Research · Sector & thematic · November 10, 2026
OpenAI and Anthropic: Compute Leasing, Investment Structure, and the Chip and Power Bottleneck
How the two laboratories lease and finance capacity, what Anthropic’s registration statement actually discloses about counterparties and chips, and why power and turn-on speed—not merely contracted chips—frame the next year or two.
Anthropic figures are from Amendment No. 2 to the Form S-1 of Anthropic, PBC, filed November 10, 2026. That amendment updates the aggregate commitment total of approximately $895.4 billion that appeared in Amendment No. 1, filed November 2. OpenAI has not filed a comparable registration statement. OpenAI items below are labelled “disclosed” when they come from a counterparty or company announcement, or “reported” when they come from news sources, and are cited. Unsourced model comparisons are omitted. This note is for information only and is not investment advice.
1. Summary
- Anthropic has disclosed approximately $900.9 billion of mostly multi-year compute and infrastructure commitments. Named contracts show a diversified chip mix: Google TPUs (including Broadcom equipment leases), Amazon capacity that the filing links to Trainium, and NVIDIA GPUs through Microsoft, SpaceX.AI LLC, Nscale, and Lambda. The SpaceXAI agreement is cancellable by either party on 90 days’ notice after initial periods ending in January and March 2027.
- OpenAI’s capacity is not visible in a registration statement. Disclosed and reported arrangements run through Microsoft Azure, Oracle (Stargate), CoreWeave, Amazon Web Services, and others. The dollar amounts are large, but several of them depend on partners building and powering sites.
- The view that frontier training and inference still lean on the latest NVIDIA systems (Blackwell, then Vera Rubin) is an analytical judgment in this note, not a filed finding. Diversification can hedge supply and cost; it does not, by itself, settle the performance ranking of chip families.
- In the same spirit, the tighter near-term constraint is treated here as power—grid connections and behind-the-meter supply—and the speed at which contracted capacity is actually turned on, rather than the headline size of chip contracts. That is analysis, not a measurement.
2. Anthropic’s leasing and commitment structure
As of and subsequent to September 30, 2026, aggregate committed spend was approximately $900.9 billion, substantially all for compute and infrastructure. Amendment No. 1, filed November 2, 2026, had put the same aggregate at approximately $895.4 billion. This note uses the later figure and does not keep $895.4 billion as current.
Approximately $70.7 billion is due in the 12 months after September 30, 2026, and approximately $451.6 billion is expected to be incurred by December 31, 2030. About $186.2 billion of the aggregate is subject to exercisable termination rights or contingent conditions before commencement. Unfixed arrangements estimated at $150–175 billion, including additional Google TPU capacity, are excluded from the $900.9 billion total. See also the earlier note on commitments and cash flow.
| Counterparty | Amount (as disclosed) | Chip / capacity (as disclosed) | Character | Termination |
|---|---|---|---|---|
| Broadcom (and financing assignees) | Aggregate rent ~$161.2B, excluding interest | TPU equipment, five-year master finance leases | Rent typically starts when equipment is ready for service | Non-cancelable during the lease term except by the lessor in a default |
| At least ~$121.3B | TPU capacity via a CSP hosting agreement. Additional unfixed TPU, estimated $125–150B, is not in the commitment table | Minimum spend regardless of usage, with a true-up if actual spend falls short | Non-cancelable except for Google’s uncured material breach | |
| SpaceX.AI LLC | Up to $120.5B | NVIDIA GPUs. About $3.1B a month once all capacity is delivered (expected December 2026); terms into 2029 | A cap, not a fixed minimum in the same sense as the hyperscaler true-up contracts | After initial non-cancelable periods ending January 2027 and March 2027, either party may cancel on 90 days’ written notice |
