OpenAI and Anthropic’s Capex Dependence: The Trillion-Dollar Compute Bill
Key Takeaways
- OpenAI Anthropic capex dependence is the defining financial fact of the AI boom: the two leading labs need hundreds of billions of dollars of computing, and the money comes largely from the same giants that sell them that computing.
- The IMF's October 2026 Global Financial Stability Report examined 14 disclosed AI deals since 2023 and found about $300 billion of equity, about $650 billion of debt and $1.4 trillion of future revenue commitments — a loop it calls "wrong-way risk".
- Anthropic's IPO filing of 26 October 2026 discloses compute commitments of more than $800 billion through 2045 — updated in its amended prospectus to up to $895.4 billion — including at least $110 billion to Amazon, about $114.5 billion to Google and at least $44.9 billion to Microsoft, much of it "regardless of actual usage".
- OpenAI's funding round was cut from $30 billion to about $20 billion at a $1.25 trillion valuation in late October 2026, with about $11 billion committed, according to Financial Times and Wall Street Journal reporting — while its compute bill runs into the hundreds of billions.
- The dependence runs both ways and the numbers are improving: Anthropic's compute spending fell from more than 10 times revenue in 2023 to 55% of revenue in the first half of 2026, but the absolute commitments keep growing faster than the cash to pay for them.
The two companies building the world's most capable AI models have the same problem, and it is not intelligence. It is rent. OpenAI and Anthropic do not own the computers their models run on, not at the scale they need. They rent computing power from the largest technology companies on earth — Microsoft, Amazon, Google and Oracle — and they pay for it with money those same companies, and their peers, invested in them. The supplier is the shareholder. The shareholder is the landlord. And the rent, by the labs' own filings, runs into the hundreds of billions of dollars over the next two decades.
That circular shape is why searches for the labs' spending spiked in November 2026, and why the question behind them is sharper than "is AI a bubble?" The question, in the terms readers are actually searching, is openai anthropic capex dependence: who pays for the computers, and what happens if they stop. The International Monetary Fund put it in formal language in October 2026: when the firms funding AI are also its biggest customers and creditors, trouble at one end of the loop hits both ends at once. The IMF calls that "wrong-way risk". This article lays out the numbers behind it — OpenAI's trimmed funding round, Anthropic's IPO filing, the hyperscalers' own deteriorating cash flow — attributes every figure to its source, and separates what is disclosed from what is merely reported.
OpenAI Anthropic capex dependence: the IMF's $1.4 trillion loop
The clearest framing of the problem comes not from a market commentator but from the International Monetary Fund. Its October 2026 Global Financial Stability Report, titled *Resilience, Risk, and Transformation* and released on 13 October 2026, warned that weaker or less durable AI investment and profitability could trigger a correction in AI-linked equities and credit, one of three risks it said could abruptly tighten financial conditions worldwide.
The report's most striking section examined what it called circular financing. IMF researchers looked at 14 publicly disclosed transactions since 2023 and found that hyperscalers — the large cloud companies — had put about $300 billion of equity into financially weaker AI firms, supplemented by about $650 billion in debt, in exchange for $1.4 trillion in future revenue commitments through services or chip orders. About $320 billion of the debt sits with the investee companies themselves, including debt backed by investor guarantees, contingent financing and other off-balance-sheet obligations. The mechanism is simple to describe and hard to unwind: the investor funds the AI company, and the AI company promises to spend the money back with the investor, buying cloud computing or chips.
The IMF's warning is about what happens if the loop breaks. Its first chapter warns that distress at a key investee firm could present significant "wrong-way risk": the equity and debt investments lose value at the same time as the revenue streams tied to them are impaired. If a lab cannot pay for its compute, the cloud provider loses both an investment and a customer in the same moment. The report also found that hyperscalers' share of global investment-grade bond issuance has risen to 10 percent since 2024, up from less than 1 percent, while their free cash flows decline under the weight of data-centre spending. And it modelled a scenario, in the IMF's World Economic Outlook, in which an unwinding of AI investment reduces global GDP, investment and consumption by as much as 0.4, 1.3 and 0.6 percent respectively by 2028. That is a scenario, not a forecast — the IMF does not predict a crisis — but it is the first time a major global institution has put numbers on how large the loop has become.
