The labs' compute obligations: how big, why they signed, what breaks

OpenAI and Anthropic have signed well over a trillion dollars of compute purchase commitments between them — with hyperscalers and neoclouds, in contracts running 5 to 20 years. The question Graham asked: how large is this really, why did the labs agree to sign at this scale, and what happens if the obligations turn out to be wrong?

OpenAI has signed ~$710B+ of announced compute commitments and Anthropic ~$355–385B — the largest private procurement program in history. The bulls say capacity-constrained demand plus scaling revenue covers it; the bears say it's a circular, duration-mismatched vendor-financing loop. The truth sits in which contracts are take-or-pay and whether the revenue run rates keep doubling.

OpenAI

Anthropic

Start with the labs' own line, because it is internally consistent. Altman: 'The risk to OpenAI of not having enough computing power is more significant and more likely than the risk of having too much.' OpenAI rate-limits products and delays features today; mega-projects take years to bring online, so capacity has to be contracted ahead of demand. 'Everything starts with compute. Compute infrastructure will be the basis for the economy of the future.' Around GPT-5 (Aug 2025) OpenAI doubled its compute fleet in five months — and Brockman, on Odd Lots, gave the operating version: 'compute is revenue.'

Anthropic's version is cost control as much as growth. Amodei (Apr 2026): 'we need to build the infrastructure to keep pace with rapidly growing demand.' The Information's reporting adds the margin motive: Anthropic's inference costs ran 23% over budget in 2025, with gross margins at 40% against a 77% pre-IPO target. Locking capacity and moving to custom silicon (Trainium, TPU, in-house chips) is the path to IPO-grade margins.

The analysts' read: the constraint is physical. 'A billion dollars of unrestricted capital is less valuable than assured access to GPUs, power, cooling, and interconnect' — and power is the binding limit ('compute follows power'). No lab wants to be captured inside a single cloud's pricing and allocation regime, so multi-cloud is crisis management, not branding. And long-dated commitments do double duty ahead of late-2026 IPOs: 'long-term capacity commitments are what public market investors will expect' (GeekWire, Apr 2026). OpenAI is targeting up to $1T; Anthropic a ~$2T valuation on up to $100B raised.

• The demand is real and capacity-constrained through 2027, per Moody's (Mar 2026): training plus inferencing/agentic workloads exceed supply, and hyperscalers are staging builds against contracted demand — which limits long-tail risk. The tape backs it: AWS revenue +37% in Q2 2026, Google Cloud +82%.
• Revenues are scaling into the stack. Anthropic's run rate went from $9B to $65B in seven months, with $190–200B projected for 2028. OpenAI projects >$280B of revenue in 2030 against its ~$600B compute plan. Expensive, but not the revenue-free fantasy of the fiber era.
• Take-or-pay backlog is a feature for the suppliers, not a bug: >75% of CoreWeave's >$30B 2027 revenue target is already contracted, and 12% of Oracle's $638B backlog converts within 12 months. Visibility is exactly what you want ahead of a capex cycle.
• Concentration is leverage. Dylan Patel's read (Dwarkesh, Aug 2026): OpenAI and Anthropic will take 40–50% of the world's new compute next year; by end-2027, half of incremental compute goes to two labs. If the winners keep winning, the obligations are priced against the scarcest resource in the economy.
• Multi-cloud is optionality, not profligacy: no lab wants to be captured inside one cloud's pricing and allocation regime. Custom silicon — TPUs, Trainium, Broadcom XPUs — cuts inference unit costs 20–40%, and OpenAI is already reselling scarcity downstream via 1–3 year 'Guaranteed Capacity' reservations.