| Amazon | At least $110.0B | The $110B agreement does not break out chip type. Elsewhere the filing describes the Amazon collaboration as providing access to Trainium chips and compute infrastructure | Minimum spend regardless of usage, with a true-up | Non-cancelable except for Amazon’s uncured material breach |
| Microsoft | At least $47.2B and up to $49.2B | NVIDIA GPUs (the filing groups Microsoft with SpaceXAI, Nscale, Lambda, and other providers of NVIDIA capacity) | Minimum spend regardless of usage, November 2026–June 2033 | Non-cancelable except for Microsoft’s uncured material breach or, on a per-tranche basis, certain undelivered tranches. (Amendment No. 2; this is not the $44.9B floor that appeared in earlier notes based on the November 2 amendment.) |
| Nscale | Up to ~$44.6B | NVIDIA GPUs, terms into 2033 | Fees accrue only as capacity is delivered and accepted | Otherwise non-cancelable except in limited cases, including failure to meet delivery or availability requirements or an uncured material breach; Nscale is also to use best efforts to obtain qualifying financing |
| Lambda | ~$41.3B | NVIDIA GPUs, terms into 2034 | Fees accrue only as capacity is delivered and accepted | An August 2026 agreement of ~$34.9B is currently terminable because a facility lease was not secured (neither party has terminated). Other Lambda commitments are non-cancelable except for material breach or, in some cases, missed delivery milestones |
| AMD | No fixed dollar amount in the table | AMD Instinct GPUs, including the MI450 series, via a hardware supply agreement | If the full initial 2027 tranche is taken, spend is expected to exceed $20B; not in the aggregate table | Non-cancelable by Anthropic except for AMD’s uncured material breach |
Source: Anthropic, PBC, Amendment No. 2 to Form S-1, filed November 10, 2026. Amounts are as stated: some are minimums, some are caps, and some remain contingent or terminable.
Chip mix — a derived estimate, not a filed table
The registration statement does not present a percentage mix of chips. The figures below are a simple allocation of named contract dollars, using the Microsoft floor of $47.2 billion:
- Google TPU, counting the Broadcom leases and the Google Cloud minimum: $161.2B + $121.3B = $282.5B, or about 44% of the $646.1 billion named subtotal.
- NVIDIA GPUs, counting SpaceXAI, Microsoft, Nscale, and Lambda: $120.5B + $47.2B + $44.6B + $41.3B = $253.6B, or about 39%.
- Amazon’s $110.0B minimum: about 17%. This is not a disclosed Trainium share. The $110B contract does not break out chip type.
These percentages omit the unfixed Google TPU estimate of $125–150 billion, the unfixed AMD expectation, and other items that sit inside the $900.9 billion aggregate. SpaceXAI, Nscale, and Lambda amounts are caps or delivery-linked. If any of the Amazon minimum is NVIDIA capacity, the NVIDIA share is understated. The 44 / 39 / 17 split should be read as an illustration of named dollars, not as a capacity or performance mix.
Structure, as filed. The Google and Amazon agreements require payment of the minimum regardless of usage, with a true-up if spend falls short. Microsoft is also a minimum regardless of usage. SpaceXAI and the neo-cloud contracts are more flexible, through cancellation rights or delivery-linked fees. The filing also describes financing and convertible-note arrangements with certain infrastructure partners, including Amazon, Google, and Broadcom—so a supplier can also be a lender or investor. Revenue generated through cloud-service-provider marketplaces was 44% of total revenue in the first nine months of 2026 (47% in 2025).
3. OpenAI’s leasing and commitment structure
OpenAI has not filed a public S-1, so there is no audited total comparable to Anthropic’s $900.9 billion. The rows below are company disclosures or news reports, not a complete inventory.
