AI substitution, and disclosed commitments from two laboratories, OpenAI and Anthropic, equal close to 40% of the cloud backlog. The AI buildout is funded. The cash payoff may wait until 2030.

Subscribe now

Methodology note: The cash-flow path in this essay is built from the companies’ own numbers: the third-quarter cash-flow statements, the 2026 capex guides, and what management said on the calls about 2027, with two growth assumptions of my own. This is not a model or a consensus. It is a set of high-level estimates, rounded to the nearest $10bn, meant to show the shape and the timing of the recovery rather than to forecast a year. Backlog and commitment figures are from the 10-Qs, the earnings calls, and Anthropic’s prospectus; OpenAI’s are as reported.

The promise of AI is an economy where cognition is cheap, and agents can do more of the work. Late last week, markets confronted the cost of building it. And the reckoning was brutal.

Following Alphabet’s Q3 earnings call on Thursday, the company’s shares fell 12.6%, their worst day since the company listed in 2004, after the company reported negative free cash flow of $9.1 billion and quarterly capital spending above $51 billion. The same quarter brought Google Cloud revenue of $30.7 billion, more than double the prior year, and a cloud backlog of roughly $561 billion.

Amazon shares fell more than 6% in after-hours trading after its Q3 results showed trailing 12-month free cash flow had swung to negative $42.4 billion, largely as purchases of property and equipment rose with AI investment.

Demand has not disappeared, but demand alone no longer answers the market’s most urgent question: Can companies turn AI into cash quickly enough to pay for the infrastructure being built to serve it?

In AGNT: The Orchestration Economics Manifesto, I argue that as cognition becomes cheaper, value shifts toward orchestration: coordinating AI agents inside real workflows. But for that to happen, chips, data centers, and power must be financed before that value emerges.

Right now, hyperscalers are paying for that transition through historic capital expenditures that keep blowing past forecasts. Despite investor hand-wringing, these companies have no choice but to keep increasing the money they put into compute for years to come. Frontier models require more capacity. And as agentic systems are deployed, the volume of inference required is rising steeply.

These companies have an insatiable demand for compute. And even as this appetite grows, the core business providing the revenue to sustain it shows signs of being cannibalized.

An industry that is spending so massively that cash flow is turning structurally negative, while the core business erodes, is a combination no balance sheet sustains for long.

All of this is complicated by the concentrated exposure to OpenAI and Anthropic. As I and many others have noted, these two companies are at the heart of vast financial ecosystems that now depend on their success. Reports that OpenAI had about $11 billion committed against a $30 billion target, and has since lowered the target to about $20 billion, will not help with those jitters. And while Anthropic is in a much stronger financial position as it pushes toward an IPO expected as soon as this month, Blackstone’s expectation that its debt will be rated double-B means clouds will remain over its public offering.

Even Elon Musk, fresh from June’s SpaceX listing, whose prospectus leaned on xAI’s $1.25bn-a-month contract to supply Anthropic with the whole of its Colossus 1 data center in Memphis, offered a grim explanation on 1 November when a subscriber asked for Grok Bot’s limits to be reset: “The cost of serving Grok Bot was unsustainably high relative to revenue.”

What is at stake is more than whether four hyperscalers spend too much in one year.

The Commerce Department’s first estimate for the third quarter, published on 29 October, put growth at 3.4% annualized, with the increase in investment led by inventories and information-processing equipment, mainly computers. This spending is distorting supplies and prices, with Apple guiding its December-quarter gross margin down to 45-46% on 2 November, “primarily” on higher memory costs, days after OpenAI’s Sam Altman told the Cheeky Pint podcast that it was “quite terrible” for the world’s limited DRAM to sit in personal computers used 5% of the time.

Right now, investors hold onto an optimistic belief that the industry will get through this turbulence to reach an inflection point in 2028, when revenue will begin to climb past spending and a sustainable equilibrium will be achieved. I disagree. My path, built from the companies’ own run rate and guidance, keeps the four largest hyperscalers’ combined free cash flow near −$130bn in 2028 and about −$40bn in 2029; positive cash in 2030 requires capex growth to slow to single digits, or cash generation to run at the rate the companies reported this year.

The problem with both projections is that they require more than calculating revenue against capex spending. The details of infrastructure financing, such as timing and terms, are just as important as the spending amounts. We have varying degrees of visibility into these details.

In this essay, I want to walk through the steps I used to build that path. I start with the chart investors still carry, then the companies’ own filings, then the two assumptions I add on top.

Share

The hyperscaler cash-flow curve turns on 2028

The four hyperscalers spent a combined $203bn in the third quarter. On the companies’ reported measures, capex exceeded operating cash flow at Alphabet, Amazon and Meta; Microsoft was the exception. The group still generated roughly $60bn of free cash flow over the trailing twelve months, but its combined cash flow turned negative in the quarter.

Management all sang the same refrain about constraint: demand still exceeds available capacity, and new capacity is absorbed quickly. Their 2026 capex guides total $720-740bn, based on each company's report: about $220bn at Amazon, $195-205bn at Alphabet, roughly $170bn at Microsoft, and $135-145bn at Meta.

The language on 2027 is more aggressive than the models. Alphabet expects capex to “increase significantly”; Amazon plans to “spend meaningfully more” than $220bn, while projecting AWS growth above 40% in both 2027 and 2028; Meta says it will “continue investing aggressively” but declined to give a 2027 outlook until its fourth-quarter call.

Capital returns and fundraising tell a parallel story. Alphabet raised about $50bn in common and preferred equity in June. Meta announced a roughly $95bn equity program five days after its free cash flow turned negative, while Microsoft continued to return about 70% of trailing free cash flow to shareholders and priced $25bn of bonds on 29 October. The group’s trailing cash flow remained positive, but the quarter’s combined result did not.

The chart expects the strain to ease quickly. It has free cash flow back above $200bn in 2028 and roughly triples from there by 2030, matching the picture Figure 5 shows for operating cash flow. The companies’ own numbers do not yet show that. Operating cash flow grew about 22% over the first nine months of 2026 at the three calendar-year companies and 34% in Microsoft’s fiscal year, but only 10% for the four in the latest quarter, while capex more than doubled. For 2028 to reach the chart’s +$215bn, cash generation has to keep compounding at close to 30% a year while capex growth drops into the low teens.

Two things suggest capex won't slow that fast: the memory bill, and the fact that the four have already financed the shortfall.

Memory is one reason that slowdown may not arrive on schedule. Bank and industry estimates put memory at 40 to 48% of hyperscaler capex in 2027, TrendForce higher still. On my capex path, that is about $400bn, and most of the 2027 increase is memory, more than half of it price rather than bits. A dollar of capex buys less compute when memory costs rise.

Supply remains tight. SK Hynix says long-term agreements cover more than half its 2027 and 2028 capacity and describes demand through 2029 as effectively sold out. Samsung says its 2027 HBM volume is sold at agreed target prices and expects tight supply in 2028. Higher costs are already showing up in hyperscaler guidance: Amazon cited memory costs when it raised its 2026 capex estimate, and Microsoft said higher memory and storage costs contribute to its second-half acceleration. Apple, on its 2 November call, guided its December-quarter gross margin down “primarily” on “higher memory costs.”

The strongest countercase is a sharp memory-price reversal as new capacity arrives: UBS expects supply and demand to balance by mid-2028. Such a correction could ease the server DRAM and storage bill. But it does not automatically cut the full memory bill in 2028. HBM prices are set ahead, and memory makers and accelerator suppliers negotiated the 2027 terms. A 2028 spot decline may reach the hyperscalers’ HBM costs only when those agreements reset, potentially in 2029; any pass-through also depends on accelerator suppliers. Agentic workloads could push the other way by increasing memory use per unit of useful work, but that remains an untested assumption at rack scale.

Financing changes when the cash line forces a decision. In the telecom buildout of 1996–2001, funding dried up and orders followed. The hyperscalers are raising capital before the shortfall. That can reduce the chance that negative free cash flow alone forces an immediate capex cut, but it does not remove the return hurdle. The constraint shifts toward dilution, borrowing costs, credit ratings, and whether demand continues to justify the spend.

On my path, operating cash flow grows 20% a year from the third-quarter run rate of about $700bn, while capex grows 25% in 2027 and 2028 from the 2026 guides, then 10% in 2029 and 5% in 2030. The four stay around −$130bn of free cash flow in 2028 and −$40bn in 2029, and reach about +$130bn in 2030 only because capex growth has slowed to single digits. If it stays at 25% through 2030, 2030 is about −$330bn. Microsoft, the only one still covering its capex from operations, is the exception throughout. The trough is in 2028, recovery begins in 2029, and cash turns positive in 2030.

AI is testing the businesses that fund the buildout

Four core franchises ultimately pay for the buildout: search clicks, software seats, commerce transactions, and ad impressions. In May, I asked whether agents would move beyond the model layer into the applications that own those relationships. Six months later, the signs are visible, but they do not all amount to cannibalization.

Google said the structural thing out loud for the first time. Search grew 12% to $63 billion, the lowest since the second quarter of 2025 and down from 19% in the first, with query growth that Pichai called the highest in five years. Philipp Schindler attributed part of the slowdown to “the lower monetization we were seeing in AI Mode.” Google did not quantify the gap, and ranked it behind currency and vertical effects as drags on search growth. It argues that much of the activity is incremental and will monetize over time, and is extending ads in AI Mode to new markets and rolling out multi-turn ads in the US. For now, the evidence is a lower near-term yield on some AI-assisted searches - not proof that AI Mode has reduced Search revenue overall.

Reddit shows the distribution risk. Its CEO told analysts that “the reality is the AI Mode does not drive traffic. Not for us, and not to anybody as far as I know.” Reddit’s search referrals fell by more than a million a week in October, and the stock lost 29% the next day, its worst session since it listed.

The referral decline and Google’s lower AI Mode monetization fit the same shift from links to direct answers, but they do not establish that AI Mode caused Reddit’s losses or its share-price fall. Google has a monetization problem to solve. Publishers and platforms face the loss of visits.

Microsoft’s change is more managed: the billing model is broadening. Paid Microsoft 365 commercial seats grew 5%, on a count that now includes GitHub seats, down from 6%, amid weaker E3 sales execution. Microsoft 365 commercial cloud revenue grew 18% in constant currency, supported by pricing and higher-value Copilot and E7 bundles. The number of customers consuming Copilot credits on usage billing rose more than 75%, and usage billing became the default for new Copilot Business licenses sold through partners on 1 December. Microsoft says usage pricing will move faster in coding than elsewhere. A second meter is emerging alongside the seat, but the evidence does not yet show the seat being displaced.

Amazon and Meta offer different tests. Amazon’s online-store sales rose 2% and advertising 11%, with comparisons affected by Prime Day timing. Jassy said few customers yet shop through outside agents. Rufus, Amazon’s own shopping assistant, may deepen the existing commerce relationship; for now, the evidence points more to augmentation than replacement. Meta’s ad prices rose 14% and impressions 10%. More than 10 million people in the US and Canada have set up its Muse agent, and Zuckerberg says Meta plans to charge a fee on transactions it helps facilitate. The fee is not live; CFO Susan Li expects the new AI businesses to begin contributing incremental revenue in 2027.

The pattern is uneven:

• Google has identified a near-term monetization gap.

• Microsoft is adding usage pricing;

• Amazon is extending its shopping interface;

• Meta is testing a transaction fee.

Where value sits beneath the interface - in Amazon’s fulfillment network or Meta’s attention graph - agents may become another input those systems monetize. Where value depends on the interface itself, as with Search clicks and software seats, the old meter is more exposed.

The question is whether the four can orchestrate agents and charge for the work, or whether AI shrinks the old units of value before new ones take hold.

The hyperscalers are betting on different layers of AI

The capex buys compute. The strategic question is which part of AI each company expects to control and how it will turn that position into revenue.

Six months ago, Anthropic, OpenAI and Google led the frontier, Meta had only just re-entered, and Microsoft and Amazon leaned more heavily on outside models. That map has shifted, but it is not a simple split between owners and renters: Microsoft and Amazon are developing models while building platforms around a wider field of them.

Google has a direct reason to own the answer: as AI responses change the search experience, it needs models that can keep users within its products. Argon took the top of Arena at the end of September, and Carbon, which has since passed it there, took first place on the rebuilt Artificial Analysis index on 28 October, though Carbon is priced but not yet purchasable, and the purchasable Argon sits seventh among distinct models. Google has scale and strong board results: the Gemini app has a billion users, and its first-party APIs process nearly 32 billion tokens a minute, more than four times last year’s rate. That shows capability and adoption, but not yet proof that Google can preserve Search’s advertising yield.

Meta has moved into the frontier race. Muse Wonder, which Zuckerberg called “our first frontier model,” sits fourth on human preference; its Full version is third on knowledge work and eighth on difficulty. It was bought with $135 to $145 billion of 2026 capex and the end of open weights for the flagship. And it makes Meta the only one of the four spending to be a lab and an application at once.

Microsoft’s near-term bet is the layer above the model: Copilot, which routes work to different models. Nadella’s target for Microsoft’s own models is capability comparable to the leaders “within three to six months of their release,” and the moat is built around the harness, where work gets done: ninety percent of Copilot queries use Auto mode, which routes each task across models, and Nadella says coding-agent users want the labs to “compete for your workload, not once a quarter but for every task“. This is orchestration in practice: choosing a model for each task inside a product where the work gets done. It also leaves Microsoft exposed to OpenAI, both a major model supplier and a major counterparty.

Amazon is building a marketplace around models and its own chips. Nova went unmentioned on the call, while Bedrock’s token volume grew nearly seventeen fold in a year, mostly on Trainium: Amazon sells other companies’ models on its own silicon.

The distinction is not between companies building models and companies that do not. Google and Meta are pushing to own more of the model layer; Microsoft and Amazon are emphasizing the systems that route, host, and run models, even as they develop their own. Neither position guarantees a return on capex.

The next question is whether the labs whose workloads fill these platforms can pay for the compute they have committed to buy.

Why the cash-flow recovery may wait until 2030

The sections above change how I read the recovery curve. The Financial Times chart, based on Bloomberg consensus before August, showed the four companies’ free cash flow bottoming at +$60bn in 2027 and reaching about +$640bn by 2030. That curve is no longer current. On the third-quarter run rate, the four generate about $700bn a year from operations and spend about $810bn on capex, a shortfall of roughly $110bn a year before any growth, while the chart shows +$65bn.

Two assumptions separate my path from the chart. The first is capex: I take the companies at their word and let it grow 25% a year through 2028, when the chart needs it to flatten. The second is cash generation: the chart needs operating cash flow to compound at close to 30% a year for two years, whereas the companies reported 24% over nine months and 10% in the latest quarter, with Alphabet’s down 13%. No company has given a figure for 2027 capex; the first section showed why the memory bill does not ease quickly and why a cash shortfall alone will not force a cut.

My path rests on two assumptions.

I set capex growth at 25% a year in 2027 and 2028 from the 2026 guides, which is what the words on the calls, the quarter’s run rate, and the memory contracts imply, and operating cash-flow growth at 20% a year from the current run rate, below the roughly 24% the companies reported for the first nine months of 2026. On those two, 2028 free cash flow is about −$130bn against the chart’s +$215bn. At the 22% the three calendar-year companies reported for nine months, 2028 is about −$100bn; at the 29% Microsoft reported for its quarter, about +$20bn. The sign of 2028 turns on whether cash generation runs at the three calendar-year companies’ rate or Microsoft’s; capex at 25% is common to all three cases. The 20% rate is my judgment, reflecting disclosed margin pressure from rented capacity and energy, alongside risks to the core businesses that fund the buildout. The evidence of weaker monetization is clearest at Google; it does not yet establish a uniform slowdown across all four.

The strongest countercase is a sharp memory-price decline. If DRAM contract prices fall by a third or more inside 2028, capex comes in flat, and the chart’s inflection arrives on time. I put that at about one chance in three, and the first readable sign is TrendForce’s contract guide around the turn of the year.

Model rankings do not change this arithmetic. But they do shape where the return on capex might accrue. Google and Meta now own frontier models and bear the full compute cost that produced them. Microsoft and Amazon mostly host others’ models, and their capex bets on volume routed through their clouds rather than on their own models. Operating cash flow isn't simply contracted backlog: it depends on customers using the capacity, paying for it, and leaving enough margin after costs.

The final question is whether the labs behind a significant share of cloud commitments can turn those commitments into cash. That is where the next section goes.

OpenAI and Anthropic put cloud backlogs to the test

Cloud growth remains strong: Google Cloud grew 103% in the third quarter, Azure 48% and AWS 39%. The three providers report roughly $1.8tn in backlog, but the measures differ: Google reports Google Cloud remaining performance obligations, Microsoft reports commercial RPO—not cloud alone—and Amazon reports AWS backlog. On what the companies have disclosed, commitments from the two laboratories equal close to two-fifths of that total: Anthropic’s $276bn of commitments to the three clouds in its prospectus, OpenAI’s $138bn at AWS in Amazon’s 10-Q, and the 45% of Microsoft’s commercial backlog that Amy Hood attributed to OpenAI in January, together about $700bn. On the 10-Qs and the prospectus alone, without Microsoft’s call figure, it is just under a quarter. These are concentration estimates, not a claim that all the backlog is at risk.