• Circular financing is the load-bearing bear argument. Vendors are funding their own customers: Nvidia holds an LOI for up to $100B in OpenAI equity while OpenAI buys Nvidia chips; Nvidia backs CoreWeave, Nscale, Nebius and Lambda, guarantees up to $105B on an OpenAI Ohio data center, and launched a $500B third-party financing initiative with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR (Aug 2026). The Economist's line: Nvidia is 'the central bank of AI.' Nvidia CFO Colette Kress answers that 'some will call this circular financing. We see it differently' — but the rebuttal is a framing, not a number.
• There is a duration mismatch no one can hedge. Contracts run 5–20 years; the GPUs inside them are superseded every 12–18 months. One September survey put average GPU utilization near ~5% across ~23,000 clusters. Debt service is fixed whether or not the silicon earns it back in time.
• The revenue math doesn't cover the obligation stack. OpenAI did ~$13.1B of revenue in 2025 against a ~$60B/year average Oracle bill alone. HSBC's estimate: only 12 of the 30 GW comes online by 2030, at annual rent of ~$620B against a $280B 2030 revenue target. OpenAI doesn't expect profit until 2029, after up to $44B of losses. Ed Zitron's version: 'trillions of dollars to build an industry with only tens of billions in revenue.'
• The headline figures overstate how firm the commitments are. The Anthropic–SpaceX $45B deal is terminable on 90 days' notice. Nvidia's $100B OpenAI equity LOI is non-binding. The Anthropic–Lambda $35B was never officially announced. A skeptic's read: some of the 'obligations' are press releases dressed as contracts.
• The telecom parallel is the cleanest historical rhyme. Reuters BreakingViews (Sept 10, 2026): neoclouds are the new 'altnets' — high-yield and GPU-backed bonds today vs. PIK and zero-coupon bonds then, with Nvidia playing the vendor-financier role the telecom equipment makers played in the late 1990s. One difference: fiber eventually got used. GPUs depreciate.
• Counterparty credit is already straining. CoreWeave carried $72B of total liabilities at June 2026 (+56% year-to-date) with $640M of Q2 net interest, and 72% of its revenue comes from three customers. Oracle's FY2026: capex $55.7B (+162%), free cash flow −$23.7B, borrowings $137B — guiding ~$70B of net cash capex in FY2027 funded by roughly $40B more debt and equity.

• Totals are sums of announced deals and may double-count overlapping capacity — e.g., Crusoe-built sites serving the Oracle contract.
• Dollar values were never disclosed for the Nvidia (10 GW), AMD (6 GW) and Google Cloud deals at either lab; only Broadcom's 10 GW has third-party estimates, and they conflict wildly ($350–500B FT vs. 'multiple billions' WSJ).
• The Anthropic–CoreWeave ($12B/5yr) and Anthropic–Riot ($9.1B/20yr) figures come from secondary outlets only.
• Take-or-pay is confirmed for CoreWeave but not explicitly for the hyperscaler deals; exact reserved-vs-take-or-pay splits are not public.
• Both labs are private: their own purchase-obligation totals appear in no public filing. The first real disclosure may be Anthropic's IPO prospectus.

Assembled Sept 14, 2026 from 14 web sources (press announcements, filings, analyst notes), X and Threads discussions of the circular-financing debate, and podcast moments — including Greg Brockman's 'compute is revenue' on Odd Lots and Dylan Patel's concentration thesis on Dwarkesh. Figures are flagged announced / reported / estimate where the source warrants it.

• TechCrunch — OpenAI–Oracle $300B (Sept 10, 2025)
• GeekWire — AWS $100B expansion (Feb 27, 2026)
• TechCrunch — Anthropic–AWS >$100B (Apr 20, 2026)
• Reuters — Anthropic–Azure $30B (Nov 18, 2025)
• Bloomberg Law — Anthropic–SpaceX $45B terms
• Reuters BreakingViews — data centre upstarts vs. the broadband craze (Sept 10, 2026)
• Data Center Knowledge — Moody's: hyperscaler capex toward $700B (Mar 17, 2026)
• The Information (via Vidya) — Anthropic $517B commitments vs. $180B guidance
• Gizmodo — Nvidia defends against circular-dealm accusations (Aug 2026)