| Partner | Amount or scale | Basis | Notes |
|---|---|---|---|
| Microsoft Azure | Incremental $250B of Azure services | Disclosed by Microsoft, October 28, 2025 (company blog and SEC disclosure) | Microsoft also said it would no longer have a right of first refusal to be OpenAI’s compute provider. This is a counterparty disclosure, not an OpenAI registration statement. |
| Oracle (Stargate) | Up to 4.5 GW of additional U.S. capacity; a later computing agreement of about $300B has been reported | 4.5 GW: OpenAI announcement, July 2025. ~$300B: reported by the Wall Street Journal and Reuters, September 2025 | The 4.5 GW figure is not the same thing as the Abilene campus alone. A planned Abilene expansion was later reported as dropped (Bloomberg / Reuters, March 2026); Oracle said the 4.5 GW agreement remained on track. |
| CoreWeave | Up to $22.4B | Reuters and Bloomberg, September 25, 2025, consistent with CoreWeave announcements and an 8-K for a $6.5B increment | The total combines earlier contracts of up to $11.9B and $4B with the September 2025 order. At least the earlier reserved-capacity terms have been described as take-or-pay in character. Reported / disclosed by the supplier, not by an OpenAI filing. |
| Amazon Web Services | A $38B, seven-year agreement, later expanded by $100B over eight years, including about 2 GW of Trainium capacity | $38B: Reuters, November 3, 2025. Expansion and Trainium: OpenAI and Amazon announcement, February 27, 2026 | Company announcements. This replaces a vaguer “tens of billions plus Trainium” formulation. |
| NVIDIA | Guarantee of up to $105B | Reuters and Bloomberg, August 17, 2026, describing a NVIDIA disclosure | This is a contingent residual-value guarantee tied to leases on an Ohio data-center campus, not a $105 billion commitment to buy chips. The cap applies if specified default or insolvency events occur. |
OpenAI items are disclosed by a counterparty or company, or reported in the press, and should not be added together as if they were a single audited commitment schedule.
A fair structural contrast, made cautiously: Anthropic’s named NVIDIA share, on the derived dollar basis above, is lower than the picture often drawn for OpenAI, which remains NVIDIA-heavy in public descriptions while also adding AMD, Trainium, and other custom silicon. OpenAI’s capacity is, in the arrangements above, largely leased from partners that must build and power the sites—so timing risk sits with those partners’ construction and interconnection schedules. That is a reading of the disclosed structures, not a forecast of delays.
4. Investment structure — a comparison
| Anthropic | OpenAI | |
|---|---|---|
| Overlapping investors and suppliers | The S-1/A describes compute supply and, separately, financing or investment ties involving Amazon, Google, and Broadcom | Microsoft holds an equity stake and has disclosed a large Azure commitment. NVIDIA has been reported as an investor and as a guarantor of certain leases. SoftBank appears in Stargate and related announcements |
| Minimum-spend terms | Google, Amazon, and Microsoft: payment regardless of usage | Varies by partner; CoreWeave reserved capacity has been described as take-or-pay in character |
| Chip diversification | Filed access to Trainium, TPUs, NVIDIA GPUs (including Blackwell and Vera Rubin), and AMD GPUs. The derived dollar mix above is TPU-heavy among named contracts | Public descriptions remain NVIDIA-centered, with disclosed Trainium capacity at AWS and other diversification. No filed mix is available |
| Flexibility | SpaceXAI cancellation rights after early 2027; Nscale and Lambda fees and termination are partly delivery-linked | Several large agreements look less easy to resize, though terms are not all public |
| Principal risks, as framed here | Honoring take-or-pay minima, and the cost of short-term compute (the S-1 notes that short-term commitments have recently cost more than multi-year supply) | Dependence on partners’ construction and power schedules, where those schedules are what make contracted capacity real |
5. NVIDIA systems, SpaceX, and the power bottleneck
Anthropic’s filing states that it trains and serves models across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, and that its NVIDIA arrangements include Grace Blackwell and Vera Rubin systems. Alternatives exist, and Anthropic is using them. The further claim that the highest-performance frontier work still depends most on the latest NVIDIA generation, and on the CUDA software stack and cluster scale that go with it, is the analysis in this note, not a finding in the S-1. Diversification can reduce vendor and price risk; it is not, by itself, evidence about peak model quality.