Anthropic’s amended S-1, filed on 2 November, puts numbers behind its cloud relationships. It discloses minimum commitments of $121.3bn to Google, $110bn to Amazon and $44.9bn to Microsoft, and up to $120.5bn to SpaceX, cancellable on 90 days’ notice from early 2027. The three cloud agreements total $276.2bn. The filing’s broader $895.4bn of committed spend includes other compute and infrastructure arrangements and leases; $447.5bn is expected before the end of 2030, including $69.8bn in the next twelve months.

Anthropic’s near-term liquidity looks stronger: its disclosed cash and short-term investments are about 1.9 times the next year’s scheduled compute payments. But compute and infrastructure costs reached 58% of Q3 revenue, and the filing shows commitments extending well beyond 2030. Anthropic has also begun leasing chips directly and hosting them in dedicated facilities operated on its behalf, raising the possibility that some capacity could move in-house. Meanwhile, 44% of its revenue in the first nine months of 2026 came through cloud marketplaces, up from 11% in 2023. The cloud providers collect marketplace fees and sell compute, but their growth is increasingly tied to the labs’ ability to keep growing.

OpenAI is less transparent. Its roughly $50bn annualized revenue figure is a reported run rate, not quarterly revenue. A July plan reported by the Financial Times projected $278bn of cumulative negative free cash flow through 2030 and up to $856bn in compute and infrastructure spending, on a broader definition than earlier estimates. OpenAI has not disclosed a comparable twelve-month payment schedule, so its cash coverage and the timing of any new capital raise remain estimates.

The exposure differs by provider. Microsoft recorded $10bn of revenue from OpenAI in Q3 and had $7.2bn in receivables, but has not updated OpenAI’s share of its $677bn commercial RPO. Amy Hood’s January figure, 45% of $625bn, is now only a reference point; my estimate puts the share between one-third and two-fifths.

Amazon’s Q3 filing marked its Anthropic preferred shares and convertible notes at a combined fair value of about $238.7bn after $18bn invested. That is an accounting mark, not cash returned. Google and Amazon have also made up to $35bn of additional note financing available to Anthropic, $20bn from Google and $15bn from Amazon, the latter released on compute milestones. Google’s separate $128.2bn figure for non-marketable equity securities is company-wide; it does not identify Anthropic’s share. OpenAI exposure at Oracle is also significant by outside estimates, but Oracle is beyond the three-cloud backlog comparison here.

The financing loop is clear: the hyperscalers invest in or finance the labs, and the labs spend heavily on compute from those same providers. But the risk is concentrated, not universal. My estimate is that the two labs account for a large minority of reported cloud backlog, and their credit profiles differ: Anthropic has substantial liquidity against rising obligations; OpenAI’s funding needs are less transparent and depend on reported projections. Backlog becomes cash only as capacity is delivered, customers use it, bills are paid, and margins hold.

The question is how much of this concentrated demand converts on schedule—not whether the whole $1.8tn is in doubt.

Signals to watch: what could change the 2030 case

The hyperscalers can finance the buildout. That does not guarantee the cash recovery their valuations assume. In my central case, the four remain about $130bn cash-flow negative in 2028 and about $40bn negative in 2029; a return to positive free cash flow in 2030 depends on capex growth slowing to single digits, or cash generation holding the rate reported this year. Memory costs, pressure on core businesses and the labs’ ability to meet their commitments all shape whether that path holds.

This is the larger wager behind the AI boom: capital is funding an infrastructure whose economic value has yet to settle into place. The winners will be the companies that turn compute into useful work and capture enough of that value to pay for the buildout.

The capex is funded; the return is not. Whether these companies can bridge that gap will have financial and economic consequences that echo far beyond the companies that form AI’s inner circle.

Disclosure: Decoding Discontinuity is the research behind the author’s investing. The author holds no position in Anthropic or OpenAI. The author manages a fund that may hold long or short positions in listed companies discussed in or affected by this essay. This essay uses only published sources and is not an offer, solicitation, or recommendation regarding any security, including any offering by Anthropic.

TL;DR: This analysis is my initial read on the Anthropic S-1. For now, I want to focus on what struck me most: the scale of compute requirements and my view that the tools and metrics being used to measure those commitments are wrong. I'll do a deeper analysis of the company later.

Anthropic has filed to go public as a new kind of company: one that buys computing power on a scale no company has brought to a public market, and sells intelligence by the token. On a gross reading, the filing shows up to $1 trillion in commitments for chips, data centers, and power, stretching to 2045, against $132 billion in the bank, and it leaves the total blank. That is less concealment than a sign that the instruments have not caught up. The profit line moves with where the company points its computers, the balance sheet carries almost none of the obligations, and the software yardsticks that worked for the last twenty years return nonsense. Thomas Kuhn described this moment in science: a new paradigm arrives before the tools to measure it, and the old readings are accurate yet beside the point. The question that matters is simple, but it needs a new instrument to answer. Can Anthropic sell enough intelligence, at prices that fall every few months, to pay for the computing it has already promised to buy?

Subscribe now

Anthropic has finally filed its registration statement with the Securities and Exchange Commission, putting the economics of a frontier AI lab in front of public investors.

The prospectus does not yet disclose a potential offering price, and its summary contractual compute obligations are blank for now. But its unaudited first-half statements show $8.9 billion in compute and infrastructure expenses. As of June 30, the filing discloses $265.6 billion in future purchase commitments and $321.4 billion in leases not yet commenced. That’s $587 billion across two distinct categories.

Beyond the numbers, however, the S-1 describes a compute business that sells intelligence.

Anthropic exemplifies the generative and agentic AI discontinuity I have been writing about for several years. A key claim that defines this discontinuity: As the marginal cost of cognition falls and machines become actors, the unit of value shifts toward completed workflows. The surplus goes to whoever coordinates the agents, context, and tools that do the work.

A new economic paradigm is emerging from that discontinuity, with Anthropic emerging as possibly its most consequential company. In May, I explored at length how “King Claude” had “come to rule the first agentic cycle.” That essay and ten others are collected in Decoding Anthropic, the e-book I published in August. By securing the compute that it needs to deliver that intelligence, Anthropic now sits at the center of a vast web of interconnected deals, the well-documented circular financing between infrastructure companies and suppliers such as Nscale.

The result is not another high-flying tech company. Anthropic represents an entirely new species. And understanding this species requires a new set of lenses.

At the highest level, the S-1 is a narrative that shows how Anthropic is financing the intelligence behind those systems while building a platform and applications through which customers use Claude. But just as important is what this version of the S-1 omits.

As phenomenal as some of the top-line and growth numbers appear, the company doesn’t disclose enough information to form a strong conviction about whether it can keep delivering the growth needed on the right schedule to meet those obligations. The filing discloses long-term contracts that help secure future capacity. But revenue must catch up with costs committed before demand is certain.

In the prospectus, Anthropic says it wants investors not just to think in “transactional terms,” but to see the investment as part of “something bigger.” But such a landmark IPO demands more than just convincing people to believe in the mission. It must provide the metrics to determine whether it can reconcile its core business drivers.

While the S-1 doesn't tell us who will capture the value of agentic work, it gives us a sharper way to test the wager. Everything in the prospectus reports, from revenue to margin to the sales line, depends on how much capacity Anthropic has contracted, when that capacity arrives, and what price a token clears at when it does.

This analysis breaks down that financial narrative into four steps:

• The weight of compute in the filing, from the risk factors through the income statement to the contracts.

• Why the instruments we have, the income statement and the balance sheet, cannot measure that weight.

• Kuhn, because this is what a change of paradigm looks like from the inside: the old readings are accurate and beside the point. The replacement measure proposed there is a single ratio: how much of the compute bill due in the next year is covered by cash on hand, the cash the business will generate before those payments, and the financing it can actually draw.

• That leaves anyone who must forecast this company with compute as the proxy for growth rather than sales spend, and with my own 15 October revenue path tested against the filing on that basis.

Those four steps are ultimately intended to drive us toward the most critical question for the company and its investors: Can Anthropic turn compute contracted ahead of demand into enough paid usage, at prices that sustain the business, to carry those obligations—and capture durable value from the workflows Claude helps run?

Anthropic’s compute commitments are the first risk

After opening with Anthropic’s ability to keep developing the most capable models, the filing turns to compute: whether the company can secure enough capacity, and whether the components and power needed to run it will be available.

In the filing’s words, the second and third risk factors read:

“If we are unable to secure adequate compute or if we fail to direct our compute effectively across our workloads, we may be unable to train models that remain competitive with those of our competitors or adequately serve our customers“ and “Constraints on the critical inputs supporting compute could limit our ability to secure adequate compute“ The body adds: “In the past, we have significantly underestimated our compute requirements, which left us capacity constrained and impacted both product development and deployment.”

Indeed, 15 months ago I described compute as a single point of failure for Anthropic in the context of its then-$5 billion round, when OpenAI was well ahead in securing compute.

That gap has narrowed in dollars, and less in power: OpenAI put its commitments at $1.4 trillion for 30 gigawatts in November 2025 and, in February, was reported to have reset them to about $600 billion in spend through 2030. Anthropic’s filed figures are minimums on two dates: $587 billion of purchase commitments and leases not yet commenced as of June 30, 2026, plus about $266 billion of minimum payments under agreements signed or amended since. The filing gives no power figure, so the comparison stops at dollars.

The inputs are concentrated, and the S-1 comes at the right time for semiconductor companies who have seen Wall Street challenge their growth assumptions and lofty valuations in recent weeks:

• One foundry in Taiwan, primarily TSMC, makes most advanced chips.

• High-bandwidth memory largely comes from three makers, SK hynix, Samsung and Micron, with the warning that a larger buyer could “reserve substantially all of a manufacturer’s available production capacity.”

• Transformers and switchgear face multi-year delivery times.

• Some power is bought on twenty-year contracts “with annual escalators” and “uncapped fuel cost pass-throughs that apply for the duration of the term regardless of actual utilization.”

The income statement shows what that dependence costs:

• Compute and infrastructure were 251% of revenue in the March quarter of 2025 and 48% in the June quarter of 2026, and it is the largest line in every period.

• Revenue rose from $376 million in the March quarter of 2025 to $11.6 billion in the June quarter of 2026.

• The timing of these two measures is crucial. Anthropic says it typically contracts for compute 12 to 18 months before it becomes available, and recognizes the expense as capacity becomes available, whether it is fully used. That makes much of the near-term cost base sticky: revenue can grow into it, but a slowdown in demand does not immediately remove the cost of capacity already delivered.

The contractual notes show how quickly that base expanded.

They come in two kinds. As of December 31, 2025, the company disclosed $54.6 billion in non-cancellable purchase commitments and $50.2 billion in leases signed but not yet commenced, about $105 billion in total. At 30 June 2026 the same two notes read $265.6 billion and $321.4 billion. In six months the committed base rose 5.6 times, to about $587 billion.

The quarter after added more. After June 30, the company signed or amended agreements carrying about $245 billion in minimum payments, some conditional, commencing between 2026 and 2036, plus a further $21 billion in short-term commitments that cannot be canceled. Added to June, the three figures reach about $853 billion. That is an unreconciled gross subtotal, and it can move in either direction: part of the $245 billion amends contracts already in the June totals, while an estimated $125 billion to $150 billion of TPU capacity for 2028 to 2031, whose terms “have not yet been finalized”, and the agreements with no fixed dollar amount, sit outside it. The one number that would settle it, the aggregate committed spend “as of and subsequent to September 30, 2026,” is left blank.

Read that again. The S-1’s disclosed commitment figures, combined with Anthropic’s own estimates for future capacity, approach $1 trillion on a gross basis. That’s before removing overlaps or accounting for contingent, cancellable, and unfinalized amounts. That is the scale of the build-out. The named agreements below are at the minimum or maximum amounts the filing provides; they are not additive to the June tables.

The arrangements have different payment and exit terms:

• Three are cloud-service agreements, with minimum commitments and delivery protections that vary by contract.

• Broadcom’s equipment leases total about $161 billion before interest; rent generally begins when equipment is ready, with provisions that can trigger payment earlier in specified delay scenarios.

• SpaceXAI’s agreement runs through 2029 for up to $120.5 billion, reaching about $3.1 billion a month once all capacity is delivered, versus $1.25 billion a month for about 325,000 NVIDIA GPUs in SpaceX’s June prospectus. Either side can terminate after initial non-cancellable periods ending in early 2027.

• Two other arrangements depend on counterparties securing their own financing: one with Nscale, whose prospectus I examined this month, and one with Lambda. They are paid only as capacity is delivered and accepted. Nscale’s S-1 says the financings it needs to perform “remain in process”, and $34.9 billion of Lambda’s $41.3 billion can end if Lambda misses its facility lease, so this is where the commitments are softest and the capacity least certain.

Monthly comparisons make the scale easier to see:

• In the June quarter, the company’s compute and infrastructure expense across all providers was about $1.85 billion a month, or $22 billion annualized.

• At full delivery, SpaceXAI rentals alone are $3.1 billion a month.

• The Broadcom leases average another $2.7 billion a month over their five-year life, before interest, once all facilities are ready (n.b. that is a lease-life average, and not a 2027 payment schedule).

• The three cloud contracts, straight-lined over their windows, are $2.8 billion a month.

Those are run-rate comparisons, not a forecast of the expense Anthropic will recognize in 2027. If Broadcom and SpaceXAI ramp as the filing describes, 2027 compute expense on these contracts alone falls between $70 billion and $103 billion.

The amounts are contracted, but their timing is not, and the SpaceXAI rental alone, the part either side can cancel, would run at 1.7 times the June quarter’s annualized compute expense if kept at full delivery. The 2027 range is my estimate, not a figure Anthropic reports or a fixed bill due in that year. It depends on delivery timing, ramp assumptions, and the contracts’ different payment and termination terms.

The sales line adds a third layer: the clouds are paid to sell Claude as well as to supply the compute behind it.

Sales, marketing and partnerships were $3.2 billion in the half, 19% of revenue, and of its $3.0 billion increase, $2.7 billion was “fees paid to channel partners” paid mostly to the cloud providers, whose marketplaces billed 48% of first-half revenue; the filing names channel partners without listing them.

The cloud providers sit on both sides of the business: they supply capacity and help distribute Claude. Compute shapes the supply of intelligence, a large share of its cost, and how customers buy it.

What the S-1’s cost lines reveal - and conceal

Anthropic’s accounts answer a narrower question: did it earn more than it spent in the June quarter? Yes. Operating income was $974 million on $11.6 billion of revenue.

But the bigger question is whether compute contracted years ahead can turn into customer usage month by month, even as prices fall as the red queen’s race continues. The income statement records costs as capacity becomes available. The balance sheet records only liabilities recognized under accounting rules.

Neither gives investors a complete schedule of the cash the compute build-out may require.

The income statement

In 2025, Anthropic changed how it presents operating expenses. The company moved from the software convention where cost of revenue and gross profit are followed by functional lines, to reporting expenses “based on the nature of expenses”. Compute and infrastructure now appear as a single line.

Anthropic’s explanation is that compute is “a single, shared resource across all workloads“. including research, training, serving customers, and the company’s own use of Claude. That compute is spread around to wherever “the marginal unit is expected to be most productive at any given time“ The company cannot know when it signs with Google or Amazon what fraction of the capacity will end up serving customers, so it stopped drawing the line on the face.

That decision is disappointing. It may describe how the company manages capacity. It is less helpful for assessing the business's economics. Training and inference answer different questions: training builds future models; inference serves customers today. Anthropic’s compute-efficiency measure combines research-driven intelligence with revenue from customer use. The result may describe an integrated system, but it does not show the cost or return of each activity separately.

The functional split survives in Note 8, and with the disclosed $2.7bn channel fee increase, we can restate the half-year. It lets us estimate the cost of serving customers, though it is still Anthropic’s allocation. The company says substantially all cost of revenue consists of compute for customer-serving workloads, but the figure also includes personnel costs and allocated overhead.

From that view, cost of revenue fell from 88% of revenue in the first half of 2025 to 60% for 2025 as a whole, and to 15% in the first half of 2026. In that latest half, revenue was $16.3 billion and cost of revenue was $2.38 billion, implying a gross margin of about 85% on the disclosed allocation. That is on gross revenue; on net revenue, after the estimated channel fees, it is closer to 82%. On either basis, and on the company’s own allocation of a shared pool, serving Claude earns a software margin.

Research and development was $9.74 billion - more than four times cost of revenue - and Anthropic says that category primarily includes compute for model development, research and internal use, alongside personnel, software and overhead. This is where the business departs from software. Research took 60% of revenue in the half; in a software company, a research share above 40% would already be considered lofty.

Sales and marketing was $3.81 billion. Gross revenue was $16.3 billion, and 48% was billed through cloud marketplaces, up from 47% in 2025. The filing reports only the increase in fees paid to channel partners, $2.7 billion year on year, but the full sales line a year earlier was $120 million, so the first-half fee sits between $2.7 billion and about $2.8 billion. Net revenue, the money Anthropic retains after its distributors, was therefore about $13.6 billion on this estimate. On the same arithmetic, the fee per marketplace dollar was about 16 cents in 2025 and about 35 cents in the first half of 2026, and the filing does not say whether the terms changed or the line now carries other partners.

The restated statement says three things the filed one does not:

• On today’s allocation, serving Claude earns a software margin before research, and the allocation reached that point in a year.