Two cautions belong with the headline figures. First, the IMF names no companies in its analysis of the 14 deals, so the $300 billion, $650 billion and $1.4 trillion figures describe the pattern across the industry, not the exposure of any single lab or supplier. Second, $1.4 trillion is a promise of future revenue, not cash received. It is what AI companies have committed to spend, over years, if the contracts run their course. Treat it as a measure of dependence, not as money already spent.
Anthropic: $800 billion of promises, filed with the SEC
If the IMF described the pattern in aggregate, Anthropic's IPO filing put a single company's version of it on the public record. Anthropic publicly filed its Form S-1 with the US Securities and Exchange Commission on 26 October 2026, planning to list on the Nasdaq under the ticker ANT. The filing — more than 300 pages, with a risk-factors section running over 90 pages — is the first full public look at the finances of a frontier AI lab, and the compute numbers in it are the story.
Anthropic calls compute a "critical input", and its signed commitments total more than $800 billion, extending as far as 2045. An amended prospectus filed in early November 2026 raised that figure further: cumulative data-centre spending commitments of up to $895.4 billion through 2045, according to reporting of the amended filing. The commitments are not a single contract. They are a web of deals with the same companies that fund Anthropic, and the filing spells out the terms with unusual bluntness.
With Amazon, Anthropic committed in April 2026 to spend at least $110 billion between May 2026 and April 2036, "regardless of actual usage" — the filing's own words, quoted in coverage of the S-1. Amazon had already invested $8 billion in Anthropic between 2023 and 2025, and in April 2026 agreed to invest up to a further $25 billion. The same agreement secured up to five gigawatts of capacity to train and deploy Claude models, and Anthropic committed to spend more than $100 billion on Amazon Web Services technologies over the next decade, including Amazon's custom Trainium chips.
With Google, the pattern repeats at a similar scale. Anthropic committed in March 2026, in an agreement amended between May and August 2026, to spend at least approximately $114.5 billion between April 2026 and July 2033, again regardless of actual usage. It also expects to procure additional TPU compute from Google for delivery between 2028 and 2031, with aggregate future payments estimated at approximately $125 billion to $150 billion. Google, meanwhile, had invested roughly $3 billion in Anthropic across earlier rounds, and in April 2026 the companies said Google planned to invest a further $40 billion. Google is also, in Anthropic's own telling, one of only two specific competitors named in the prospectus — alongside OpenAI — even as it builds Gemini, a rival model family.
With Microsoft, Anthropic committed to spend at least $44.9 billion between November 2026 and May 2033, under an agreement that cannot be cancelled except for an uncured material breach by Microsoft or an undelivered tranche. Microsoft announced plans in November 2025 to invest $5 billion in Anthropic, bringing Claude models to Azure. The relationship is doubly tangled because Microsoft is also the largest investor in OpenAI, Anthropic's biggest rival, with about a 27% stake.
Add the chip suppliers and the total climbs further. Broadcom and Google jointly develop TPUs and lease them to Anthropic under a five-year lease totalling $161 billion, with up to an additional $150 billion of TPU spending expected from 2028. Anthropic also committed to spend more than $20 billion on AMD chip systems, with AMD holding an option to invest up to $5 billion back into Anthropic as chips are delivered. Nvidia, which supplies the GPUs that power Anthropic's models, said it would commit up to $10 billion to Anthropic. And the filing discloses a contract with SpaceX worth up to $120.5 billion running to 2029, with monthly fees of around $3.1 billion once all computing capacity comes online — a contract either party may terminate with 90 days' notice from early next year.
Set against those commitments, Anthropic's own finances show a company growing extraordinarily fast and still spending more than it earns. Revenue was $41 million in 2023, $386 million in 2024 and $4.585 billion in 2025. In the first half of 2026, revenue reached $16.286 billion — more than 14 times the same period a year earlier — alongside a net loss of $80.7 billion, of which $72.4 billion came from non-cash mark-to-market adjustments on convertible notes and other financial instruments. The operating loss, a cleaner measure of the underlying business, narrowed to $693 million. In the first half of 2026, Anthropic spent $8.95 billion on computing and infrastructure, $3.15 billion on personnel and $3.17 billion on sales, marketing and partnerships.