SpaceX’s role is clearer on one point and more inferential on others. The S-1/A names SpaceX.AI LLC, a wholly owned subsidiary of Space Exploration Technologies Corp., as a supplier of NVIDIA GPU capacity up to $120.5 billion. On this site’s SpaceX third-quarter 2026 results note, Super Intelligence was the largest segment by revenue, nameplate compute was 1.7 gigawatts, and management pointed to 2.5–3 gigawatts by year-end—outlook, not a historical result. Claims that SpaceX is ahead of Meta or other hyperscalers in construction speed, permitting, or behind-the-meter power, or that it has a preferred allocation of Vera Rubin systems, are analytical judgments, not disclosed facts. They are not used here as if they had been measured.
The SpaceXAI contract does let Anthropic buy capacity from a supplier that is also building compute for sale. After the initial periods ending in January and March 2027, either party can cancel on 90 days’ notice. That flexibility is in the filing. Whether it produces faster turn-on than Anthropic’s other contracts is a separate, and unproven, question.
6. The next one to two years — a scenario, not a forecast
- Chip access. The S-1 confirms that Anthropic’s NVIDIA path includes Vera Rubin systems. Any suggestion that SpaceX will have a relative advantage in Rubin allocation, and that Anthropic will inherit it through SpaceXAI, is an inference, not something the filings show.
- Turn-on. This note treats power and construction speed as the likely binding constraint. A ranking sometimes offered in market commentary—SpaceX first, then the Microsoft/Oracle–OpenAI axis, then Meta—is a qualitative observation, not an official or measurable league table, and it can vary by site and year.
- Positions. Anthropic’s disclosed structure combines diversification with a large, cancellable NVIDIA contract at SpaceXAI. OpenAI’s disclosed and reported scale is larger in places, and more of it sits with partners’ build schedules. Neither statement is a prediction of who will have more usable capacity.
7. Closing
Both laboratories are pre-empting capacity with multi-year minima and leases rather than relying only on pay-as-you-go. The documented difference is that Anthropic’s S-1/A shows a broader named chip set and a specifically cancellable, NVIDIA-based contract with a SpaceX subsidiary. Whether that combination actually turns capacity on faster is the open question, not a result.
If the binding constraint over the next year or two is power and turn-on speed rather than the face value of chip contracts, then suppliers that can energize sites—and laboratories that have genuine options among those suppliers—matter more than a raw commitment total. That is the framing of this note. It is not a ranking of investment merit.
This note is for information only and is not investment advice. Anthropic has filed a registration statement, but the offering is not completed and the company is not yet listed. OpenAI figures that are not in a public filing may be incomplete or may change. Figures should be checked against the original disclosures.
Sources
- Anthropic, PBC, Amendment No. 2 to Form S-1, filed November 10, 2026 (File No. 333-299633; Accession 0001628280-26-072484). Primary source for Anthropic commitments, counterparties, termination terms, chip platforms, marketplace revenue, and cash-flow context.
- Anthropic, PBC, Amendment No. 1 to Form S-1, filed November 2, 2026, which stated an aggregate of approximately $895.4 billion, subsequently updated.
- Microsoft, “The next chapter of the Microsoft–OpenAI partnership,” October 28, 2025, and Microsoft’s related SEC disclosure: OpenAI contracted to purchase an incremental $250 billion of Azure services.
- OpenAI, “Stargate Advances with 4.5 GW Partnership with Oracle,” July 2025. Reuters, “OpenAI, Oracle sign $300 billion computing deal, WSJ reports,” September 10, 2025. Reuters, “Oracle and OpenAI drop Texas data center expansion plan,” March 6, 2026.
- Reuters, “CoreWeave expands OpenAI pact with new $6.5 billion contract,” September 25, 2025; Bloomberg, same date; CoreWeave Form 8-K on the September increment.
- Reuters, “OpenAI turns to Amazon in $38 billion cloud services deal,” November 3, 2025. OpenAI, “OpenAI and Amazon announce strategic partnership,” February 27, 2026.
- Reuters, “Nvidia to provide up to $105 billion guarantee for OpenAI’s Ohio data center,” August 17, 2026; Bloomberg, same date.
- This site’s SpaceX Q3 2026 results note and Anthropic financials data file, compiled from the earnings release and the S-1/A.