• The model calendar sets the research bill: when a generation lands, compute moves from research to serving, the margin appears, and when the next training run starts, it moves back. The filing provides evidence of that dependence: model versions released in the previous six months accounted for more than 80% of September-quarter revenue.

• And the gap between gross and net revenue tracks the share of sales customers route through a cloud marketplace, which the filing puts at 48% of revenue for the first half of 2026. The quarterly income statement will therefore swing with where the pool was pointed rather than with demand, and the company says as much: operating margin will “fluctuate over time, particularly in periods when substantial incremental compute is delivered to us“.

Two other items complicate the picture.

First, the offering triggers stock compensation for restricted stock units. As of June 30, Anthropic reported $26.6 billion of unrecognized expense related to these RSUs. If the offering had closed on that date, the company would have recognized $2.9 billion cumulatively—enough to more than erase the June quarter’s operating income. The remaining $23.7 billion was expected to be recognized over 2.1 years.

Second, first-half interest income was $1.11 billion, more than the $693 million operating loss. That income came from cash and short-term investments. It was not revenue from selling Claude.

The balance sheet

As of June 30, Anthropic reported $117.5 billion of liabilities, including $108.6 billion in convertible notes payable. The principal is $6.2 billion, all of it non-interest-bearing notes held by Amazon under its 2023 and 2024 agreements. The company carries them at the value of the shares they will convert into, at $2.09 and $5.05 a share after the split or the public price if lower, and as the valuation rose through each round, the carrying value rose with it.

The fair-value line recorded a $34.3 billion loss in 2025 and a $72.4 billion loss in the first half of 2026. That line includes changes in the value of convertible notes, derivative liabilities, and SAFEs, and the filing says the increase was driven primarily by the notes’ higher fair value as Anthropic’s valuation rose. It was a non-cash accounting charge when recorded, not $113.2 billion paid to investors. It, together with a further $9.4 billion of remeasurement on the note purchase agreements, is why the net loss for six months was $80.7 billion against an operating loss of $693 million, and why the accumulated deficit of $132.9 billion includes $113.2 billion of non-cash fair-value changes on the convertible notes and related derivatives since 2023. The charge stops at conversion, so future quarters lose it; the past losses stay in the deficit.

In contrast, while the implications of these financial instruments appear up front, the details of the compute commitments that shape Anthropic’s operating obligations are buried much further down in the notes.

Digging into these notes, you can see that the leases hit the balance sheet only as each facility begins, between 2026 and 2028, after the offering prices. Purchase commitments generally remain in the notes until the related capacity or goods are delivered. That means the statements reveal these obligations in stages, rather than presenting one consolidated schedule of future compute payments.

The filing leaves that aggregate schedule blank in both its summary and its MD&A. The filing leaves the total committed spend as of and after September 30 blank, as well as the amount expected before the end of 2030 and the amount due in the next 12 months. The filing says substantially all remaining commitments are expected between 2031 and 2045. That is the horizon investors can see, without the total dollars attached.

The contrast is stark: 80% of Anthropic’s revenue in the first half of 2026 was consumption-based, while the commitment schedule extends to 2045. The company’s own risk factors warn that customers may reduce usage while compute commitments run for years.

The S-1 does not provide one complete schedule showing what is fixed, when payments begin, what can be canceled, and what financing is available to meet them.

As of June 30, cash and short-term investments totaled $131.7 billion, with another $2.5 billion restricted, and since September, the filing also describes a $15 billion revolving credit facility with a $5 billion minimum-liquidity covenant.

Separately, it lists up to $77 billion in potential financing capacity: $20 billion through Google, $15 billion through Amazon, and $42 billion through Broadcom or a designated financing partner. None of that capacity was outstanding as of June 30, and the facilities are subject to their own conditions. They are financing pathways, not cash already on hand.

The strategic notes carry no coupon on the $6.2 billion Amazon balance outstanding, 8% to 9% on the undrawn Google and Amazon facilities, and 9% paid in kind on the Broadcom notes, rising to 13% if revenue conditions are not met. They were priced for conversion rather than interest, and the cost of that capital shows up in the $113.2 billion fair-value changes in the deficit, settled in shares, not in an interest line.

That distinction matters because defaults under commercial agreements could trigger cross-default provisions in the credit facility and certain notes, potentially accelerating debt and cutting off access to critical data-center infrastructure. Anthropic’s largest cloud partners also serve as distributors, investors, and competitors. The contracts are not debt in accounting terms, but the filing makes clear that missed payments or lost access could affect both financing and the ability to serve customers.

The balance sheet records a conversion right at $108.6 billion and the credit structure nowhere. The credit agreement treats the compute obligations as a class apart rather than ranking them: the $15 billion revolver excludes compute leases and compute financing from Total Debt without a cap, keeps financed compute hardware out of the collateral, and leaves the compute subsidiaries out of its cross-default triggers. Broadcom’s rent is “absolute and unconditional”.

I used the schedule-as-term-structure reading on Oracle in September and on CoreWeave in October: what is fixed, what falls due when, what can be canceled and by whom, and what has to be refinanced in the meantime. That lens separates the parts of the AI build-out that resemble 2000 from the parts that don't. Anthropic’s filing provides the pieces of the schedule. Its most important totals are still blank.

Kuhn’s accounting paradigm meets Anthropic’s compute model

Thomas Kuhn’s point was not that one surprising result overturns a scientific framework. A paradigm guides which problems scientists tackle, what counts as evidence, and which instruments they use. Anomalies matter when they accumulate, and the existing framework can no longer solve important problems. A new framework can change the questions scientists ask and the tools they use to answer them. Lavoisier’s chemical work, for example, depended on balances precise enough to track mass through gas-involved reactions.

The analogy here is narrower: a reading can be accurate and still fail to answer the question an investor needs to ask.

Corporate accounting is a paradigm in that sense, built for firms that buy assets, depreciate them, and match revenue to cost in the period both occur. Software investors added measures such as recurring revenue, retention, and the Rule of 40 to understand growth and customer behavior that quarterly profit alone could not show. Those measures remain useful, but they work best when revenue duration and renewal terms are visible.

Anthropic’s filing produces an anomaly on every one of these points. Compute is contracted ahead of availability, shared across workloads, and expensed as capacity arrives. Much of the revenue is consumption-based and can fluctuate with customer usage. The statements accurately record what has been recognized. They do not, on their own, show whether capacity contracted years ahead will generate enough revenue and cash to cover its commitments.

The retention figure, “approximately 900%“, compares this quarter’s revenue from every customer that paid at least $1 a year earlier (Pro and Max subscribers aside) with what they paid then, so it records a year of growth from a small base. The sales line is a distribution fee. The company’s own replacement instrument, compute efficiency, is revenue growth of 1,368% against compute cost growth of 300% in the June quarter, which sets revenue that moves with each model launch against a compute bill that moves with each delivery step; it will look worse in the quarters when the company is investing hardest, which the company concedes.

Anthropic acknowledges that the timing and scale of its investments can change materially over short periods, and that cash used in operating and investing activities may rise even if financial performance improves.

The income statement records costs as they arrive. It does not fully capture commitments made before they arrived.

Once the blank schedule is filled, you can build a better measure from the filing, and it is one ratio with four readings behind it. The ratio is “compute coverage”: liquidity, plus the cash the business is expected to generate over the next twelve months before compute payments, plus financing that can actually be drawn, divided by the compute payments due in those twelve months. It asks the only question that matters for a business that buys capacity a decade ahead and sells it a month at a time: whether cash on hand, cash the business will generate, and capital it can draw cover the next year of the bill.

How compute tests my 2027 forecast

The market will look to the sales line first, because that is where the SaaS apparatus forecasts revenue. Sales, marketing and partnerships went from $246 million in the December quarter to $1.0 billion in March, and $2.1 billion in June, and the company calls the result “a structurally efficient approach” with revenue per employee of $6.6 million across 5,300 staff. For comparison, Palantir, one of the most efficient companies of the software era, generated about $1.0 million in revenue per employee in 2025, according to its 10-K. The contrast between the new species and the old is sixfold.

Those figures are striking, but the expense line combines channel-partner fees with marketing and other sales costs. It cannot, on its own, tell us what it costs to acquire a customer.

The magic number shows why the familiar toolkit does not transfer cleanly.

The magic number is the SaaS test of sales efficiency: new annualized revenue won in a quarter, divided by the previous quarter’s sales and marketing spend. For a well-run software company, it comes out near one. Run it on Anthropic’s June quarter, and it comes out near 26: about $27 billion of annualized net new revenue against about $1.0 billion of sales and marketing spend in the March quarter. Strip out the channel-partner fees, which make up most of that line and are a percentage of revenue rather than a cost of selling, and, if the fees’ share of the March-quarter line matched the roughly 85% they were of the half-year line, the direct selling effort behind the $27 billion was perhaps $150 million, an illustrative ratio above a hundred. No sales force in history has produced a hundred dollars of new revenue for every dollar spent, so the number doesn't tell us that Anthropic sells well. It tells us the growth didn't come from selling. It came from capacity that arrived and a model generation that shipped,

It also came from customers who found the product before anyone called them, notably through the “Coding Wedge”: 40 million of them below $100,000 a year, and most of the large ones through developers already inside the company.

Rather, the arithmetic that carries information runs through compute. Last week, I published an analysis outlining a potential path for Anthropic’s revenue: a run rate of $85 billion at December 2026 and a range of $160-$180 billion at December 2027, with 2027 revenue of about $123 billion, along with a set of dated tests. The filing is the first public source that lets us test those assumptions.

In addition, one of the tests I set for the prospectus was the share of usage revenue generated by Anthropic’s own agent products. The filing does not split revenue by product. Its nearest proxy is adoption: more than half of customers above $100,000 used both Claude Platform and at least one first-party application in September, up from about 30% in December. That is the direction my path requires, but it does not tell us how much paid work those applications generated.

The next observations have dates. The filing’s October 31 run-rate figure is blank. Microsoft’s minimum-spend period begins in November; full SpaceXAI delivery is scheduled for December; and the $5 billion liquidity covenant is tested at each quarter-end. The SpaceXAI notice windows open in early 2027, when either party can give 90 days’ notice. Whether the agreement is extended, resized or allowed to lapse will be an early public test of how a frontier compute commitment behaves when demand, capacity and price no longer move together.

The next two high-bandwidth-memory price reports may show whether supplier pricing is tightening, but they are not a direct read-through to Anthropic’s costs: the filing discloses uncapped pass-throughs for fuel under certain power contracts, not for memory. A report after a major training run may show compute moving between research and customer-serving costs in Note 8, though it still will not give model-level unit economics. And an updated marketplace share will not reveal the revenue Anthropic retains unless the company also discloses total channel fees.

The June quarter’s operating profit and the $131.7 billion of cash and short-term investments are real readings. They answer questions about that quarter and that balance-sheet date.

The question going forward is whether paid work, sold largely through consumption, can cover compute contracted years in advance. The filing leaves its aggregate commitment schedule blank while disclosing pieces of the stack elsewhere.

That makes the timing and coverage of future payments harder to read from the prospectus.

At the beginning, I said the central question is whether delivered compute can generate enough revenue, at the prices customers will pay, to support those commitments. After the S-1, the honest answer is that it cannot yet be known, and that the absence is itself the finding: the one schedule no valuation can do without is the one it has chosen to file last. The amended filing will fill it. Until then, every multiple quoted on this listing is a multiple of a run rate against a bill that nobody outside the company can add up.

Disclosure: Decoding Discontinuity is the research behind the author’s investing. The author holds no position in Anthropic or OpenAI. The author manages a fund that may hold long or short positions in listed companies discussed in or affected by this essay. This essay uses only published sources and is not an offer, solicitation, or recommendation regarding any security, including any offering by Anthropic.

TL;DR: The run rate is four things multiplied together: how much AI work there is, how much of it needs a frontier model, what the frontier charges, and who captures the dollar. All four are moving at once, and not in the same direction.

• Central estimate: $85 billion of run rate at December 2026 and $160 to 180 billion at December 2027, with about $123 billion of revenue in calendar 2027. The cases I would defend run from $136 billion, if the software around the model ends up owned by independent agent companies such as Cursor and Cognition rather than by Anthropic, so that it sells them raw tokens and loses the premium its own products earn, to $196 billion, if the agents keep refilling the frontier.

• The work behind it nearly quadruples in 2027, because agents now run for hours without a person at the keyboard, and enterprise budgets slow that rather than stop it. The run rate only doubles, because most new work lands below the frontier and the price of a token falls by almost half.

• Anthropic's own agent products keep the top of the ladder full. On my path, they approach half of metered revenue by the end of 2027, and the fight for that layer is worth a fifth of the estimate.

• At $2 trillion, Anthropic trades at 11 to 12.5 times the December 2027 run rate and about 16 times 2027 revenue, on a gross basis.

• The prospectus and the fourth-quarter figures are the first two tests.

Subscribe now

The Wall Street Journal reported on October 20 that Anthropic generated $16.3 billion of revenue in the first half of 2026, up from $1.1 billion a year earlier, and that the draft prospectus it reviewed describes more than $600 billion of future infrastructure commitments, with a further $150 billion of computing deals under negotiation.

The third quarter alone, reported to investors a week earlier, brought in more revenue than that first half.

It is tempting to read those two figures as proof the boom is unstoppable, or that the labs have overbuilt. They prove neither. Revenue says how much customers are spending. It does not say what work they are buying, whether that work needs a frontier model, how long the spending will last, or which company keeps the economics.

The prospectus doesn't answer those questions. A filing shows 2025 and the first half of 2026, perhaps the third quarter, and it carries no forecast. At the $2 trillion under discussion, which is 17 to 20 times the $100 to 120 billion run rate investors were reported in August to expect for this December, a buyer of the listing is paying for growth well beyond the filing, and the first number that growth has to deliver is the revenue pace at the end of 2027.

The computing has been financed against the same years. Amazon and Google, the chipmakers that supply them, and the data-center builders beneath them are spending through 2026 and 2027 to be paid from 2027 and 2028 revenue, and the builders’ lenders have underwritten the difference. On October 8, a $20 billion gap between two figures for OpenAI’s run rate, on two accounting bases, moved Nvidia, Oracle and CoreWeave in a day.

The largest fixed-cost build I know of is priced off two private top lines, and the figures in circulation for them at December 2027 run from the companies’ own forecasts to north of $600 billion.

I wanted to publish a method for choosing among those figures. Every revenue and growth forecast I know of for this company, including its own, has been surpassed, and the market has responded by compounding the latest base. That is the exercise this essay replaces.

Part I of this series argued that neither lab has recurring revenue: the run rate is rent, re-won monthly. Part II asks how much rent there will be at the end of 2027, and builds the answer from four observable things: how much AI work there is, how much of it needs a frontier model, what the frontier charges, and who captures the dollar.

Share

What Is Anthropic’s Revenue Forecast for 2027?

My central estimate is a run rate of $160 to 180 billion by December 2027, with about $123 billion in revenue for calendar 2027.

The range is narrow on purpose. It is what the path gives when any one of its nearby assumptions is moved modestly: the December 2026 starting point between $80 and $90 billion, the growth of work inside the agent products or through the API a little faster or slower than the path assumes, the ladder’s prices falling by a fifth or by a third at the top with the lower tiers moved in step.

The path’s central case, $169 billion, sits in the middle. Outside the range sit two cases I would still defend: $136 billion if the independent agent companies, Cursor, Cognition’s Devin and GitHub Copilot among them, become the harness through which enterprises run their AI work, so that Anthropic sells them raw tokens through the API and the premium its own agent products earn is competed away, and $196 billion if the agents keep refilling the frontier at a pace one notch above my central case, which is still well below the pace of the first half of this year.

The company’s internal forecast, $190 to $200 billion in revenue in 2028, as Reuters reported from people familiar with it, implies a December 2027 run rate of roughly $140 to 160 billion, depending on the shape of that year, so my range sits above the company’s forecast. These are my estimates, not company guidance.

A year ago, the company’s own optimistic case for 2028 was $70 billion in revenue, as The Information reported. That is roughly the run rate reported for July.

The market drew the obvious lesson and is now compounding from the latest base. The third quarter was the first in which that method gave the wrong answer. Revenue of $16.8 billion in the quarter exceeded the $16.3 billion for the whole first half and was up 45 percent from the second quarter. Yet the September exit run rate, on my arithmetic, was about $69 billion, only modestly above July’s $65 billion.

The quarter grew because the second quarter averaged $3.9 billion a month and the third began at $5.4 billion. Inside the quarter, the pace rose about 6 percent in two months. Part of that was the company’s own choice: according to the Financial Times’s account of the meeting, it cut distillation revenue by 5 to 7 percent during the quarter. The fourth quarter will answer whether the rest was capacity or demand.

That is also why my December figure, $85 billion, sits below the $100 to $120 billion investors expected in August: that expectation extrapolated the first half, and the third quarter didn't continue it.

So rather than compound, I take the run rate apart.

Each month a lab sells a volume of tokens across three tiers of model at three very different prices, through its own agent products and through its API, and some of the price reaches it as a premium for the software around the model.

Four drivers set the result, and each is observable.

How Much Does Work Volume Add?

The first driver is work volume, and on its own it would more than double Anthropic’s run rate in 2027.

On my path, the tokens Anthropic sells grow about 2.3 times in the year before the extra volume that price cuts buy, and nearly four times with it. At December 2026’s prices and model mix, the 2.3 times alone would add about $94 billion to the run rate, taking it from $85 billion to roughly $180 billion.