The direction of travel is genuinely encouraging for Anthropic, and the filing shows it. Compute spending exceeded revenue by more than 10 times in 2023, stood at 653% of revenue in 2024, 160% in 2025, and 55% in the first half of 2026. The Financial Times, cited in coverage of the filing, put it more vividly: in the first quarter of 2025, every $1 of Anthropic's revenue corresponded to more than $2.50 in compute and infrastructure spending; by the second quarter of 2026, that figure had dropped to under 50 cents. Revenue is catching up with the compute bill. The problem is the size of the bill still to come: even at 55% of revenue, a commitment of up to $895.4 billion through 2045 is roughly 55 times Anthropic's entire first-half 2026 revenue, and it must be paid whether or not usage — or revenue — keeps pace, because the largest contracts are fixed regardless of actual usage.
The financing behind those commitments is itself fragile. The $63 billion chip-financing deal known as Project North Star, arranged by Broadcom and anchored by Blackstone, is the mechanism funding Anthropic's next batch of chips — and Blackstone has now told investors it expects Anthropic to receive a double-B, or junk, credit rating after its IPO, rather than the investment-grade rating it had previously expected by year-end, according to the Financial Times. The junior debt resting on Anthropic's own credit already pays more than 10%, up from about 8.5% in June. A junk rating shuts out the cheapest debt — pension funds and insurers that can only hold investment-grade bonds — at exactly the moment Anthropic needs to borrow at scale. Under the North Star lease alone, Anthropic has committed $125.2 billion, with a $42 billion senior loan covering about a third of it, and it may issue up to $42 billion in convertible notes to Broadcom to pay rent, with interest rates rising from 9% to 13% and conversion terms worsening if revenue targets are missed. Anthropic is not short of cash today — reporting of the amended prospectus puts its liquidity at $146.6 billion as of 30 September 2026, and the IPO itself could raise as much as $100 billion — but as one analysis of the filing put it, that covers the next few years, not the next twenty.
OpenAI: a smaller round, a bigger bill
OpenAI's version of the same dependence looks different in form but identical in substance. Where Anthropic's exposure is visible in an SEC filing, OpenAI's is visible in its fundraising — because a private company that needs hundreds of billions of dollars of computing must keep raising money to pay for it, and in late October 2026 the market stopped giving it quite as much as it asked for.
The Financial Times reported on 30 October 2026 that OpenAI had cut the target for its latest funding round from $30 billion to about $20 billion, and lowered the valuation it was seeking from $1.4 trillion to $1.25 trillion. The Wall Street Journal reported that about $11 billion had been committed by the 31 October deadline — less than half the original ask — and described the response from some large investors as tepid. The first close is expected to come in below the $20 billion target. Investors from the United Arab Emirates are expected to lead the round, according to TechCrunch's account of the reporting. OpenAI did not comment on any of the figures, and every number in this paragraph comes from press reports citing unnamed sources rather than from the company itself. That matters: a target is not a result, and a round that stands at $11 billion in October can close at $20 billion in December.
Even trimmed, the round would be enormous. A $1.25 trillion valuation would make OpenAI the most valuable private company ever, up from the $852 billion post-money value of its $122 billion round in March 2026. And the Financial Times attributed the trim to an AI stock sell-off — the Nasdaq 100 fell 3.5% on the Friday of the deadline week — and to competition from Anthropic's listing, which was drawing the same pool of capital. The round is a bridge to an IPO that reporting suggests is likely in the first half of 2027. What the trim shows is not that OpenAI is short of money. It shows that the price of money for AI labs has risen, and that investors are no longer writing whatever cheque the labs ask for, at whatever valuation they name.
The reason OpenAI needs the money is its compute bill, and it is larger than Anthropic's in the near term. OpenAI's Stargate programme, announced on 21 January 2025 as a joint venture with Oracle, SoftBank and MGX, was a $500 billion, four-year commitment to build 10 gigawatts of AI capacity in the United States, with $100 billion to be deployed immediately. Its largest single contract is with Oracle: an agreement for up to 4.5 gigawatts worth more than $300 billion over five years, starting in 2027. On top of that, OpenAI agreed to buy $250 billion of Azure cloud services from Microsoft in its 2025 restructuring — Microsoft having invested about $13.8 billion in OpenAI since 2019 for its roughly 27% stake — and expanded its Amazon commitment from $38 billion to $100 billion, after Amazon agreed in February 2026 to invest up to $50 billion in OpenAI, starting with $15 billion and a further $35 billion conditional on a stock-market listing or a technical milestone. Nvidia's much-publicised commitment of up to $100 billion to OpenAI turned out to be a letter of intent; in the March 2026 round, Nvidia actually put in $30 billion. DatacenterDynamics, citing Bloomberg, reported on 14 October 2026 that OpenAI now projects about $750 billion of compute spending through 2030, up from an earlier estimate of roughly $600 billion.
