Independent ownership research on U.S.-listed companiesMethodology  ·  Disclosures
Home / Research / Sector & thematic

Research · Sector & thematic · November 10, 2026

Anthropic’s SpaceXAI Contract and a Comparison of Frontier Compute Strategies

The filed terms of Anthropic’s agreement with SpaceX.AI LLC, where that contract sits in the commitment stack, and a qualitative comparison with OpenAI, Google, Meta, xAI/SpaceX, and Amazon.

SpaceXAI terms and Anthropic commitment totals are from Amendment No. 2 to Anthropic’s Form S-1, filed November 10, 2026, which updates the $895.4 billion aggregate in Amendment No. 1 (November 2). OpenAI, Meta, and Google items that are not in that filing are cited to company announcements or news reports, or are omitted. Rankings of turn-on speed and “NVIDIA dependence” are the author’s qualitative assessment, not measured official statistics. This note is for information only and is not investment advice.

SpaceXAI cap
up to $120.5B
NVIDIA GPU capacity, into 2029
Run-rate if full
~$3.1B / month
once all capacity is delivered, expected Dec 2026
Cancellation
90-day notice
after initial periods ending Jan and Mar 2027
Share of aggregate
~13%
of $900.9B; a ratio of the cap, not fixed spend

1. The SpaceXAI contract, as filed

Between May and October 2026, Anthropic entered into and amended several agreements with SpaceX.AI LLC (formerly X.AI LLC), a wholly owned subsidiary of Space Exploration Technologies Corp., for compute powered by NVIDIA GPUs. Under those arrangements the company agreed to spend an aggregate of up to $120.5 billion over terms extending into 2029, consisting of approximately $3.1 billion in the aggregate per month once all capacity has been delivered, which it expects in December 2026, with different fees during capacity-ramping periods.

Following initial non-cancelable periods ending in January 2027 and March 2027 under the respective agreements, each agreement may be cancelled by either party upon 90 days’ advance written notice. That is the central flexibility in the contract, and it is in the S-1/A—not an inference.

By contrast, the Google, Amazon, and Microsoft agreements require minimum spend regardless of usage (with true-ups at Google and Amazon) and are non-cancelable except for the counterparty’s uncured material breach, or, for Microsoft, certain undelivered tranches. Calling SpaceXAI the most flexible of Anthropic’s large named contracts is a fair reading of those termination clauses. It is not the same structure as Nscale or Lambda: those fees accrue only as capacity is delivered and accepted, and a $34.9 billion Lambda agreement is currently terminable because a facility lease was not secured. SpaceXAI’s flexibility is the post-initial-period cancellation right, not an identical delivery-contingency clause.

Where it sits in the portfolio

As filed or derived
CapUp to $120.5 billion
Share of Anthropic’s aggregate commitmentsAbout 13% of the $900.9 billion aggregate in Amendment No. 2 ($120.5 ÷ $900.9). Amendment No. 1 had put the aggregate at about $895.4 billion; this note uses the later total. Because $120.5 billion is a cap, and because part of the $900.9 billion is contingent or terminable, this is a ratio of disclosed maxima, not a share of irreducible spend.
Chip typeNVIDIA GPUs, as stated in the filing
CharacterAn “up to” cap with cancellation rights after early 2027, rather than a use-or-pay minimum of the Google/Amazon/Microsoft type
Strategic reading (analysis)A way to buy NVIDIA capacity from a supplier that is also building compute for sale, while retaining an exit after the initial periods. Whether that produces faster turn-on than Anthropic’s other contracts is not something the filing demonstrates.

Source for contract terms and the $900.9 billion aggregate: Anthropic, PBC, Amendment No. 2 to Form S-1, filed November 10, 2026. See also the notes on commitments and cash flow and leasing structure.

2. A qualitative comparison of compute strategies

The table below is a compact comparison, not a scoring model. “NVIDIA dependence” for Anthropic uses the derived share of named contract dollars (about 39%) explained in the leasing-structure note: SpaceXAI, Microsoft, Nscale, and Lambda, over a $646.1 billion named subtotal, excluding unfixed Google TPU and AMD amounts and not treating the Amazon $110 billion minimum as a disclosed Trainium percentage. Other companies’ “high” or “low” labels are qualitative. They are not filed percentages.