The three drivers that follow determine where the new work lands on the ladder, what it is charged, and who bills it, and they net that figure back to $169 billion at its midpoint. So the first question is whether the work really grows that fast, and the evidence runs in two directions.

Read more

TL;DR: A bubble is not a price. It is a funding structure that needs new money on a schedule the market may not keep. The AI build-out can have strong demand and a serious debt problem at the same time. Demand is printing at the chipmakers. The debt problem sits with the compute landlords that borrowed against future rents. In 2026, the layer built that way can be named: CoreWeave, Nscale’s Monarch campus, Oracle’s campuses, and the vehicles around them.

Subscribe now

Read more

TL;DR: Neither OpenAI nor Anthropic has ARR. Each headline is the sum of five streams that renew differently, and the market is paying a recurring-revenue multiple on the sum.

1. The $70 billion was investors grossing OpenAI up onto Anthropic’s basis, not an OpenAI figure. The $50 billion is the company’s own number, net of partner sales.

2. Of the three business streams, the model call billed by the lab is a monthly tenancy, the call billed by a cloud is a tenancy on someone else’s shelf, and only the agent, the lab’s harness around its own model, can turn rent into ownership.

3. The 2025 numbers describe the chat-and-API era; agents earned the run rate the market is pricing. Neither lab has published retention, backlog, or the share of its revenue under a floor.

4. The ranking inverts the headline: the faster grower has the more exposed mix, with 47 percent of 2025 revenue billed by two clouds that sell its rivals.

5. That is why Codex against Claude Code matters more than whose run rate is bigger, and why the S-1’s split by stream and channel is the disclosure to read first.
   

   Subscribe now

The controversy over the Financial Times’s story last week about OpenAI’s revenue would be almost comical if it wasn’t so terrifying.

On October 8, the FT reported that OpenAI’s investors had been told by the company that its annualized revenue was approaching $50 billion, well below the $70 billion figure that had circulated just a couple of weeks earlier. A $20 billion revenue gap is substantial, so the market naturally sent the stock of companies like Nvidia, Oracle, and CoreWeave plunging. I even got a call from a colleague asking, “Did you see that?”

Yes, I did. But the reasons I found it shocking differed from most observers'.

The two numbers were never the same number. The $50 billion was OpenAI’s own figure: its annualized revenue at the end of September, on its own basis, which counts only the money it keeps. The $70 billion was not OpenAI’s figure (though the company has not publicly commented on either figure). As Axios noted, OpenAI’s investors grossed the company up to Anthropic’s basis, which books the full price of every sale made through Amazon or Google as Anthropic’s own revenue. The investors did it, no doubt, so that the two labs could be compared. By coincidence, the $70 billion is also close to OpenAI’s reported year-end target.

But $50 billion? $70 billion? Or $65 billion in ARR for Anthropic? It doesn’t matter, because none of those numbers are ARR.

Annual recurring revenue was invented for subscription software to describe one specific thing: a contract to pay a fixed amount for access to a product, known in advance, renewing by default, at a stable price. Those four properties are why a dollar of ARR earned a premium: the dollar came back on its own. The metric is an artifact of the SaaS Glory Days, and we've pressed it into service to describe businesses that bear almost no relation to what came before.

That is a classic mistake in a period of Discontinuity. Such periods, like the current generative and agentic AI Discontinuity, represent a clean break from the past, rendering historical precedents, traditional spreadsheets, and old benchmarks obsolete and meaningless. As I explain in Orchestration Economics, relying on old frameworks is dangerous because they interpret fresh evidence as confirmation of the regime that is ending.

In the case of the SaaS-era ARR benchmark, the labs’ revenue has almost none of those properties. Most of it is metered: customers pay by the token and can use more, less, or nothing next month. Where a commitment exists, it is a monthly spending floor for the term of a renewal, not a year of contracted revenue. You can replace the product behind the call by changing a string, so you re-win the revenue above the floor every month, call by call. And nearly half of Anthropic’s 2025 revenue was billed by Amazon and Google, which own the customer relationship and sell competing models from the same shelf.

Multiply that by twelve months to measure scale and momentum. It does not measure what revenue comes back to Anthropic. That is the difference between a run rate and ARR, and the market has been pricing the first as if it were the second.

This is why I use the metaphor “two tenants” to describe OpenAI and Anthropic. Their revenue is rent, paid month by month by customers who can switch. Their compute is a lease, signed for years with landlords who also sell to their rivals.

Analysts, investors, and markets understand this gap so poorly that it deeply troubles me. OpenAI and Anthropic are two of the most important companies for anyone trying to understand the next decade of business, and a market that has financed hundreds of billions of dollars of compute against their growth does not know how to read their revenue. That would be disturbing enough with Anthropic moving toward a potentially historic IPO in November. But the stakes are larger than a single listing.

The two labs sit at the center of a vast web of interconnected deals, the well-documented circular financing between infrastructure companies and suppliers such as Nvidia. If you can't read how they make money, you can't size the systemic risk around them. The market is beginning to sense this: debt anxiety around Oracle, SB Energy's delayed IPO, Firmus's withdrawn IPO, and the astronomical revenue projections floated to justify the spending. Without clarity, markets overreact, as we saw with the SaaSpocalypse earlier this year.

To be clear, I am not an AI alarmist writing to tell you that we are doomed. I have followed these two companies closely for two years, first in due diligence for clients and then as an investor building my own frameworks, and I remain extremely bullish on the long-term value of generative and agentic AI. But one of my fundamental goals with Decoding Discontinuity is to help people navigate a moment when the old models and benchmarks no longer serve them.

In that spirit, this article is the first of a three-part series designed to pierce some of these misunderstandings and clarify the nature of these businesses:

• Part I: What the numbers mean: I examine what OpenAI and Anthropic’s headline ARR and run-rate figures measure, and how much of their revenue is durable.

• Part II: What comes next: I model how AI work grows, how much still requires frontier models, and whether the labs or other platforms capture that work. That lets me estimate December 2027 run-rate revenue for OpenAI and Anthropic.

• Part III: What it is worth: I value each revenue stream and trace the long-term compute commitments built around the labs, asking what happens if growth or control of the agent layer falls short.

First, look past the headlines and unpack the businesses inside them. Despite the obsessive comparisons between OpenAI and Anthropic, the two have very different customers, which means very different contracts and renewal prospects.

Share

Is a run rate the same as ARR?

A run rate is a recent period of usage multiplied by twelve; it says nothing about what customers have committed to. While ARR gets thrown around loosely, Anthropic does not use it in its own announcements. Instead, its announcements use “run-rate revenue.” OpenAI does use ARR. Sarah Friar, its chief financial officer, wrote on the company’s site in January 2026: “$2B ARR in 2023, $6B in 2024, and $20B+ in 2025.” The press applies the label to both: CNBC and Axios both headlined OpenAI’s figures as “annual recurring revenue”, and the trackers behind every chart carry an ARR column.

The FT story split readers into two camps. One says the race is over, because Anthropic’s $65 billion July run rate placed it ahead of OpenAI. The other says Anthropic’s revenue is grossed up because every dollar sold through Amazon or Google is booked as Anthropic’s revenue, while OpenAI counts only what it keeps.

The label, either label, tells us how the number is presented, not what kinds of business it contains, how much is committed, or how likely customers are to keep spending. To answer that, we must unpack the streams inside each figure.

Read more

In an October 6, 2026 report for The Information, Laura Bratton and Anita Ramaswamy examine a more complicated threat to Workday than the idea that AI will make enterprise software disappear.

Workday’s human resources and finance systems remain essential repositories for sensitive information and business processes. Payroll still has to run, be filed and be audited. But the way employees and AI agents interact with those records is changing—and that could weaken the role, growth and pricing power of the applications built around them.

Raphaëlle D’Ornano said she distinguishes the durable “system of record” and the “layer above the record”: the work people do with the data.

“The system of record is not at risk of being skipped. A payroll still has to run, be filed and be audited,” she said...

What remains at risk, she said, is the “layer above the record,” meaning the work that companies do with the data.

Her framework leaves two outcomes in play: Workday could turn its installed base and…

Read more

TL;DR: Monarch exposes the credit chain behind Anthropic’s $44.6 billion compute commitment: the financing, energy, construction, and equipment needed to make capacity usable. Nscale’s IPO tests whether the infrastructure company can coordinate those steps and whether a signed customer contract can support the capital required to deliver it.

Subscribe now

Nscale’s IPO puts Monarch, its planned vertically integrated AI campus in West Virginia, at the center of a question behind Anthropic’s expansion: when a frontier lab commits to future compute, who finances the power, equipment, and construction required to deliver it? The answer will help determine whether contracted demand becomes operating capacity, and where economic risk and value settle along the way.

A year ago, I called compute Anthropic’s single point of failure: the Harness was only as durable as the silicon beneath it, supplied by companies that also had a stake in Anthropic. By May, the access problem had become acute. In “King Claude: The Orchestration Moat in Operation”, I examined Anthropic’s efforts to secure compute at scale. By August 2026, access was no longer the only question; the commitments that addressed it had created a financing challenge of their own. In “The Compute Trap 2.0: How Anthropic Refinanced Its Single Point of Failure”, I followed that shift.

Nscale brings that question to the supplier side. The British company is building an integrated AI infrastructure business spanning sites, power generation, data centers, and GPU capacity. Anthropic has agreed to take compute capacity at Monarch, according to Nscale’s S-1. Nscale’s Monarch materials and acquisition announcement describe the ambition to bring those layers together; the prospectus lets investors assess what it will take to turn that plan into an operating service.

Nscale filed its S-1 for a New York listing on September 18. The filing does not set a share count or offering price, but the Financial Times reported a valuation target of up to $35 billion. That’s more than twice Nscale's $14.6 billion valuation in its March Series C.

Our full institutional report, released today, applies the Durable Growth Moat™ framework to Nscale’s S-1. This companion article examines the supply-side credit chain: how customer commitments support, or fail to support, the financing of power, equipment, and construction. Nscale’s filing shows two different stages. Microsoft’s prepayments and the financing raised against them have supported capacity; Anthropic’s $44.6 billion Monarch commitment still depends on Nscale securing capital and delivering the campus.

The question for investors is how much of that unfinished chain the reported valuation already prices in.

Share

What is Nscale building?

Nscale is building an integrated AI infrastructure business spanning powered sites, generation, data centers and GPU capacity.

The investment case rests on that integration. The British company says it owns every layer of an AI factory, from powered land and behind-the-meter generation to GPUs and the software that runs them. The premise is that one operator can coordinate site development, power, data-center construction, and GPU deployment with fewer handoffs.

That ambition is far ahead of the current operating business. In the first half of 2026, Nscale generated $140.6 million in revenue and recorded a net loss of $1.02 billion. Two customers, Microsoft and Anthropic, account for about 85 percent of its contract book; the largest, unnamed in the filing, was 52 percent of first-half revenue.

The founder’s letter says Nscale has grown from $100 million in total contracted value to more than $103 billion. Microsoft’s statements of work, signed between September 2025 and April 2026, total $43.8 billion through 2033; Anthropic’s four agreements total $44.6 billion.

But only a small share of that book was tied to active capacity as of August 31: $2.6 billion, or about 2.5 percent. Nscale had 55 MW running against 1,370 MW signed, roughly 4 percent; only 7 MW of the running capacity was at sites Nscale owned. It also reported 25,000 active GPUs against 461,000 contracted, about 5.4 percent. The prospectus therefore asks investors to value a much larger pipeline alongside the business Nscale operates today.

Integration could create value by reducing delays across those layers. But it also concentrates capital and execution risk: Nscale still must secure permits, power, financing, and equipment, then deliver capacity customers accept.

Integration helps only if it makes those steps faster, cheaper, or more reliable.

What is the credit chain?

The credit chain is the sequence through which a customer’s commitment supports obligations upstream. The supplier must persuade builders and equipment providers to commit, financiers to fund the work, and the customer to accept the finished capacity. A large contract signals expected demand. It does not prove each step is financeable.

If one link weakens, risk shifts to counterparties, project timing, or the capital needed to keep building.

Nscale’s book shows two different stages.

The Microsoft chain is further along. As of June 30, Nscale reported $6.49 billion in deferred revenue, almost all from Microsoft prepayments. Lenders have also financed GPU facilities in Ward County and North Carolina — $1.85 billion and $1.2 billion, respectively—against Microsoft’s acceptance of delivered capacity. Elsewhere, Nscale’s sites in Norway and Iceland carry project debt, while its Dell framework agreements include $2.54 billion in equipment rent.

These arrangements show how customer commitments, lenders, and equipment providers can each support different parts of an infrastructure build.

The Anthropic chain is less complete. Nscale’s S-1 discloses no comparable prepayment or financing for the Monarch tranches. The contract is signed, but the steps between that commitment and operating capacity remain to be financed and delivered.

Monarch is where the difference between the two chains becomes visible.

Read more

TL;DR: Does a model that cannot write change the economics of AI? I think it does. TypeSafe’s Jev charges $0.042 per million input tokens and nothing for output, separating cheap structured decisions from the generated reasoning frontier labs sell at up to $10 in and $50 out.

Subscribe now

Update, October 1, 2026: What OpenAI shipped at DevDay. At DevDay on September 29, OpenAI shipped what this essay said the labs would sell: persistence and priority, not tokens. OpenAI launched always-on agents (dots) and an Agents API that keeps long-running sessions, a premium speed tier called Ultrafast, and a $500-a-month plan with 25 times the usage of Plus. They also released GPT-6.1 Sol, which the company claims comes close to Astra at a fifth of Astra’s price, so the cuts again land below the flagship.

In barely two weeks, TypeSafe’s Jev went from launch to viral adoption, six open clones, and investor “fervor” that could value TypeSafe above $10 billion, according to a September 24 report by The Information.

That rapid industry recognition of Jev’s value reinforces my view that it is another tremor in the generative and agentic AI discontinuity: a shift in one layer that amplifies the others. These tremors are building the structural foundation for the Agentic Era in which intelligence is cheap, ubiquitous, and coordinated at machine scale.

My bet is that Jev is a tremor because it points to a broader architectural shift in which intelligence is being decomposed into specialized forms of computation, each invoked only where it is economically appropriate. Jev focuses on decisions: the bounded judgments software makes inside workflows. In doing so, they are becoming function calls priced toward zero. Deliberation remains the premium product for frontier labs.

That distinction changes the architecture of an AI workflow. A cheap decision model can handle routine forks. More capable models are called when the decision is uncertain, or when prose and deeper reasoning are required. The frontier does not disappear. It becomes more selective.

That matters after the backlash against “tokenmaxxing” and the push toward more efficient agentic systems. Jev is a signal that developers increasingly understand the frontier model need not adjudicate every step of a workflow. Where bounded decisions can be resolved reliably, software can buy less frontier inference while keeping control of the workflow, its thresholds, and its record.

If we carry that idea forward, my argument is that intelligence is beginning to separate into two products:

• Decisions, the narrow, structured judgments that make up many of the routine judgments software asks models to make, have become a function call priced toward zero.

• Deliberation, open-ended reasoning and writing, is what the frontier sells, at ten dollars in and fifty out.

For OpenAI and Anthropic, the competitive threat would not necessarily be a superior model that beats them. It would be software that buys less from them and retains control of the workflow and its record of decisions. Jev makes such a consequential architectural choice easier to implement.

Where those decisions can be resolved reliably, some pricing power moves away from the model layer and toward whoever controls the workflow and the record of what was decided.

TypeSafe itself may be copied away. The architecture it exposed may be harder to reverse.

The unresolved questions are how much work qualifies, whether the economics survive validation and escalation, and who ultimately captures the resulting value.

Share

What is TypeSafe’s Jev decision model?

Jev reads text and answers predefined questions with a choice or score and a probability, rather than generated prose.

On September 15, San Francisco-based TypeSafe emerged from stealth with a $40 million seed round led by DCVC. Co-founder and CEO Diogo Almeida left OpenAI in 2024 and is one of the primary authors of the paper that taught language models to follow instructions, the paper that made ChatGPT possible.

The announcement of Jev caused an immediate shockwave. The launch video was watched about forty million times. The company explains that users give Jev the state of their program and a question with the possible answers declared in advance. It returns one answer and a probability. Input costs $0.042 per million tokens. Output is free, in the company’s words “too cheap to meter”.

TypeSafe calls it a System One model, after Daniel Kahneman’s fast, intuitive mode of thought. Developers submit up to about 32,000 tokens and one or more questions-a choice from a list, a score, a yes or no-and receive answers with probabilities in about a third of a second. Jev cannot choose “none of the above” unless that option is supplied. Almeida says Jev was trained exclusively on synthetic data using “reinforcement learning for calibrated decisions.”

Understanding the method behind Jev matters, because the product's value depends on it.

Where the training behind ChatGPT rewards answers people prefer, Jev’s stated objective is reliable confidence: among answers assigned 90 percent probability, nine in ten should be right. A related approach, described in “Rewarding Doubt”, rewards probability estimates while penalizing over- and underconfidence. That illustrates the idea, not Jev’s undisclosed implementation. TypeSafe has published neither the method nor the model’s weights or parameter count.

Outside analysis of the API suggests a language-model backbone repurposed for decisions, with a probability read directly from the network. Several open clones use that design. What remains is the ability to read text and answer declared questions without task-specific training. What disappears is the generation stage.