Set against that, OpenAI's revenue is large and growing, but an order of magnitude smaller than the commitments. Reuters reported that OpenAI told investors it had reached $50 billion in annualised revenue and expected $70 billion by the end of the year. Fifty billion dollars of annualised revenue against $750 billion of projected compute spending through 2030 is the gap the funding rounds exist to fill — and it is why a $10 billion cut to one round, less than two months of the company's projected compute spending at that run rate, still mattered enough to move markets. Our earlier coverage of the round itself, in the report on OpenAI's funding round cut to $20 billion, covers the fundraising in detail; this article is about what the money is for.
There is also a gap between what OpenAI has contracted and what has actually been built. OpenAI said in April 2026 that it had passed its 10-gigawatt US capacity goal, with more than 3 gigawatts added in the previous 90 days — but most of that capacity is contracted, not built. The largest dated, delivered figure published is far smaller: Oracle's fiscal first-quarter 2027 earnings call in September 2026 said six of eight buildings at the Abilene, Texas flagship campus, 618 megawatts and 75% of the campus, had been delivered and accepted. Roughly 6% of the headline capacity is verifiably online. The rest runs on 2027 and 2028 dates, and Oracle has already issued a force majeure notice on its New Mexico campus, citing potential power delays. The compute is coming, but it is not here yet — and OpenAI is paying, in commitments and in fundraising, for capacity that does not exist.
The other side of the loop: the hyperscalers are stretched too
The dependence is not one-directional. The cloud giants need the labs as much as the labs need them: a data centre with no customer is a stranded asset, and the labs are the largest customers there are. But the hyperscalers' own finances show the strain of building for them. Alphabet, Amazon, Meta and Microsoft have now outlined more than $740 billion in capital expenditure this year, and the market's reaction shifted during the autumn of 2026 from applause to dread.
The third-quarter 2026 results, reported in late October, tell the story in cash flow. Alphabet reported negative free cash flow of $9.1 billion in the quarter, its second consecutive quarter in the red, while capital expenditure reached $51.4 billion and guidance points to as much as $205 billion for the year. Amazon posted negative free cash flow of $42.4 billion over the past twelve months, against $14.8 billion positive the year before, with $220 billion of capital expenditure planned. Meta was negative $2.1 billion in the quarter, with guidance up to $145 billion. Microsoft was the outlier: $18.6 billion of positive free cash flow and $45.2 billion of capital expenditure, below its own $50 billion projection, which it blamed on third-party capacity arriving late. S&P Global Market Intelligence counted $326 billion raised in debt during 2026 by Microsoft, Alphabet, Amazon, Meta and Oracle combined — triple last year's total. Bank of America data, cited in coverage of the funding squeeze, put AI-related US corporate bond issuance at $655 billion by mid-October 2026, against $394 billion in all of 2025, with nearly $1 trillion expected next year.
Reuters' analysis of LSEG consensus forecasts found that capital expenditure at Microsoft, Alphabet, Amazon, Meta and Oracle is expected to exceed their free cash flow by 2027, with every additional $1 of operating cash flow corresponding to about $1.57 of new capital expenditure. What the companies earn is not enough to fund what they are spending, so the gap can only be filled by issuing stock or debt. That is the structural fact behind the autumn's market wobble: the buildout is no longer self-financing, and its continuation depends on credit markets staying open and cheap. Nvidia, meanwhile, has extended roughly $170 billion in investments and credit support to neoclouds and large customers, becoming a creditor to the ecosystem it supplies — another strand of the same loop.
None of this means the spending is irrational. Data centres are real assets, the models running in them earn real revenue, and both labs' revenue curves are steep. Anthropic's compute-to-revenue ratio has fallen by more than an order of magnitude in three years. OpenAI's annualised revenue has risen from roughly $40 billion in August to about $50 billion by October, according to Bloomberg reporting cited in coverage of the round. The question the market started asking in October 2026 is narrower and harder: whether the revenue will arrive fast enough, and in large enough amounts, to service the specific debts and fixed commitments already signed — the "regardless of actual usage" contracts, the 9%-to-13% convertible notes, the $63 billion chip lease — before the credit that funds them gets more expensive.
