SpaceX S-1 Reveals Anthropic Compute Deal Details
SpaceX's S-1 filing, excerpted in coverage by WIRED, discloses a commercial computing arrangement giving Anthropic access to GPUs at the Colossus and Colossus II data centers. WIRED reports the contract will require Anthropic to pay $1.25 billion per month through May 2029, which WIRED frames as roughly $15 billion per year; WIRED says Anthropic confirmed those figures. The S-1 states SpaceX expects to "enter into additional similar services contracts" and that it has capacity to provide compute for its own AI models; an S-1 quote in the original filing references supporting proprietary AI applications such as Grok 5, reportedly being trained at COLOSSUS II. Reporting emphasises how access to large-scale GPUs has become a central bottleneck in advanced-model development.
What happened
SpaceX's confidential registration statement (S-1), quoted in reporting and in the filing text provided to press, discloses commercial use of the company's Colossus and Colossus II data centers to serve external AI customers. Per WIRED's coverage of the S-1, Anthropic agreed to pay $1.25 billion per month through May 2029 for access to the facilities, which WIRED frames as roughly $15 billion per year; WIRED reports that Anthropic confirmed those figures. The S-1 also indicates SpaceX expects to "enter into additional similar services contracts" and that it has capacity to provide compute for its own AI models. The S-1 text included in the original filing references supporting proprietary AI applications, naming Grok 5 and stating it is being trained at COLOSSUS II.
Technical details
Editorial analysis - technical context: The deal centres on wholesale access to large-scale GPU capacity at hyperscale data centers, in this case the Colossus family. For practitioners, running advanced generative models at training scale typically requires sustained multi-megawatt power envelopes, dense GPU racks, and high-bandwidth interconnects. Industry-wide, long-term, high-dollar contractual access to colocated GPU capacity is an established approach to secure throughput and predictable scheduling compared with spot or on-demand cloud pricing.
Context and significance
Industry context: Reporting places the contract size in stark relief against public estimates of compute demand for frontier models. The WIRED-reported payment levels illustrate how compute access can be monetized by infrastructure owners and how compute scarcity translates into multi-billion-dollar bilateral deals. For model developers and ML infrastructure teams, the revelation underscores the degree to which training pipeline economics can depend on privileged access to physical facilities rather than solely on software or cloud procurement strategies.
What to watch
For practitioners and observers: monitor future S-1 disclosures and regulatory filings for more granular operational metrics, such as utilization rates or contracted capacity; watch public statements from Anthropic and SpaceX for clarifications on billing cadence or service levels; and observe pricing and contract patterns across other data-center owners for signs of a broader market reprice for training-scale GPU access.
Attribution
All numbered contract figures and the confirmation are reported by WIRED referencing the SpaceX S-1 and conversations with Anthropic; quoted language and the Grok 5 mention appear in the S-1 text as provided in the filing materials.
Scoring Rationale
The filing reveals a rare, very large private contract for training-scale GPU access, which matters to ML teams and infrastructure planners because it highlights compute scarcity and a commercial route to secure capacity. The story is notable but not a model-paradigm shift.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