CompanyChip strategy (as documented or described)NVIDIA dependence (qualitative, except Anthropic’s derived share)Notes and sources
AnthropicFiled access to Google TPUs (including Broadcom leases), Amazon/Trainium infrastructure, NVIDIA GPUs, and AMD GPUsIntermediate; about 39% of named contract dollars, derived, not a filed mixSpaceXAI is the large, cancellable NVIDIA contract. S-1/A Amendment No. 2.
OpenAIPublic descriptions remain NVIDIA-centered, with disclosed Azure, Oracle, CoreWeave, and AWS/Trainium arrangements and other diversificationHigh, as a qualitative label; no filed mixSee the leasing-structure note for disclosed and reported amounts. Actual turn-on depends on partners’ sites.
GoogleCustom TPUs designed and used in-house; Google Cloud also offers NVIDIA capacity to customersLow for Google’s own frontier training, on the public record; not zero in the cloud businessCNBC, November 2025, reported that Gemini 3 was trained entirely on TPUs. Business Insider, April 2026, noted that Google Cloud sells access to NVIDIA systems, including planned Vera Rubin instances.
MetaLarge self-owned clusters built around NVIDIA systems, with a parallel AMD partnershipHigh, as a qualitative labelNVIDIA announced a multiyear partnership on February 17, 2026, covering millions of Blackwell and Rubin GPUs (Reuters). Meta’s October 19, 2026 engineering post describes the Prometheus supercluster, scaling toward about a million GPUs, with Vera Rubin NVL72 arriving. Speed versus SpaceX is not measured here.
xAI / SpaceXNVIDIA-focused clusters, self-built, with capacity also leased to AnthropicVery high, as a qualitative label; no filed mixThe S-1/A names SpaceXAI as an NVIDIA supplier to Anthropic. SpaceX’s Q3 2026 results on this site show Super Intelligence as the largest segment. Construction-speed claims are analysis.
AmazonTrainium plus NVIDIA in AWSIntermediate, as a qualitative labelAnthropic’s S-1 describes Trainium access through the Amazon collaboration and a $110.0 billion minimum. OpenAI has announced AWS/Trainium capacity. AWS also sells NVIDIA instances. No filed percentage split.

Turn-on speed — a qualitative ranking, not a league table

The following order is the author’s assessment of relative ability to energize and use new capacity in the near term. It is not an official ranking, a measured statistic, or a consensus of “the prevailing view.” It can vary by site and year.

CompanyWhy it is placed here (assessment)
1SpaceX / xAISelf-built clusters and an emphasis, in public commentary, on rapid construction and behind-the-meter power. This is an inference from that posture and from Anthropic’s choice of SpaceXAI as a large NVIDIA supplier—not a comparative audit of interconnection queues.
2AnthropicIndirect access to that supplier through the SpaceXAI contract, plus a diversified set of other contracts. Diversification reduces some risks; it does not, by itself, prove faster turn-on.
3OpenAI (Microsoft / Oracle axis)Very large disclosed and reported commitments, much of them sitting with partners that must still build and power sites.
4MetaVery large self-owned infrastructure (Prometheus and related). A claim that Meta is slower than SpaceX on permitting, power, or construction is an assessment in this note, not a documented finding, and is not attributed to “many observers.”
5GoogleTPU vertical integration gives supply and cost control for Google’s own models. The placement reflects a narrower role for the latest NVIDIA systems in Google’s own training, not a judgment that Google is slow to energize TPUs.

3. Meta, in more detail, and with sources

On chips, the public record supports a NVIDIA-centered, self-owned build-out. NVIDIA’s February 17, 2026 announcement, reported by Reuters, described a multiyear partnership for millions of Blackwell and Rubin GPUs. Meta’s own October 19, 2026 engineering post on Prometheus describes a gigawatt-scale supercluster of hundreds of thousands of GPUs, intended to scale toward about one million, using GB200 and GB300 systems with Vera Rubin NVL72 arriving. That is a large, self-directed NVIDIA cluster. It does not, by itself, establish that Meta is slower than SpaceX at turning the latest generation on.

Power and permitting comparisons with SpaceX’s behind-the-meter approach, and any claim that Anthropic will have a one- to two-year advantage in “turning on the newest chips” via SpaceXAI, are analytical judgments. They are not filed facts and are not presented here as a dominant external consensus.

Meta’s longer-term advantage, if the assessment holds, is control of its own sites and a very large installed base. Anthropic’s nearer-term option is a cancellable contract with a supplier that is building NVIDIA capacity to sell. Those are different structures, not a completed race.

4. What follows, cautiously

This note is for information only and is not investment advice. Anthropic’s offering is not completed and the company is not yet listed. Qualitative rankings can be wrong and should not be read as precision.

Sources

  1. Anthropic, PBC, Amendment No. 2 to Form S-1, filed November 10, 2026, for SpaceXAI terms, the $900.9 billion aggregate, other counterparty terms, and chip platforms.
  2. Anthropic, PBC, Amendment No. 1 to Form S-1, filed November 2, 2026 (aggregate previously stated as approximately $895.4 billion).
  3. NVIDIA, “Meta Builds AI Infrastructure With NVIDIA,” February 17, 2026; Reuters, “Nvidia to sell Meta millions of chips in multiyear deal,” same date.
  4. Meta Engineering, “Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster,” October 19, 2026.
  5. CNBC, “Gemini 3 — and the custom chips that power it,” November 25, 2025. Business Insider, “Google’s New Chips Are a Shot at Nvidia in AI Inference,” April 2026.
  6. This site’s SpaceX Q3 2026 results note.