What would break, and what would not
It helps to be precise about what "dependence" means here, because the word is doing a lot of work. OpenAI and Anthropic are not insolvent. Anthropic holds more than $100 billion in liquidity and may raise as much as $100 billion more in its IPO. OpenAI, even after trimming its round, is raising $20 billion at a $1.25 trillion valuation. Both companies' revenues are growing at rates most businesses never see. If the story ended there, there would be no story.
The dependence is structural rather than immediate. First, neither lab can train or serve a frontier model without its suppliers' hardware, and the suppliers are a small group — Nvidia for GPUs, Broadcom and Google for TPUs, the big three clouds for data centres. Anthropic's own prospectus says its relationships with Microsoft, Amazon, Google and Broadcom "create dependencies and potential vulnerabilities". Second, the contracts are fixed and long-dated: ten-year, "regardless of actual usage" commitments that must be paid even if a model underperforms, a market turns, or a cheaper way of delivering AI capability appears — the exact risk the IMF flagged when it noted growing evidence that AI capabilities can be delivered at lower investment levels than assumed. Third, the financing is layered: equity from the supplier, debt raised against the supplier's guarantee, and revenue promised back to the supplier, so that a single failure impairs the asset, the loan and the income stream at once.
What would actually break the loop is not a single bad quarter. It is a sequence: revenue growth slows, a fixed commitment comes due, refinancing costs more because the credit rating has slipped, and the supplier — itself carrying the debt that funded the buildout — pulls back. The most likely outcome, on the evidence available in November 2026, is neither collapse nor vindication but a slower, more expensive buildout: more debt, higher rates, smaller rounds, and a longer wait for the revenue to cover the rent.
The bottom line
OpenAI Anthropic capex dependence is not a slogan. It is a set of signed contracts: more than $800 billion of compute commitments at Anthropic, rising to up to $895.4 billion in its amended prospectus; about $750 billion of projected compute spending at OpenAI through 2030; $1.4 trillion of future revenue commitments across the 14 circular deals the IMF examined. The money flows in a circle — from hyperscaler to lab, from lab back to hyperscaler as rent — and the circle only works while revenue grows fast enough and credit stays cheap enough to keep it turning. In October 2026, both of those conditions started to look less certain: OpenAI cut its round by a third, Anthropic's backers began pricing its debt as junk, and the hyperscalers funding the buildout slipped into negative free cash flow. The labs are not failing. They are dependent — on each other, on their suppliers, and on a credit market that has only just started charging them for the privilege. The cheapest AI models in history, as we covered in our look at OpenAI and Anthropic's cheaper AI models and their real 2026 costs, are cheap for users precisely because someone else is paying the capital cost. The question this article leaves open is how long that someone else is willing to keep paying.
Sources
- IMF, Global Financial Stability Report, October 2026, *Resilience, Risk, and Transformation* (released 13 October 2026), via Unite.AI — https://www.unite.ai/imf-warns-ai-market-correction-could-trigger-global-financial-turmoil/
- Anthropic Form S-1, filed 26 October 2026 (Nasdaq: ANT), via Invezz — https://invezz.com/news/2026/10/26/anthropic-ipo-ai-labs-investors-include-rival-google-vcs-also-backing-openai/
- Anthropic amended prospectus, up to $895.4 billion in commitments through 2045, and Blackstone junk-rating expectations, via HTX — https://www.htx.com/news/the-hidden-risks-behind-anthropics-nearly-1-trillion-bet-i8vbADw6
- OpenAI funding round cut ($30bn to ~$20bn, $1.25tn valuation, ~$11bn committed), Financial Times and Wall Street Journal reporting via The Inference — https://theinference.org/article/openai-trimmed-its-round-to-20-billion-at-a-1-25-trillion-valuation-with-11-billion-committed
- Hyperscaler capital expenditure and free cash flow, Q3 2026, and debt issuance, via n8nlab — https://n8nlab.io/news/ai-compute-credit-market-builders-pay-interest
- Anthropic's $63 billion Broadcom chip financing — https://aiera.blog/anthropic-broadcom-63-billion-chip-financing/