Anthropic’s amended IPO filing lists about $825.6 billion of compute and infrastructure commitments that fall due after the next 12 months, by our arithmetic, against about $3.5 billion of booked customer revenue it expects to recognise after that same 12-month window. That is about 239 times, from figures in the Form S-1/A filed on 2 November.
Theory Ventures investor Tomasz Tunguz put that gap at the centre of an essay on X on 3 November, titled “The Financial Engineering Essential for AI in 2027”. His conclusion: “In 2027, the winners will be the labs that match the duration of their customers’ contracts to the duration of their compute.”
Business Pill · Backlog Explained
A Business Engineer explainer of backlog: revenue that is signed but not yet earned. It teaches the general idea only and does not discuss Anthropic’s figures.
The key insight: As we read it, the filing already concedes Tunguz’s premise. Anthropic says short-term compute “currently costs significantly more than comparable compute procured under our multi-year commitments”, and that much of its revenue “is not subject to long-term contractual commitments”.
Tunguz wrote that GPUs “have increased in rental price per hour from $3 to $12 to $13 on short-term contracts, pushing Anthropic’s blended cost per GPU-equivalent hour from $3.10 to $5.65 (within a $3.45 to $8.29 confidence band), according to data from the updated Anthropic S-1/A.”
He described the mechanism this way: labs rent accelerators “on multi-year contracts that fix prices around $2 to $4 per hour”, and “when that long-term capacity is exhausted, AI vendors must buy access to GPUs in the short-term burst market, where prices have surged.”
His fix is on the revenue side: “selling a Fortune 500 buyer a multi-year minimum commitment at a locked-in rate to match the duration of the GPU lease underneath it.”
Anthropic’s compute and infrastructure commitments against its remaining performance obligations (booked, non-cancelable customer revenue), split by when they fall due. Source: Anthropic Form S-1/A, 2 November 2026. Commitments include deals signed after 30 September; obligations are as of 30 September. The $825.6 billion, the ~$10.4 billion / ~$3.5 billion split and the 239x are our arithmetic.
Where His Per-Hour Prices Come From
The per-hour figures are Tunguz’s own estimates. His chart, credited to Theory Ventures, is headed “Estimated cost per hour by supplier contract” and gives each contract a range. It puts the $5.0 billion of short-term compute, on a bar labelled with a range of 175k to 325k, at $12.94 an hour, and the amended SpaceXAI deal at $13.07.
The chart prices a “GPU-equivalent hour” across Nvidia GPUs, Google TPUs and Amazon Trainium contracts, not one chip. We searched the S-1/A text for hourly prices and GPU counts and found none; the filing gives contract totals and terms instead. For SpaceXAI, for example, it gives about $3.1 billion a month once all capacity is delivered.
A market reference point: SemiAnalysis chief executive Dylan Patel put a one-year Nvidia H100 rental at about $2.80 an hour, up from a trough of $1.70 in October 2025, as we reported from Patel’s H100 rent data.
What the Filing Says About Duration
The compute side is long. The S-1/A says Anthropic has about $895.4 billion of commitments, about $447.5 billion of them expected before the end of 2030 and substantially all the rest between 2031 and 2045. About $69.8 billion is due in the 12 months after 30 September.
Its operating leases, for offices as well as compute and infrastructure, carry a weighted-average remaining term of 10.3 years; its finance leases for compute and infrastructure, 4.8 years. The filing says most compute is contracted “under multi-year commitments that we typically procure 12 to 18 months in advance of availability.”
The customer side is short. Remaining performance obligations, the revenue customers have committed to under non-cancelable contracts, were $13.8 billion at 30 September, and Anthropic expects to recognise about 75% of that within 12 months. Subscriptions run for a term the filing calls “generally 12 months or less”.
The risk factors say it in one line: “our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years.” Consumption-based revenue was $26.3 billion, or 79% of revenue, in the nine months to September.
How We Worked Out the 239x
Compute: $895.4 billion minus the $69.8 billion due within 12 months leaves $825.6 billion due after that. Customers: about 25% of the $13.8 billion of remaining performance obligations is about $3.45 billion due after 12 months. $825.6 billion divided by $3.45 billion is about 239. All three steps are our arithmetic.
Inside the next 12 months the gap is narrower: $69.8 billion of commitments against about $10.4 billion of committed revenue, about 6.7 times. Neither figure counts usage revenue that is not under a non-cancelable contract, which is most of what Anthropic earns.
The two sides are not measured on the same date. The commitments include deals signed after 30 September; the obligations are as of 30 September. The filing also says about $213.8 billion of its contractual obligations is subject to termination rights or contingent conditions before commencement.