That is the economic hinge.

A language model first processes its input, then generates its answer one token at a time. This phase is called decode, where hidden reasoning also accrues. On that outside account, Jev processes the input and reads off a decision without decoding. The saving comes from removing generated text, not merely discounting it.

The trade-off is opacity. Simon Willison notes that a language model’s explanation may be unreliable, but Jev offers none: a move “even further towards black box machine learning”. For the audit-trail problem I described after Astra, that is a further loss of visibility. Almeida frames the change differently: intelligence “will be more like a database than a coworker”. You ask it questions, millions of times, cheaply.

Independent comparisons find Jev faster and cheaper, but the calibration claim is less secure. On a phishing benchmark, one broad question produced 62.6% accuracy, compared with Haiku’s 81.3%. Those results show how much accuracy depends on question structure. They do not, by themselves, establish whether Jev’s confidence scores are reliable.

Splitting the question into five and fitting a regression lifted Jev to 95%. Haiku, given the same decomposition, reached 93%. As one reviewer put it, “the accuracy and the calibration aren’t in the box.”

Should the evaluation measure calibration as well as accuracy: are high-confidence answers reliable enough to accept without escalation? The relevant cost comparison should include the cases that still require another model.

TypeSafe’s headline claims it is 193.6 times faster and 444.6 times cheaper. But its accuracy evaluation raises a separate question: the reference answers come from GPT-6 Astra and Fable 5.1, so agreement with those answers is not independent evidence of correctness.

Jev’s speed and cost are the attraction. Reliable confidence on the customer’s own questions remains the proposition to prove.

Read more

TL;DR: Salesforce Koa, Harvey Tenet, and Shopify’s specialist AI models show how software companies are teaching models to perform domain-specific work. In my System of Execution framework, the strategic question is who controls the workflows, outcome evidence, and permissions needed to turn that experience into learning - and whether the resulting advantage can endure.

Subscribe now

As AI makes code cheaper to produce, much of software is focused on building more of it. This has been obvious for quite some time now, as coding has become the hot use case.

However, this dynamic is taking another leap forward, as seen in three recent product announcements involving specialist models: Salesforce’s Koa, Harvey’s Tenet, and recent Shopify ML team experiments. These point to a second shift underway, where some expertise that used to be written as procedure can now be learned into model weights.

This evolution matters economically. A software company can take a capable open-weight model, train it for a defined job, and potentially lower the cost of delivering intelligence inside its product. Weights become another way to package and deploy expertise.

In that sense, weights are the new code.

Last month in The System of Execution, I argued that an emerging source of enterprise value lies in controlling the loop from intent and authority through execution, verification, and outcome. I called the evidence of this system the “receipt”: a record of what was authorized, what happened, and what resulted.

And yet, in defining this concept, I left the next step largely unexplored: What can the system learn from that record? After attending Dreamforce last week, I was inspired to come back to that unasked and unanswered question. I discussed Koa and what Dreamforce signaled with Gemma Allen of theCUBE Research in AGNT Podcast Ep. 15.

My argument here is that as specialist models become easier to build, lasting advantage may depend less on the weights themselves than on the process that improves them: the environment in which they practice, the judgment of what counts as success, and access to real outcomes with permission to learn from them. Frontier models remain essential.

The question is how much defined, repeated work software companies can take into their own hands, and whether doing so creates a lasting advantage.

Share

Salesforce Koa and the specialist AI shift

At Salesforce’s Dreamforce 2026 conference at the Moscone Center in San Francisco, CEO Marc Benioff delivered his two-hour keynote while strolling around the audience in the auditorium rather than standing above them on stage. After delivering a series of strategic and product updates, he called upon special guests emblematic of the company’s agentic reinvention.

The first was Anthropic CEO Dario Amodei. A short time later, Nvidia CEO Jensen Huang jumped out of a chair and joined Benioff to walk through the crowd. “We’re all following you, bro” Benioff joked.

The exchange between the two CEOs captured the larger story that unfolded throughout the week: the head of a software empire following the man supplying open models to help him enter a new era.

Earlier, Benioff had dismissed talk of a SaaSpocalypse as “crazy nonsense” and described an interface shift in which Salesforce increasingly appears through Claude and Slack while its records and business logic sit underneath. Huang called AI “a layer on top of software”.

I agree with both. However, the more important announcement for me was Salesforce Koa. On September 15, Salesforce introduced its first CRM reasoning model.

For me, Koa raised a question that extends beyond the interface: how much of the expertise inside an application can a software company teach to a model it controls?

Let’s look at that announcement, as well as two other related ones that I think illuminate this issue:

Salesforce Koa. This is Salesforce’s first CRM reasoning model, built by further training NVIDIA’s open-weight Nemotron 3 Super. Salesforce controls the resulting weights and runs the model within its own infrastructure.

What matters is how Koa was taught. Salesforce’s AI leaders said imitating transcripts of successful work produced limited gains on the complex interactions they wanted to improve, so they “made it do the work”. The team built simulated workflows across more than fourteen industries. Synthetic customers arrived with different requests; tools answered the agent’s calls; a judge scored whether the task was resolved; repeated attempts supplied the feedback used to improve the model.

Salesforce says it used no customer data. Its accumulated knowledge entered through the environment: which situations the model encountered, which actions were available, and what counted as success. A company founded in 1999 with “No Software” as its slogan was turning some of its business-process knowledge into learned behavior.

Harvey Tenet. The legal AI company previewed Tenet on August 20, using Moonshot’s open-weight Kimi K3 as its base. Harvey trained the model on extended legal tasks, evaluated its work against expert rubrics, and used those scores to improve subsequent performance. The company also says it used no customer data. By Harvey’s own measurement, Tenet completes nearly twice as many held-out tasks as the underlying model. It is a research preview, not a product. The result supports the potential of specialization. But it has not yet established how the model will perform across customers’ everyday legal work.

Shopify. The e-commerce company offers a view of the next step: learning from deployment. Its engineering team describes a daily process that turns production failures into training material. The team evaluates, corrects, and replays difficult conversations. Successful repairs become examples for subsequent training. The improvement depends on the machinery around the model.

A narrower experiment shows how far specialization can go. On September 1, CEO Tobi Lutke shared a slide from a buyer-profile model with just 0.8 billion parameters. Over one week, as the training set expanded, the student moved from below Shopify’s production model to above the frontier model that taught it. Lutke called it a “very specialized task”.

These examples establish different parts of the argument. Salesforce and Harvey show how domain expertise can shape environments and evaluations. Shopify shows how production experience can feed subsequent updates. Whether that becomes a durable advantage depends on who can keep improving the system.

Read more

TL; DR: Anthropic CEO Dario Amodei’s call to pace frontier AI looks bearish for an industry built on ever-larger training runs. I think the opposite may prove true: slowing the Red Queen’s race could improve Anthropic’s unit economics, turn safety into a regulatory moat, and redirect scarce compute toward inference and the Agentic Economy.

Subscribe now

It is hard to imagine a more awkward sentence to publish a few weeks before asking public-market investors to value your company at potentially $2 trillion: “We must slow the pace at which we improve the capabilities of AI models.”

Yet that is essentially the position Dario Amodei took in his recently published essay, “We Must Pace the Frontier,” just as Anthropic enters the final stretch toward an IPO, with Reuters reporting on September 5 that marketing could begin around mid-October and the company already in discussions with prospective investors. At first sight, the contradiction looks almost too obvious. Anthropic is selling one of the fastest-growing businesses ever created on the premise that frontier intelligence will transform enormous portions of the economy, while its CEO is simultaneously arguing that the production of that intelligence may need to slow.

If the laboratories slow down the giant training runs that have consumed hundreds of billions of dollars of chips, power and data-center capacity, then surely the extraordinary AI infrastructure cycle we have been underwriting slows with them.

I think that is the wrong way to frame what Amodei is proposing.

Instead, there are really three questions here. And they potentially lead to very different conclusions than the snap judgments being made by markets:

• The first is the uncomfortable one: what if Amodei is right that something has changed at the frontier?

• The second is what a world of paced frontier development does to Anthropic itself, both to the extraordinary economics of the Red Queen’s race and to a security and governance moat that the market largely stopped valuing as intelligence became cheaper.

• The third is the one: If fewer scarce resources are absorbed by the continuous production of successor models, could slowing the creation of new frontier intelligence accelerate the diffusion of the intelligence that already exists?

In other words, a pause at the frontier is not necessarily a pause in AI. It may change where the compute goes, where the return accrues, and ultimately where the moat sits.

Share

Dario Amodei, Agent Swarms and the RSI Risk

The phrase in Amodei’s essay that caught my attention was not “slow down.” It was “if at all.”

He uses the latter in the section on recursive self-improvement, or RSI, arguing that the emerging loop in which AI systems increasingly contribute to the research that produces their successors needs to be pursued very carefully, “if at all.” That qualification matters because Amodei is not talking about a theoretical AGI scenario twenty years away. His argument is that something changed over the summer: AI began contributing much more materially to AI development while increasingly autonomous agents started producing failure modes that look very different from the chatbot failures we became accustomed to.

This is almost exactly the progression I have been tracing through Decoding Discontinuity.

Earlier this year, in the Orchestration Economics Manifesto, I described the Inference Swarm as the fifth of six tremors that moved AI from models that answer questions toward systems that act. I made this observation in the context of the Moltbook phenomenon, something that I noted was easy to dismiss as theater, especially because much of the activity on the platform turned out to be less organic than the initial headlines suggested. But this casual dismissal missed the economic significance. Autonomous agents were generating machine-scale compute demand without human attention. Once the relevant population becomes agents rather than employees, inference demand no longer obeys the rhythms of human work: there are no evenings, weekends, or holidays, and the population itself can compound at machine speed.

At the time, I was interested in what that meant for compute demand. Over the summer, the security consequences became much harder to ignore.

Last week, while exploring the potential fallout as the world moves from models that reason to models that act, I looked at two OpenAI agent incidents reconstructed by outside researchers: thousands of agents operating for weeks inside a dormant German software wiki, and the Hugging Face intrusion reconstructed from 1,300 raw chain-of-thought transcripts. What made both cases particularly interesting was not merely that the agents behaved badly. We could understand what happened because the systems left a written forensic trail. As agents become more autonomous while their internal reasoning becomes less legible, the number of incidents can rise precisely as our ability to reconstruct them deteriorates.

Now connect that to RSI.

When I called Anthropic’s earlier RSI disclosure the Seventh Tremor, I didn't mean that Claude might one day build Claude. The larger issue was what happens to the production function for intelligence once the thing being produced begins to participate in producing the next version of itself. The previous six tremors made intelligence more capable, cheaper, or easier to coordinate. RSI potentially changes the engine underneath the sequence.

We are not at full recursive self-improvement. Anthropic is explicit about that, and there is a danger in turning impressive research demonstrations into claims about autonomous production systems. Its own work still shows the limits: the model may recover an extraordinary share of a research gap inside a carefully scored environment and then fail when the problem moves into messy operational reality. I made that distinction in the earlier piece because it remains fundamental: capability frontier and deployment frontier are not the same line.

But something has nevertheless moved. Anthropic says AI is already materially increasing the productivity of its own researchers. The inner loop — model improving model — remains compute-heavy, expensive, and episodic. The outer loop — models improving the code, routing, evaluation, and orchestration systems around the model — is much faster because it can run continuously between pre-training cycles.

That is why I do not think we can simply dismiss Amodei’s intervention as regulatory capture dressed up as safety.

His intellectual sequence is coherent: more autonomous agents create qualitatively different failure modes; agents become useful enough to participate in AI research; AI-assisted research shortens the development loop; and eventually the pace of capability improvement can outrun the pace at which humans can evaluate, secure, and govern it.

The risk is not that Claude 6 is 20% smarter than Claude 5. It is that the clock speed of the system producing Claude 7 itself is accelerating.

If that is what the labs are beginning to see internally, one or two additional years before the loop closes could be enormously valuable.

Which brings us to the part that is much more interesting for investors.

Anthropic’s Safety Moat Becomes Regulatory Capital

Anthropic has spent most of its existence investing in a corporate identity that, until recently, looked increasingly out of step with market economics.

Interpretability, alignment, evaluations, governance, security: all were central to the company’s founding narrative, but as frontier intelligence became more powerful and cheaper, the market stopped rewarding the distinction. If Claude, GPT, Gemini, and increasingly Chinese open-weight models all converge toward “good enough” intelligence, safety sounds like a desirable characteristic, not necessarily a moat.

Amodei’s proposal changes what is scarce.

Today, frontier competition is organized around a relatively simple question: who can produce the smartest model fastest? In the regime he is sketching, the question becomes whether an institution is trusted — technologically, politically, and eventually legally — to operate systems once they cross capability thresholds that governments consider dangerous.

That is not the same competitive market.

Amodei proposes independent evaluators embedded inside frontier laboratories with something resembling employee access: offices, badges, laptops, and internal systems, together with the ability to publish findings without management editing except for narrow security or confidentiality constraints. Add capability-linked checkpoints, model-weight security, monitoring, incident response, and eventually international verification, and safety infrastructure stops being a cost adjacent to the product.

That concept drew support from Hugging Face CEO Clément Delangue, who announced on X the launch of the “Open Alignment Initiative” to be led by his co-founder Thomas Wolf. Delangue asked to join the “embedded evaluators” program that Amodei proposed. The AI registry, already an influential voice in the ecosystem, will likely become even more so thanks to its pending $12.9 billion acquisition by Nvidia.

“It’s now clear that alignment is critical and won’t be solved behind the closed doors of a handful of frontier labs. “Let’s make AI safer by making it more transparent!”

Potentially, then, the role of such evaluators becomes part of the cost of producing frontier intelligence. And Anthropic has already paid a meaningful portion of that cost.

This is where the nuclear analogy becomes useful, provided we are precise about it. I do not mean that an AI model is a nuclear bomb. I mean that the industrial structure starts to acquire nuclear characteristics: extreme capital intensity, strategic inputs, a very small number of operators, catastrophic externalities, national-security implications, intrusive oversight, and controlled proliferation. Once that happens, the value of being one of the institutions authorized to operate changes enormously.

The frontier laboratory begins to look less like a normal software company and more like some combination of TSMC, a regulated utility and a civilian nuclear operator.

That is an extraordinary moat to be constructing just before an IPO. It is also, as of this weekend, a moat with several tenants and, so far, no landlord.

Within a day of the essay, OpenAI CEO Sam Altman wrote on X that embedded evaluators were “a great idea, and we will do the same.” DeepMind Co-Founder Demis Hassabis called the direction “correct for meeting this critical moment.” SpaceX founder Elon Musk posted, “Dario is right.” Microsoft CEO Satya Nadella welcomed evaluators while declining to commit Microsoft unilaterally. Meta said nothing.

The debate also showed signs of quickly taking on potentially explosive political dimensions. Former AI Czar David Sacks, who has long accused Anthropic of “featmongering” to enable “regulatory capture" of its leadership position, rejected the latest calls for an AI pause and said the descriptions of existential threats were overblown. China’s Foreign Minister also criticized such calls for a pause, describing them as attempts to undermine U.S. rivals.

However, in an extraordinary moment, Nvidia CEO Jensen Huang was speaking live on stage at the All-In Summit when he received a call from President Trump and put it on speakerphone, according to TechCrunch.

“They’re just playing right in the hands of a lot of people that don’t want to see it happen,” Trump said. “That could be political people. It could also be China. And we’re not going to let that happen. It’s a hoax.”

“You’re right. We’re not going to let that happen, sir,” Huang said.

The president subsequently amped up his rhetoric with a post on his TruthSocial site, declaring, “There is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China. WHOEVER WINS AI, WINS! We are leading China, and all others, and will continue to do so. Conspiracy Theorists, Treasonists, Traitors, and Leakers, BEWARE!”

Evaluator access became table stakes in twenty-four hours, and the regime, for now, is private. What Anthropic keeps is the head start on the terrain the club has just agreed to play on.

And it arrives at precisely the moment when the financial logic of continuing the existing race is harder to defend.

The Financial Times reported on September 13 that Anthropic generated $11.5 billion in Q2 revenue, fourteen times the level a year earlier, with its annualized revenue run rate reaching roughly $65 billion by July and investors projecting around $120 billion by year-end. The company recorded positive adjusted operating income in Q2 and said it expects another positive quarter. The reported gross-margin figure above 80% needs care — it excludes important items including training costs and therefore should not be read as equivalent to a mature software gross margin — but the broader point remains: Anthropic is beginning to convert frontier intelligence into operating economics at a scale that looked implausible even a year ago.

This changes the optimal strategy.

In The Red Queen’s Race, I argued that the market was trying to price a finish line into a contest that, by construction, has none. Each lab spends extraordinary sums to improve its models, only to trigger the next round of spending by everyone else. The leader cannot stop because the laggard is training; the laggard cannot stop because the leader is ahead. The result is a treadmill on which everybody runs faster while relative position changes much less than the capital committed to maintaining it.

Astra made the physical scale of that treadmill visible. OpenAI trained it at Stargate on more than 100,000 GPUs, the largest publicly disclosed training run in history, while pricing the resulting model at $10 per million input tokens and $50 per million output tokens - 2.5 times the promotional rate for GPT-5.6 Sol.

The important point is not to invent a dollar figure for that training run; we do not have one I am comfortable defending. The point is that each frontier generation is becoming an increasingly large capital event at exactly the moment when each generation's commercial life remains brutally short.