Where Short-Term Compute Shows Up
The filing ties the third-quarter operating loss of $567 million mainly to “$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments”, and to $2.0 billion of non-cash donation-matching charges. Adjusted operating margin fell from 19% to 15%.
The company says it expects its use of short-term compute “to continue to increase in the near term”. It also says its own build-out, Anthropic Direct, lets it “backfill the short-term compute we procure during periods of high demand with more cost-effective, long-term compute as our dedicated infrastructure comes online.”
The SpaceXAI arrangements sit at the short end. After initial non-cancelable periods ending in January and March 2027, either party can cancel each agreement on 90 days’ written notice, according to the filing.
The Other Side of the Trade
CoreWeave, which rents out GPU cloud capacity, books its customers for longer. Its 10-Q for the quarter to 30 June, the latest on EDGAR when we checked on 3 November, reports $103.7 billion of remaining performance obligations: 41% due within 24 months, 39% in months 25 to 48 and the rest in months 49 to 78.
By our arithmetic, about $61.2 billion of CoreWeave’s obligations falls after the first two years. Its operating leases carry a weighted-average remaining term of 12 years and its finance leases 5 years, per the same filing.
What Tunguz Says Labs Are Doing
Tunguz lists three responses: efficiency gains, tighter subscription allowances, and “pushing enterprise buyers from serverless API consumption into multi-year committed contracts.” He cites Opus 5.5’s “40% reduction in cost per task over Opus 5” and OpenAI’s move to pool ChatGPT and Codex usage into one shared token limit.
The S-1/A describes the third route already under way: “A growing number of enterprise customers agree to non-cancelable annual or multi-year contracts with minimum spend commitments in exchange for a committed price list.”
Connecting the Coverage
Our related pieces on Anthropic’s filing and GPU prices:
As we read the numbers, the gap is mainly one of timing: inside the next 12 months, commitments are about 6.7 times committed revenue; past 12 months, about 239 times, both by our arithmetic.
As we read the CoreWeave comparison, the GPU seller already books on the long side: 59% of its $103.7 billion of remaining performance obligations falls after two years, by our arithmetic from its 10-Q.
As we read Tunguz’s chart, his widest ranges sit on the short-term rows: $6.97 to $12.94 an hour for the third-quarter burst and $7.78 to $13.07 for the amended SpaceXAI deal, against $1.45 to $2.65 for the Broadcom TPU lease.
Tomasz Tunguz, Theory Ventures, on X
“A doubling of GPU cost must be met by a doubling of model efficiency, a doubling of price to the customer, or a combination of the two to preserve margins.”
Three Implications
FOR IPO INVESTORS The S-1/A says the expense of new compute is recognised in advance of the full revenue it will support, and that operating margins may compress in those periods.
FOR ENTERPRISE BUYERS Anthropic says a growing number of enterprise customers sign non-cancelable annual or multi-year contracts with minimum spend commitments in exchange for a committed price list.
FOR GPU CLOUDS CoreWeave’s 10-Q recognises obligations out to month 78, while Anthropic’s SpaceXAI agreements can be cancelled on 90 days’ notice after initial periods ending in early 2027.
The framework states: “For two years, AI was a CIO line item and a CEO talking point. In 2026 it became a CFO problem, because the economics finally turned legible — and unforgiving.”
As we read it, duration matching is that CFO problem inside a lab: Anthropic’s filing sets compute commitments running to 2045 against customer obligations mostly due within a year. The framework does not discuss Anthropic’s S-1/A or Tunguz’s essay.
What Is Not Established
The per-hour prices and the $5.65 blended cost are Tunguz’s estimates, not figures in the filing. The S-1/A does not split the 25% of obligations due after 12 months by year, and it does not give the length of its enterprise contracts. CoreWeave’s figures are three months older than Anthropic’s.
The $825.6 billion, $3.45 billion, 239 times, 6.7 times and $61.2 billion figures are our arithmetic. We did not contact Anthropic, Theory Ventures or CoreWeave.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Anthropic’s own filing describes the mismatch Tunguz writes about: compute commitments running to 2045 against revenue that is mostly pay-as-you-go. On the filing’s figures, the compute due after the next 12 months is about 239 times the booked customer revenue due after it, by our arithmetic.
A note on sourcing. This piece is based on Anthropic’s Form S-1/A filed with the SEC on 2 November 2026, Tomasz Tunguz’s X article of 3 November 2026 and its Theory Ventures charts, and CoreWeave’s 10-Q for the quarter to 30 June 2026. Per-hour GPU prices and the blended cost per GPU-equivalent hour are Tunguz’s estimates, not figures in the filing. The $825.6 billion, $3.45 billion, 239 times, 6.7 times and $61.2 billion figures are our arithmetic. We did not contact Anthropic, Theory Ventures or CoreWeave.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