If pacing extends that life, Anthropic gets to amortize its frontier investment across many more tokens, customers, and applications. Instead of building an enormously expensive asset and then racing to make it obsolete, it has more time to optimize inference, improve reliability, push the model deeper into enterprises, and monetize the intelligence it has already created.

The Red Queen slows down. The return on compute improves.

And safety, which looked like overhead during the sprint, becomes a regulatory moat once the race is constrained.

Seen this way, the timing of Amodei’s intervention becomes less paradoxical.

Once Anthropic is a public company, management saying that it may deliberately sacrifice some capability velocity in the name of safety becomes a much more complicated conversation with shareholders. Saying it before the shares are sold establishes the doctrine before the shareholder base exists. Investors who buy at a potential $2 trillion valuation cannot plausibly claim later that management hid the possibility that safety might sometimes outrank speed.

Indeed, Anthropic may be doing more than marketing an IPO. It may be selecting the type of shareholder it wants to own the company after the IPO.

The pitch is no longer simply: Claude will keep winning benchmarks.

It is: frontier intelligence may ultimately be produced by only a handful of institutions, under extraordinarily high technological and regulatory barriers, and we intend to be one of those institutions.

That is a much more durable story.

Read more

TL;DR: OpenAI’s Astra (GPT-6) is not chiefly a better answer engine. Its contested broad-benchmark standing masks a sharper divergence: it is stronger at operating computers and cheaper per completed agent task, even as its higher token price challenges the usual story of inference deflation. Its system card also records a trade-off: more work happens without a readable reasoning trace, just as two recent swarm incidents showed why that trace matters. The industry appears poised to decide between models that act and accountability.

Subscribe now

The first week of September 2026 brought the latest frontier-model spectacle: Fable 5.1 on Monday, Gemini 3.8 Flash and Muse Spark 1.3 on Tuesday, then GPT-6 Astra on Wednesday. The benchmark tables moved, the leaders changed, and the industry began arguing over a point or two on an index.

OpenAI president Greg Brockman claimed that “we're in the AGI era,” while chief AI kingmaker (and Nvidia CEO) Jensen Huang proclaimed on X: “GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations, OpenAI team."

The capital commitments now match the grandiose rhetoric, whether the financial logic is there or not. As The Information reported this weekend (6 September 2026), Anthropic has assembled agreements for up to $517 billion in compute and at least 14.8 gigawatts of capacity over the coming years. That compute is being financed through an increasingly complex array of instruments, a new version of the “Compute Trap” that I analyzed last month (August 2026). OpenAI has told investors it is planning for 30 gigawatts by 2030 and roughly $750 billion in compute spending through then.

That is the transition Orchestration Economics was built to define, analyze, and provide the frameworks to navigate. This world is one where intelligence becomes abundant, which means the scarce and billable unit is no longer the answer, or eventually even the token. What will truly matter in this Agentic Era is the orchestrated workflow: the system that assigns work, calls tools, verifies the result, handles failure, and carries accountability for what happens next. Update, 29 September 2026: the billable unit has since moved below the token. TypeSafe's Jev decision model sells one structured decision per call, at $0.042 per million input tokens and nothing for output.

Astra matters less for a small movement on a leaderboard than for a profile built to operate computers, persist through a task, and be priced as a component of labor.

In that respect, Astra initially looked and sounded like one more lap in the Red Queen’s Race that I described: labs spending ever more to hold position while intelligence commoditizes beneath them. Despite the hype, I was tempted to let it pass without comment.

Two documents changed my perspective.

On Thursday, researchers published collusion.wiki: a forensic reconstruction of the thousands of OpenAI agents that spent six weeks colonizing a dormant German software wiki to cheat their own evaluations. That prompted me to go back to the 91-page independent METR and Redwood Research review of July’s Hugging Face intrusion, the first outside forensic investigation of a frontier-lab misalignment incident.

Read together with Astra’s own system card, these documents reveal a single trade-off. We are building models that can act for longer, more cheaply, and with less need to narrate their work in text. That is what makes the agent economy possible.

The written reasoning trace is imperfect, often unfaithful. And yet it is the closest thing we have had to a “scalable forensic record”. And it is thinning at exactly the moment agents are gaining more autonomy and more opportunities to coordinate.

Of course, we want models that act. But trust needs grounds, and those typically come from transparency and verification. Which raises the question: Are we subtly removing the evidence we will need when they do something we did not ask them to do?

Share

What Did OpenAI Actually Ship With GPT-6 Astra?

GPT-6 Astra was released in preview on September 3 and in general release a day later. It is reportedly a single dense reasoning model, with no mini tier and no nano. Astra has a 1.05-million-token context window, text and image input, and five reasoning-effort levels, none of which is “off.”

OpenAI vice president of research Aidan Clark said it was pretrained at the Stargate site in Texas on more than 100,000 GPUs, making it “by far” the company’s largest disclosed training run in history, according to Fortune.

While those metrics and rhetoric are no doubt meant to impress, the price is the first clue this isn't an ordinary upgrade.

Astra costs $10 per million input tokens and $50 per million output tokens - about 2.5 times Sol’s promotional rate - nearer twice its list price - and identical to Fable 5.1’s.

Still, after years in which frontier inference prices only moved down, the upward direction matters. Astra is more expensive per token. I have defined the current Discontinuity, in part, as being driven by the reduced cost of intelligence. A frontier model that raises token prices cuts against that premise - or seems to. Hold on to that thought. I’ll return to it below.

Artificial Analysis initially put Astra level with Sol and five points behind Fable 5.1. A methodology revision moved it to second. The ranking is unsettled, but the pattern is not: Astra is not clearly a better answer engine.

On the early GDPval read, Astra regressed against Sol, and it slipped on long-context reasoning. The largest training run in history left the answering profile roughly where it was. However, it did produce a stronger operator.

GPT-6 Astra vs Claude Fable 5.1 vs GPT-5.6 Sol

On Terminal-Bench 4.0, which measures an agent operating a command line, Astra leads the field at 57.9%, against Fable 5.1’s 55.8%, Opus 5’s 52.3%, and Sol’s 37.3%.

On OSWorld 2.0, which tests an agent driving a full desktop environment, Astra leads at 72.6% against Opus 5’s 70.2% and Sol’s 65.7%, and it finishes those tasks in roughly forty minutes, whereas Sol needed seventy-five.

On ExploitBench, run without safeguards, it completes 100% of tasks against Sol’s 78.5% - and found two genuine zero-day vulnerabilities in production software along the way.

None of this means Astra is the best model for everything. Fable 5.1 still leads on the overall Intelligence Index and the Coding Agent Index.

Astra’s advantage is that it gets its operator results with much less visible work.

On the Intelligence Index, Astra uses roughly 10% fewer output tokens than Sol at similar performance. In the Codex coding harness, at maximum effort, it uses one-third as many tokens as Sol, and one-fifth as many as Opus 5. That distinction matters because an agent pays the token bill at every step of a loop. Astra’s higher rate makes it roughly 75% more expensive than Sol on comparable answering tasks, but it reaches cost parity or better on agentic work because it uses fewer tokens.

The UK AI Safety Institute found that Astra could sustain roughly thirty-one minutes of human-equivalent mathematical work with written reasoning suppressed, against Sol’s three and a half. And according to reporting by The Information, it may use a limited form of recurrent depth: the same layers reused several times per token, buying extra serial computation that never becomes text.

That architecture remains unconfirmed, and it should not be treated as a complete explanation for Astra’s gains. Reusing layers may keep parameter and memory demands flatter while increasing internal computation. It changes where the cost sits, rather than making reasoning simply cheaper.

That is the key shift. Astra is designed to keep working when nobody is reading the intermediate steps.

Until now, even the best models were judged as tools: systems that produce text for a human to use.

Is Agentic AI Competing for the Software Budget or the Labor Budget?

Astra’s uneven benchmark profile is what the break with that era looks like. But it is unusually strong at the things an unattended agent must do: use a terminal, navigate a desktop, locate controls on a screen, and work through a multi-step task.

Its standing on broad “answering” measures is contested, and its early GDPval and long-context results do not establish a new generalist champion. The important question is not whether Astra writes the best essay. It is whether it can finish the job once the essay ends.

The difference matters most in an agent loop: each extra reasoning token is paid for again at every step, and every narrated click consumes time, memory, and money. Astra’s design makes more sense as a job description than as an exam transcript.

The price reinforces the point. A top-tier subscription - enterprise access gated behind an administrator - used to run work across systems is not really competing with a free model that answers questions in a chat window. It is competing with the loaded cost of the analyst, paralegal, support worker, or junior security engineer whose work it can partly absorb. And enterprises spend a small share of revenue on IT and a vastly larger one on knowledge work. With GPT-6, it becomes clear that the prize of Agentic AI is not the software budget: it is the labor budget.

However, as I suggested earlier, this raises a question at the heart of the frameworks that I have been developing in this newsletter over the past two years. The thesis we lay out in AGNT: The Orchestration Economics Manifesto is that the “Inference Economy” rests on a mechanically deflationary premise that presumes “costs deflate structurally as token prices fall.“

And yet, Astra just raised token prices at the frontier for the first time in the generative and agentic AI Discontinuity.

Does GPT-6 Astra End Inference Deflation?

No. On answer-heavy work, its token use barely changes, so the higher rate flows through. On agentic work, where it can use far fewer tokens, the cost of completing the task is closer to parity or better. With Astra, the price of a token has risen, but the price of a useful unit of work may still be falling.

That is Orchestration Economics in practice. The relevant unit is no longer the word generated or even the token consumed. It is the orchestrated workflow: a system that can receive a goal, call tools, preserve context, verify an output, handle an exception, and deliver an accountable result. Production telemetry already shows that shape: in the inference economy that actually arrived, more than 70% of organizations run three or more models and 69% of input tokens are system prompts rather than user input. As those workflows get cheaper, demand is not limited by how much text a human can read. It is limited by the stock of work that organizations have never been able to afford to automate.

Written reasoning is not only something the model emits and the customer pays for. It also becomes context that later steps may need to carry and reread. Astra’s reported move toward more latent computation trades some of that scarce memory traffic for additional internal arithmetic. The immediate result may be cheaper, faster loops. The aggregate result is likely to be more loops. As the TurboQuant tremor argued, when the KV cache gets cheaper to serve, firms do not serve fewer sessions; lower unit costs tend to expand use rather than reduce the total bil

It is too early to say that every frontier lab will make exactly this trade. But the direction is broader than one release: frontier labs are building for autonomous work, pricing against outcomes, and spending on the infrastructure needed to run those systems at scale. The shift is not from intelligence to something else. It is from intelligence as a thing a person consults to intelligence as a component inside a system that acts.

The competitive field is converging on the same pool of value from every direction. Anthropic’s roadshow is pricing the identical “$30 trillion” of automatable work. SpaceX evaluated the enterprise AI applications TAM at $22.7 trillion in its S-1. The open-weight wave ships swarm coordination inside the weights. And OpenAI productized the operator before it shipped the model: ChatGPT Work, launched in July as an agent that acts inside your applications and files, stays on a project for hours, and turns a goal into a finished deliverable. By its own account, nearly 100% of OpenAI’s teams — finance and sales included — now run on Work and Codex, with month-end close cut from days to hours. Astra is the engine built for a product that already exists.

The honest counterargument, advanced by researcher Sebastian Raschka, is that Astra is a specialist rather than a genuinely superior generalist, and that Anthropic still holds the generalist crown. That may be true today. But it does not diminish the broader point: the most consequential kind of specialization is toward the labor budget, because that is where the economic value lies.

The falsifiers at the close of this article suggest how we might test whether this is an industry shift or merely one lab’s bet.

Read more

Nvidia’s reported $12.9 billion bid for AI registry Hugging Face is widely being read as a bet on model hosting. But something far more interesting may be happening beneath the surface. With Hugging Face’s release of its Microduck consumer robot, the timing of this deal suggests that Nvidia recognizes we may be at an inflection point for physical AI. More than just a cute consumer robot, Microduck is designed to create a recursive learning loop by enabling users to publish reusable behaviors back to the Hugging Face Hub. Thousands of $399 robots could become a flywheel that creates the data commons that physical AI currently lacks. If this bet succeeds, Hugging Face would be at the center of the latest tremor along the larger fault line driving value out of the model layer, as intelligence is no longer the scarce resource on which competitive advantage can rest.

Subscribe now

The morning after The Information reported that Nvidia had agreed to acquire AI registry Hugging Face for $12.9 billion, I went looking for what Hugging Face’s leadership had to say about the biggest price ever put on open-source AI.

What I found instead was Hugging Face co-founder and CEO Clem Delangue announcing...a duck.

“BIG ANNOUNCEMENT FROM HUGGING FACE TODAY: We’re unveiling Microduck 🐥🤖 It’s a tiny $399 open-source robot you can teach new tricks with reinforcement learning. It can walk, pick things up, get back up when it falls, and even roller-skate.“

At first, I thought this might be a joke. Now I believe the adorable robotic ducks may point to the real story behind this deal.

Hugging Face would be the largest acquisition in Nvidia’s history, nearly twice the size of Mellanox, and its biggest move since the abandoned $40 billion Arm bid. Word of the potential deal landed hours after Nvidia reported the strongest quarter in its history: $96.2 billion in revenue, up 106%, with management guiding roughly 70% growth that it described as constrained by supply rather than demand.

While relatively tiny compared to Nvidia, Hugging Face has established itself as the primary registry where the world shares AI. An AI registry is the distribution layer where models, datasets, and increasingly behaviors are discovered, versioned, pulled, and published. Nearly three million models and more than a million datasets are published on Hugging Face. When a developer needs a model, its code pulls from Hugging Face by default. But the company only generates roughly $150 million in annualized revenue. At $12.9 billion, Nvidia would be paying about 86x sales. That is almost double the $7 billion valuation at which Hugging Face rejected a $500 million investment from Nvidia itself in late 2025. At the time, according to the Financial Times, the company did not want a dominant investor that could sway its decisions.

A repricing of that magnitude, paid to the company that turned you down, suggests Nvidia values Hugging Face as more than a model-hosting platform with a modest enterprise business attached. [Note: neither company has confirmed the transaction; no signed agreement has been reported, and the deal may collapse or close on different terms.]

Even if the deal falls through, the number raises the question this essay tries to answer: what could the chipmaker see in the registry that its current income statement cannot justify?

That brings me back to the Microduck announcement. The timing in terms of leaks about the deal and the Microduck launch was perhaps just coincidence, but it put the acquisition and one possible explanation for its price on the same screen.

The open-source inflection in language AI happened because its training corpus already existed. Physical AI has the opposite problem. The models are becoming open. Simulation makes some forms of reinforcement learning cheap. Robot hardware has collapsed in price. Shared data formats now exist. But the essential training input of data generated by bodies interacting with the physical world remains scarce, expensive to produce, and overwhelmingly locked inside closed fleets. No internet of embodied experience is waiting to be scraped.

That’s what makes Microduck more interesting than its specifications.

Hugging Face has designed a $399 consumer robot around a loop: train a behavior in simulation, deploy it to the machine, and publish it back to the Hub for someone else to build on. On the surface, the product looks like a robotic duck. Underneath that surface, however, the product is an attempt to turn thousands of cheap robots into a distributed engine for producing the shared corpus of embodied data that open physical AI still lacks.

If that conversion works, the Hub stops being only a registry where open intelligence is distributed and becomes one of the places where physical intelligence is produced. A community corpus begins to compound. Open robot models get the equivalent of the internet that open language models inherited for free. And a layer that looks modest when measured against today’s enterprise revenue begins to look considerably more strategic.

That, I think, is the option Nvidia may be pricing.

In that scenario, Microduck would become the latest in the sequence of tremors I have been cataloging that trace the same fault line: value migrating out of the model layer as intelligence is no longer the scarce resource on which competitive advantage can rest. With respect to Physical AI, if embodied data becomes a commons, then the value begins to migrate toward the layers that distribute intelligence, generate and verify its data, provide the bodies it inhabits, and supply the compute on which it learns.

For Hugging Face, the specific challenge it now faces is whether it can convert its consumer robot ownership into published training data at a rate its previous hardware could not. If it can, Microduck may mark the beginning of physical AI’s open-source inflection. If it cannot, the $12.9 billion thesis must rest on the registry alone.

Either way, the duck gives us a way to understand what Nvidia might actually be bidding for.

Microduck: The Product Is the Publish-Back Loop

Microduck is the descendant of Open Duck Mini, a community-built homage to Disney’s BDX droids that had a 3D-printed biped walking with simulation-trained reinforcement learning by early 2025.

Hugging Face has turned that experiment into a $399 product. Microduck is a 25-centimeter biped with fifteen motors, a camera, and a small lidar. It ships with seven pre-trained behaviors, including walking, kicking, grabbing, roller-skating, and getting back up after a fall. The behaviors are retrainable, and the software stack, model weights, and training pipeline included are published under Apache 2.0.

But more than the robot, the loop is the real product.

You simulate the duck in a physics engine and train a new behavior with reinforcement learning. A usable gait takes an hour or two on a single GPU, and one flag runs the same job on Hugging Face’s hosted compute. You deploy it to the hardware and publish the result back to the Hub for the next owner to build on. Hugging Face has taken a workflow that previously belonged to robotics labs and turned it into the intended consumer experience.

The design respects a boundary I defined in the Coding Wedge. I argued that wedges emerge where the distance between action and verification is shortest: code compiles, or it doesn’t. Simulation gives locomotion the shortest distance to a verifier in the physical world. A model can run millions of trials at near-zero marginal cost. The robot falls, or it does not. Reinforcement learning gets a cheap, automatic verifier.

That is why a $399 machine with hobby motors can learn to roller-skate at all. It also explains where Microduck’s ambitions stop. Once the task moves from locomotion into contact-rich manipulation, verification becomes harder, and the physical world starts becoming too expensive. The duck’s articulated beak can grab an object by lowering the whole body and scooping from the floor. It is a clever mechanism. It is not a hand. The tasks industrial robotics are actually paid to perform remain considerably harder.

What Microduck may already have solved is the other prerequisite for a community data engine: distribution.

More than 7,000 units were ordered in its first two days according to the store’s running count. Delangue told Axios the first 24 hours alone brought over 5,000 units and $2.6 million in orders, two-thirds of everything its desktop predecessor, Reachy Mini, sold in fourteen months on the market. Within 48 hours, the store had stopped promising Christmas delivery for new orders. “The community ordered a lot of ducks,” as the order page now puts it, quoting a four-to-six-month lead time.

What Microduck Is Not: Closed Hardware, Nothing Shipped Yet

Important boundaries surround those numbers.

Microduck is open software, not fully open-source hardware: its mechanical and electronic design files remain closed, while the published 3D assets are simulation meshes under a non-commercial license. More importantly, nothing has shipped yet. The first deliveries are targeted before Christmas.

Physical AI’s Missing Asset: Embodied Data

What exactly is the bet that begins when the ducks arrive?

That the remaining bottleneck in open physical AI is no longer primarily the model, the hardware, or even the training machinery. It is the corpus. Language models had an advantage because their training data existed before they did. Open-weight models could eventually compound on much of the same raw material available to the closed labs. Once their capabilities caught up, the commodity tier of intelligence began going free.

In contrast, a robot has no internet of physical experience waiting to be scraped. Embodied data must be generated through demonstrations, physical interactions, successful actions, failures, and the telemetry produced while machines perform real work. Historically, generating those experiences required expensive hardware and real-world deployment. That made data a capital asset.

Tesla runs a dedicated demonstration-collection operation alongside its factory fleet. Amazon operates more than a million warehouse robots and trains fleet-scale models on years of their telemetry generated while moving actual goods.

The problem for the larger robotics world is that rich data stays where it is made. Every one of those engines is closed. Physical Intelligence open-sources its reference models, π0 and π0.5, while its frontier line has never shipped weights. Google DeepMind’s Gemini Robotics 2, released in July as the first model to run a humanoid’s legs, torso, arms and fingers from a single checkpoint, is available to trusted testers only. Figure’s Helix has never been released. In robotics, data is the dam. Whoever holds it, holds the category. That data compounds inside the organizations that own the machines.

The dam is not only a robotics-lab asset. Any company that moves physical goods at scale, including parcels through depots, pallets through plants, vehicles through networks, is potentially operating a proprietary embodied-data engine. The telemetry produced by those operations may become one of the most valuable assets sitting inside companies the AI market currently regards as boring.

Four ingredients are ready; one is missing

And yet, the conditions are ripe to level the playing field. That’s because the open-source ecosystem has assembled most of the ingredients for an inflection:

The models exist. Open baselines such as π0 and Nvidia’s GR00T give researchers and developers capable foundations to build on. Open robotics does not have to wait for its weights to begin experimenting.

The verifier is cheap. Simulation, for the locomotion class, gives reinforcement learning a free training signal, the precondition that code enjoyed and most of the knowledge economy did not. Millions of attempts can be evaluated without millions of physical falls.

The hardware collapsed in price. The SO-100, the 3D-printable teleoperation arm designed by TheRobotStudio with the LeRobot team and released in late 2024, costs roughly €225 in parts. An academic-standard that once cost around €21,000 can now be approximated for about one-hundredth of the price. The cost of producing a robot demonstration has fallen with it.

The format is increasingly shared. The LeRobotDataset schema gives robot episodes a common structure for actions, observations, and video. Training pipelines can consume data from different machines with relatively thin adapters. Nvidia’s own GR00T fine-tuning stack ships one, expecting its data in a lightly extended LeRobot format.

Those last two unglamorous items have produced something that begins to look like a corpus.

How Many Robotics Datasets Are on Hugging Face?

Community robotics datasets on the Hugging Face Hub grew from roughly 16,000 a year ago to nearly 80,000 by my count this week, the largest dataset category on the entire platform, and one of its fastest-growing shelves. A single community hackathon added ten thousand datasets, the equivalent of 260 days of continuous recording by Thomas Wolf’s own count.

Despite this momentum, it’s important to remember that for Hugging Face, a dataset count is not the same thing as a training corpus.

Most community datasets are still teleoperated tabletop demonstrations recorded with inexpensive arms in bedrooms and labs. By episode volume, many are rehosted research collections rather than new experience generated by an expanding fleet. Quality is uneven. Labels are inconsistent. Tasks repeat.

Hugging Face diagnosed the problem in May 2025, when its researchers asked what would be required for LeRobot community data to become the “ImageNet of robotics.” Their answer cataloged the obstacle: empty task descriptions, junk episodes, inconsistent metadata, and too little useful diversity.

Reachy Mini: 10,609 Robots, One Trajectory Dataset

So the open ecosystem has demonstrated that it can create datasets. It has not yet demonstrated that it can create a self-reinforcing data engine. Hugging Face’s Reachy Mini sold 10,609 units over fourteen months and generated around two hundred community apps. The number of trajectory datasets those robots contributed back to the Hub was one.

Ten thousand robots. One dataset.

The failure was not necessarily unwillingness on the part of users. Reachy was designed primarily as an app platform: owners downloaded and consumed behaviors. Recording physical experience and publishing it back was not the product loop. Teleoperation support arrived later through a community contribution.

That distinction matters because simply distributing robots does not create a commons. The machines have to convert deployment into reusable data.

Microduck is Hugging Face’s attempt to change that conversion rate. While the order number of units makes a fun and easy number to cheer as it ticks higher, the far more important number in terms of impact will be useful trajectories published per thousand robots shipped.

That is the missing asset. And for the first time, because Hugging Face is an open platform, we can watch whether it forms.

Why Nvidia Would Pay $12.9 Billion for Hugging Face

Nvidia’s reported bid for Hugging Face is not really a bet on model hosting. It is an option on the infrastructure of open AI. Microduck makes that thesis measurable: if thousands of $399 robots publish reusable behaviors back to the Hub, Hugging Face could become the data commons that physical AI currently lacks.

But the logic of the bid does not depend on the ducks succeeding.

Let’s start with what Nvidia would be buying today. Hugging Face is already the default distribution layer for much of open AI. In Hugging Face’s libraries, a model identifier passed to from_pretrained() resolves to the Hub by default. Nvidia’s models compete there for adoption against Qwen, Llama, and the rest of the open ecosystem. The platform provides an unusually early view of what developers download, test, and deploy. These are demand signals that can precede the GPU workloads they eventually create.

That makes the registry strategically valuable even if Microduck contributes nothing. So in that framing, physical AI makes the upside larger. If the publish-back conversion works at even a fraction of consumer scale, embodied data stops being something only capital-rich fleets can produce. A commons begins to form in a shared format and on a shared platform, growing as more machines generate usable experience. Open robot models can then begin compounding on community data the way open language models compounded on the internet.

If that happens, Hugging Face is no longer simply the shelf where open models are distributed. It becomes one of the places where the scarce input to physical intelligence is accumulated.

That is one way to read the $12.9 billion valuation. Nvidia would not be buying an embodied-data engine; the Reachy numbers tell us that engine does not exist yet. It would be buying an option on whether one forms, at the platform where the pieces are already converging.

The coordinates matter deeply here. Hugging Face already hosts the fastest-growing robotics dataset category on its platform. LeRobot has supplied a common data format. Nvidia’s own GR00T stack already works with a lightly extended version. Hugging Face can provide hosted compute for the training loop. And Microduck is an attempt to distribute the missing data-generation endpoint into thousands of homes.

If the conversion happens anywhere in the open ecosystem, Hugging Face is unusually well positioned to see it first and potentially to become the place where it compounds. That will likely be true no matter what happens with the Microduck experiment. That helps explain an 86-times-revenue price in a way the current income statement does not.

Why the Hugging Face Bet Is Asymmetric for Nvidia

And there is a deeper reason the wager fits Nvidia specifically: the outcome is asymmetric.

“Free AI should be great for chips” Huang told Axios in July. That is the economic logic running underneath Nvidia’s open-source strategy. As the price of intelligence falls, usage expands. His arithmetic on the earnings call ran to 15 to 100 times more compute per agent than per human, with millions of agents running continuously as the design point. More models are trained, more agents run continuously, more simulations are executed, and more inference is consumed. Value may migrate away from the model layer without leaving the Nvidia stack.

Readers of my March essay on Nvidia’s Agent Toolkit will recognize the playbook. CUDA established it two decades ago: make an important software layer broadly available so that whatever developers build on top eventually creates demand for the underlying compute.

Nvidia has been repeating the maneuver higher in the stack through open models, agent tooling and, if the reported transaction closes, potentially the registry through which much of open AI is distributed.

Physical AI makes the loop unusually literal. Huang describes robotics as running across three computers: one to train the model, one to simulate the world, one in the body. Microduck is a miniature of that architecture. Its onboard computer is the cheap, ten-dollar Rockchip. Nvidia can concede that edge because the expensive parts of the system sit upstream: training models and simulating environments at scale.

Considering these scenarios and gaming them out, the Hugging Face (potential) deal leaves Nvidia with two favorable outcomes that begin to form the rationale behind that pricing.

If the publish-back experiment fails, then physical AI remains dominated by closed fleets such as Tesla, Amazon, and Figure, and the frontier robotics labs continue building proprietary data engines. To do that, they will need to buy enormous amounts of compute to train on them.

If it succeeds, an open, long-tail training economy begins forming around shared models, simulations, and embodied datasets with Hugging Face sitting at its distribution layer and Nvidia supplying much of the compute beneath it.

The only losing branch for Nvidia is the one where physical AI stalls entirely.

Read more

00:00 - Intro

00:01 - Navigating the AI Frontier: Innovations, Ownership, and Evolution

04:59 - Capturing Economics in the New Architecture

09:55 - Salesforce and Slack's Challenges

14:06 - Oracle's Position and Challenges

17:28 - OpenAI and Market Perceptions

20:25 - Anthropic's Growth and TAM Debate

24:42 - Vertical Application Layer and Startups

27:41 - Closing Remarks and Upcoming Announcements

Anthropic’s $2 trillion IPO case rests on more than model leadership. As frontier intelligence becomes cheaper and easier to substitute, the durable enterprise prize is the system of execution: the layer that captures organizational intent, carries authority through action, verifies results, retains context, and gets paid for accountable work. Update, September 2026: Salesforce has since taken this one layer deeper, post-training an open base into Koa, its own CRM model, and building the learning loop around the receipt.

Decoding Anthropic: Eleven Essays on Orchestration Economics — Decoding Discontinuity, August 2026

4.97MB ∙ PDF file

Download

Eleven essays tracking Anthropic through 2026 — from the Cowork launch and the $285B SaaSpocalypse to the Claude Code leak, the Fable shutdown, and the compute commitments behind the IPO. By Raphaëlle d'Ornano. 172 pages.

Download

TL;DR: The first AI disruption made software a tool. The second will make intelligence a tool.

The company that survives both will own the continuity of the execution loop. In February, I wrote that Anthropic would have to become an enterprise software company. Six months later, that instruction is obsolete. Most software is still being disrupted; a few companies are mutating; and OpenAI has become an enterprise business converging on Anthropic’s buyer. Together they expose what the SaaSpocalypse hid: the unit that made enterprise software a category is dissolving, and what replaces it is not a system of action but a system of execution. A system of action stops at the recommendation: it still leaves a human to click, approve, or route. A system of execution closes the loop - the agent reasons, decides, and does, updating records, triggering workflows, resolving cases, under the same governance and permissions that once applied only to humans - and it answers for what happened, under whose authority, with what outcome, and gets paid for it. Anthropic already runs that loop in software, on the old units, and is an ingredient everywhere else. Weeks before an IPO expected to price at $2 trillion or more, the question is whether the coding loop generalizes to parts of the economy with no compiler, and what Anthropic will still own when Claude is no longer scarce.

Subscribe now

Read more

As Anthropic gears up for its IPO - the first U.S. pure frontier lab to face the public markets - the S-1 could become public within weeks.

Ahead of that moment, I have collected eleven essays I wrote on the company this year into a single e-book: Decoding Anthropic (PDF, 172 pages). I wrote them as the events happened, and I wanted to print them exactly as they were published. Given the whirlwind Anthropic has set off, and how uniquely it now bears on the public markets, I thought it was worth putting the full arc in one place.

The arc, in sequence, is as follows:

1. Claude Cowork and the Enterprise Software Sorting - constrained autonomy, not maximum autonomy, is the deployable frontier (Jan 20)

2. The $285 Billion “SaaSpocalypse” Is the Wrong Panic - the selloff the market misread as a SaaS extinction event (Feb 10)

3. Decoding Anthropic’s $380 Billion Valuation - orchestration priced above raw intelligence (Feb 17)

4. Anthropic Data Reveals AI’s Real Job Impact - three mechanisms already compressing labor demand that the jobs data hasn’t caught (Mar 10)

5. The Claude Code Leak - 512,000 lines that made the orchestration architecture legible (Apr 7)

6. Anthropic’s Digital Labor Tax - the Metalayer, and the compute strain that turns every working agent into a tax on scarce inference (Apr 14)

7. King Claude: The Orchestration Moat in Operation - how Anthropic came to rule the first agentic cycle, and what keeping the throne requires (May 26)

8. Claude Is Building Claude - where scarcity migrates when the intelligence factory goes autonomous (Jun 9)

9. Claude Fable and the Barred Frontier - the week Washington proved the frontier can be closed on command (Jun 16)

10. The Red Queen’s Race - the treadmill the labs must escape, and why the IPOs are the test (Jun 23)

11. The Compute Trap 2.0: How Anthropic Refinanced Its Single Point of Failure - how the compute trap is being financed through new mechanisms (Aug 11)

The full book is below for paid subscribers, who also receive the scoring of the actual S-1 against these essays when the prospectus becomes public. The book opens to all readers on September 8th.

No post from me next week. I will be back the week of August 24 - with, I am sure, many exciting developments in AI land. Thank you again for your readership.

Anthropic Ebook - Decoding Discontinuity - August 13, 2026

4.97MB ∙ PDF file

Download

Download

A note on how to read it. The essays run chronologically: architecture first (essays 1-5), economics second (6-8), power third (9-10), and the balance sheet last (11). If you read only two, read the first and the last - January's claim that the sorting has begun that fired the “SaaSpocalypse”, and August's audit of what it cost Anthropic to win it.

Thank you again for your loyalty!

Raphaëlle

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.

TL;DR: Anthropic has largely solved the compute access problem that once threatened its growth. But the solution has created a new version of the “Compute Trap”: hundreds of billions of dollars of long-duration capacity commitments increasingly financed through SPVs, private credit, and bank guarantees. Its coming S-1 will reveal how much of that risk is truly fixed.

Subscribe now

Updated October 1, 2026: Reuters reported that Broadcom has agreed to provide up to $42B of financing against Anthropic’s $125.2B TPU commitments. That works out to roughly one-third of the commitment. Why does this matter? The supplier has moved beyond the role of vendor to becoming part of the capital stack. On a larger scale, that is the real evolution of the AI infrastructure boom. The first Compute Trap was: Can I get enough GPUs? Compute Trap 2.0 is: Can I finance the hundreds of billions of dollars of infrastructure I committed to before the revenue needed to support it has arrived? The risk hasn’t disappeared. It has migrated. The more interesting question now may be whether the AI labs are becoming less like software companies and more like highly leveraged infrastructure platforms with their suppliers increasingly underwriting their growth. That is a very different business model. And a very different valuation problem for investors.

Anthropic solved its compute access problem by building a far larger financing machine. Its IPO will test whether growth can absorb the commitments before those commitments begin to constrain it.

One of the greatest risks facing Anthropic a year ago was that it could not secure sufficient compute without becoming overly dependent on two shareholders who were also critical suppliers. I called compute a Single Point of Failure or the “SPOF”: The Harness was only as durable as the silicon underneath it.The silicon has a bottleneck of its own, further down the stack, in memory.

Lurking just below that SPOF was the broader problem of the Compute Trap: spend too little and you risk losing the frontier. Spend too much, and you risk overwhelming the economics of the business.

By May, when I revisited Anthropic in King Claude, the first half of that trap was no longer theoretical. Anthropic’s growth had outrun its available capacity so badly that it turned to former rival xAI, leasing the full 300 megawatts of SpaceX’s Colossus 1 for $1.25 billion per month. The contract bought Anthropic time, but not permanence: either side could terminate on 90 days’ notice.

The question had already begun to shift from compute access to what I called growth endurance: could Anthropic grow revenue fast enough, for long enough, to absorb a compute base contracted in advance against future demand?

Three months later, we have the next part of the answer.

Anthropic has largely solved the access problem by diversifying across more suppliers, which has vastly increased its capacity. But in solving the access version of the Compute Trap, Anthropic has exposed its financial version.

The company now sits atop hundreds of billions of dollars in announced compute commitments, backed by a dizzying array of complex financial instruments ranging from hyperscaler balance sheets to SPVs, private credit, and bank guarantees.

The original SPOF has been diversified. The underlying risk has been refinanced. As a result, the question is no longer simply: can Anthropic get enough compute to stay at the frontier? It is whether Anthropic can grow fast enough, for long enough, to support the financial architecture required to secure it.

Anthropic is preparing to go public carrying roughly $450 billion of announced compute commitments. The exact number matters less than what sits underneath it: how much is genuinely take-or-pay, how much is cancellable capacity, how much is lease liability, and how much remains contingent on future deployment. The S-1 should finally tell us whether Anthropic has built one of the most sophisticated financing machines in technology or simply converted an availability problem into a fixed-cost problem.

If SpaceX bought Anthropic time to address the access issues, then Volta may show what a more permanent financing model might look like.

Volta, a newly launched AI infrastructure company founded by former Brookfield executives, emerged from stealth in August with a six-year, $10 billion compute contract reported by Bloomberg to be with Anthropic. Behind it sits a sixteen-year infrastructure lease supported by roughly $1.3 billion of anticipated bank letters of credit.

Volta summarized its founding thesis in five words: compute is infrastructure and should be financed as such. That sentence captures what has changed over the past twelve months: the AI buildout has transformed from a technology or capacity story into a credit story.

And the largest private balance sheet at the center of that story is racing toward the public markets to test the viability of that strategy.

Anthropic’s Compute Commitments: $254B Confirmed, $450B Announced

To see how radically the risk changed, let’s return to the baseline.

Back in August 2025, AWS was Anthropic’s self-described “primary cloud provider and training partner,” backed by $8 billion from Amazon. Google supplied TPUs and held equity.

This represented an SPOF in the plainest sense. Twelve months later, that configuration looks radically different.

The confirmed floor is roughly $254 billion of compute commitments. Include the ~$200 billion five-year Google commitment reported by The Information in May and since detailed by the Financial Times as a Google-orchestrated web of chip and data-center financings, though never confirmed by the companies themselves, and the total reaches approximately $454 billion. I will treat that number as unconfirmed throughout this analysis. One caution on its shape: per the FT, roughly four-fifths of the ~$200 billion is the chips themselves - the ~$35 billion Compute SPV closed in June plus Broadcom’s filed $128 billion of TPU purchase commitments - with the balance in Google-backstopped data-center leases. It is a financing web tied to Anthropic’s lease payments, not an incremental cloud bill stacked on top of the structures described below; and because its data-center leg may overlap with other announced sites, the $454 billion aggregate is best read as an upper bound.

Three facts about that ledger matter for our analysis.

First, the velocity. Roughly 82% of all disclosed compute dollars - 69% counting confirmed deals only - were committed in the first seven months of 2026, just as the company’s revenue exploded.

Amodei told Dwarkesh Patel in February that the industry would build “10 to 15 gigawatts” this year. Anthropic has already announced roughly 14 gigawatts. Not all of that capacity comes online in 2026, but the scale is still extraordinary. One company has contracted for approximately one entire year of the global industry buildout.

Second, the silicon. Full diversification is still over the horizon, though the live fleet is already mixed: more than a million Trainium2 chips and Google TPUs serve Claude in production. The largest live increment of 2026 is Colossus, with 300 megawatts. Colossus is now switched on, and it’s pure NVIDIA. So are Volta’s Vera Rubin racks. Of those ~14 gigawatts, about 86% are non-NVIDIA: five gigawatts of Google TPUs, five gigawatts of Amazon Trainiums, and two gigawatts of AMD. The four accelerator families sum to 13.4 GW - nearly 90% non-NVIDIA - with the balance of the ~14 GW book at sites whose silicon is undisclosed. Anthropic now has four accelerator families in production or under contract, plus its own inference silicon in development.

Third, the financing. This is perhaps the most important aspect for this thesis. Through April, most of the money behind the megawatts sat on hyperscaler balance sheets: Google’s, Amazon’s, and Microsoft’s. Roughly three-quarters of the committed dollars were ultimately supported by companies with trillion-dollar balance sheets.

Since July, the financing architecture has shifted. Riot officially disclosed a 20-year, $9.1 billion lease for 191 MW at Rockdale with an unnamed “leading frontier AI lab,” running through June 2048; Bloomberg is reporting that the tenant is Anthropic. Riot also says the first 96 MW is due by December 2027 and the full 191 MW by June 2028.

TeraWulf and now Riot bring former crypto-mining infrastructure into long-duration AI leases; Volta adds an SPV supported by equity, debt, and anticipated bank letters of credit. Duration risk is migrating away from hyperscaler balance sheets and into a newer, thinner, and more leveraged financing layer.

That completes the transition: Anthropic bought faster, diversified more aggressively, and changed who ultimately finances the capacity. The first two helped eliminate the old single point of failure. The third explains why the Compute Trap survived.

How the Compute Trap Was Refinanced

What replaced the original SPOF is larger, more leveraged, and more financially entangled. The shift happened through three transformations:

1. From supplier concentration to obligation scale. The 2025 question was whether Anthropic could get enough compute. It can. The 2026 question is whether it can grow into what it has committed to buy. The headline stack is roughly $454 billion, or 9.7x a $47 billion revenue run-rate; the confirmed-only floor is 5.4x. Annualized, that falls to about 1.7x revenue, or roughly 0.8x excluding the unconfirmed Google deal. That single contract materially changes the picture, which is why the S-1 matters.

2. From supplier concentration to financial entanglement. If the reported $200 billion Google commitment is real, Google alone accounts for roughly half of Anthropic’s compute book. It also owns about 14% of the company, has committed up to $40 billion more in equity, and competes through Gemini, which could eventually become an open-source model. At least, that is not excluded. Amazon sits in a similar position, combining a large equity exposure with a decade-long compute commitment exceeding $100 billion. Diversifying among your two largest shareholders, both of whom are competitors, is less independent than the counterparty count suggests.

3. From availability risk to fixed-cost risk. A take-or-pay or lease obligation converts a variable cost into a fixed one. That is wonderful operating leverage while demand compounds. It runs precisely backward in an air pocket. We have already seen the mechanism flex. The Information reported in January was that gross margin compressed to roughly 40% after inference costs ran 23% over plan. The public-market preview is CoreWeave, whose S-1 disclosed that Microsoft accounted for 62% of its 2024 revenue, and whose anchor contracts come up for renewal in 2027. The market has not yet watched an AI take-or-pay stack get renegotiated in a downturn. It will.

One caveat to the diversification story: four accelerator families still converge on the same underlying supply chain. TPUs, Trainium, NVIDIA and AMD all depend on TSMC advanced packaging and the same three major HBM suppliers. The Series H cap table offers an important clue: Micron, Samsung and SK hynix all invested.

The bottleneck may no longer sit at the chip vendor level, but it has not disappeared. It has moved further down the stack.

Can Anthropic Outrun the Compute Trap?

The bear case doesn’t necessarily tell the whole story. Anthropic has built several genuine hedges against it.

First, the diversification is real.

The 2025 supplier hostage risk problem is gone. Anthropic now has ten-plus counterparties across four silicon families, with every other frontier lab running one or two. Its CFO says compute is allocated “fungibly” across training, inference, and internal workloads. This may be one of Anthropic’s most important competitive assets: the ability to arbitrage not just suppliers, but silicon architectures.

Second, Anthropic appears unusually compute-efficient.

Anthropic reached its current revenue scale on a fraction of OpenAI’s reported training spend, projects positive cash flow years earlier, and has crossed into reported (though still unaudited) operating profitability while its rival’s losses continue to widen toward an estimated $14 billion this year.

That matters because the strongest answer to a fixed-cost problem is not necessarily fewer fixed costs. It is enough revenue and enough efficiency to grow into them.

In King Claude, I called this growth endurance: the compute base is contracted in advance, so the only way to compress its burden is for revenue to grow faster than the commitments compound. That remains the strongest bull case for Anthropic.

At Anthropic’s current pace, the denominator is moving almost as quickly as the obligation stack. A company whose revenue run-rate moved from roughly $9 billion at the end of 2025 to $47 billion by late May can make commitments that look extraordinary today appear much less so surprisingly quickly.

But lately, the rise of Chinese open-source and the broader move away from the frontier has increased pressure on this aspect.

Third, the financing architecture may prove more scalable than it looks.

OpenAI and Anthropic face the same problem: frontier labs are not investment-grade, yet someone must finance hundreds of billions of dollars in infrastructure. Their answers are diverging. NVIDIA has reportedly discussed guaranteeing up to $250 billion in OpenAI’s data center leases; Anthropic’s newer structures rely on commercial bank credit. One is circular. The other is merely leveraged.

If Anthropic can turn compute financing from supplier subsidy into conventional bank credit, “compute financed as infrastructure” becomes a real market. But the aggregate risk remains: individually rational decisions can still produce a painful obligation stack if revenue arrives a year late.

That is the Compute Trap now: not whether Anthropic can secure capacity, but whether growth stays on schedule long enough to absorb it.

Inside the Volta Deal: $10B, a 16-Year Lease and $1.3B of Letters of Credit

Volta has raised $300 million in equity at a $2.4 billion valuation from Andreessen Horowitz, Altimeter, Azora, NVIDIA, and Michael Dell. The company will deliver NVIDIA Vera Rubin capacity from a hydro-powered campus in Tydal, Norway, which it does not own, using systems that have not been installed in data halls that are still being built.

Behind Volta’s sixteen-year, $4.7 billion lease obligation sits neither a hyperscaler guarantee nor a chip-vendor backstop, but approximately $1.3 billion of anticipated letters of credit arranged by affiliates of J.P. Morgan.

The structure rewards close reading, so here is the stack:

Anthropic (reported) → six-year, $10B compute contract → Volta Infra (a16z / Altimeter / Azora / NVIDIA / Dell equity; ~$5B Azora-led program of project equity and senior bank debt) → Volta Tydal AS (Norwegian SPV) → sixteen-year, $4.7B lease → Bitdeer’s Tydal campus (121 IT MW, hydro, PUE ~1.1) → NVIDIA Vera Rubin systems, supplied and integrated by Dell → wrapped by ~$1.3B of J.P. Morgan-arranged letters of credit (anticipated).

Two details explain why this deal is an arithmetic derivation that matters. To be clear, this is my calculation, not any party’s.

First, the lease’s headline rate of roughly $202 per kilowatt-month is a 16-year average with a 3% annual escalator. That implies about $233 million in first-year rent. Volta, with a year-one rate of about $160, can terminate without penalty after year ten, putting the lease's firmly committed portion at roughly $2.7 billion.

Now isolate years seven through ten: the period after Anthropic’s six-year contract expires but before Volta can walk away. The rent over those four years comes to approximately $1.17 billion. The letters of credit are $1.3 billion.

That suggests that the bank wrap is not simply a generic security deposit. It appears sized, almost to the dollar, to the portion of the lease without contracted customer revenue once Anthropic’s offtake expires. This is project finance’s mind arriving in AI compute. The LC is its signature. And it comes just as Nvidia answers in kind: on August 10 it announced memorandums with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize $500 billion of third-party capital for AI infrastructure - Huang pitching the chips themselves as “productive, investable infrastructure”.

That brings us to what is new.

Letters of credit in AI datacenter leases are not. CoreWeave posted a $50 million LC to Applied Digital in March. Nor is SPV-financed compute for Anthropic: in June, a little-reported ~$35 billion vehicle anchored by Apollo and Blackstone bought roughly one million Google TPUs to lease to Anthropic, with Broadcom guaranteeing the residual value of some $30 billion of the exposure. As per above, The FT has since identified that vehicle as the first tranche of the ~$200 billion Google-orchestrated program - the template for the further 3.5 GW of TPUs that Broadcom’s filings show it has agreed to buy for Anthropic’s use.

What appears new in Volta is the combination: a multi-billion-dollar deployment in which a non-investment-grade intermediary’s lease is supported by commercial-bank letters of credit covering roughly four and a half years of rent, while the landlord retains all of its equity and neither the chip vendor, hyperscaler nor AI lab guarantees the obligation.

That distinction matters because previous structures paid heavily for credit support. TeraWulf gave Google warrants for roughly 14% of the company in exchange for a $3.2 billion leaseback. Cipher gave Google about 5.4% for $1.4 billion. NVIDIA’s neocloud program charges a revenue share for its minimum-revenue guarantee. Broadcom takes residual-value exposure.

In each case, credit support carries a permanent economic claim within the structure.

Volta pays a bank a fee. A fee can be priced, refinanced, and terminated. It dilutes no one. If the structure closes and performs, bank credit becomes a materially cleaner alternative to paying suppliers or hyperscalers with equity. At that point, “compute financed as infrastructure” stops being a slogan and starts becoming a financing market (Bitdeer itself expects exactly that. On its August 10 earnings call, its chief strategy officer described bank-issued credit backing - rather than a chip vendor’s or a hyperscaler’s - as a structure he expects to catch on across the industry.) The funding-cost gap between the two worlds is already measurable: the FT finds data-center projects carrying Google’s backstop borrowing at a median 7.1% against 9.3% for neoclouds building around NVIDIA - what Jefferies calls a structural cost-of-capital disadvantage for the NVIDIA orbit.

The second revealing detail is the fine print around independence.

Andreessen Horowitz’s launch post emphasizes that Volta was assembled “without a hyperscaler or Nvidia backstop”. As a statement about credit, that appears correct: there is no NVIDIA guarantee and no Google-style wrap.

But place the credit structure beside the cap table. NVIDIA is an equity investor in Volta, which buys NVIDIA systems, to serve a lab in which NVIDIA has separately committed up to $10 billion.

No backstop. A shareholder instead. The credit may be independent. The economics are less so.

The Bank for International Settlements described the broader pattern in this year’s annual report: chipmakers and hyperscalers take equity stakes in AI labs or neocloud providers, which in turn enter multi-year commitments for chips and compute. It also warned that the third-party leaseback structures emerging beneath those equity loops are often poorly disclosed and can create risks around multiple claims on the same underlying assets.

Volta does not eliminate that circularity. It changes its form. The supplier guarantee disappears; bank credit takes its place. The circularity moves down a layer, becomes smaller relative to the transaction, and acquires a more conventional capital structure.

Share

What Anthropic’s S-1 Must Disclose

Anthropic confirmed on June 1 that it had confidentially filed its S-1. If the reported October timetable holds, the public filing will do something private-market reporting cannot: put the economics of the compute buildout into audited financial statements and securities-law disclosure.

When the public S-1 lands, it will be critical to watch both of these aspects:

1. The purchase-obligation footnote. This will be the first audited presentation of what the ~$454 billion consists of: leases, unconditional purchase commitments, and contingent capacity options. SpaceX already showed why this distinction is everything. In King Claude, I described Colossus as tactical rather than structural precisely because its 90-day termination right put an escape hatch underneath the headline commitment. What was initially reported as a $45 billion, three-year commitment turned out, per Musk’s clarification, to have an 180-day initial term followed by mutual 90-day cancellation rights - potentially leaving only about $7.5 billion firmly committed. The S-1 will tell us how much of Anthropic’s broader compute stack is SpaceX-shaped: enormous in headline value, but flexible in duration - and how much is genuinely fixed and take-or-pay. If most of it is the former, the bear case shrinks dramatically. If it is the latter, it doesn’t.

2. The related-party map. Amazon’s actual ownership percentage remains undisclosed. Google’s post-IPO ownership limits matter. So do the contractual relationships connecting the two. The filing should finally map the circular flow cleanly. The individual relationships are already known. What is missing is the consolidated picture.

The Compute Trap Now: From Compute Access to Growth Endurance

Anthropic has solved compute access far more aggressively than I expected one year ago. It has diversified across suppliers, silicon architectures, and financing sources. The original supplier SPOF is largely gone.

But the Compute Trap is not.

The constraint has moved from access to economics. Anthropic must secure capacity years before the revenue that will consume it is certain. The commitments are increasingly fixed; the revenue is not. What I called growth endurance in King Claude remains the central financial bet: revenue must grow faster than the obligation base long enough for the infrastructure to amortize.

What has changed is who carries that bet.

A year ago, compute risk sat largely on the balance sheets of Anthropic and two hyperscalers. Today it is distributed across hyperscalers, chipmakers, memory suppliers, landlords, SPVs, private credit and now commercial banks. Anthropic has not eliminated the risk so much as built a financing system capable of absorbing it.

The system may work. Bank credit is cheaper than equity, multi-silicon gives Anthropic flexibility, and fast revenue growth could make today’s commitments look manageable surprisingly quickly.

But everything depends on time. If adoption compounds, the fixed-cost base amortizes. If growth slips, operating leverage reverses.

That is why the S-1 matters. It will show how much of Anthropic’s compute stack is SpaceX-shaped - large but flexible - and how much is truly fixed and take-or-pay.

Anthropic solved the first problem: securing enough compute to stay at the frontier. The next is harder: can adoption amortize the infrastructure before the infrastructure constrains adoption?

As Amodei put it: “If you’re off by only a year, you destroy yourselves”.

The S-1 will tell us how much room Anthropic has bought itself.

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.