Anthropic is paying Elon Musk’s SpaceX $15 billion a year for compute

Anthropic has committed $1.25 billion every month to SpaceX in exchange for GPU capacity to train and serve its frontier models, a sum disclosed when SpaceX filed for its initial public offering and that totals roughly $15 billion a year. The arrangement nearly doubles SpaceX’s current annual revenue from a single customer, and it folds the AI lab behind Claude into the orbit of Elon Musk’s holdings at a moment when his own model, Grok, is one of its loudest competitors.

If neither side pulls out, SpaceX will collect more than $40 billion over the life of the deal, a figure that lands somewhere between a hyperscaler cloud bill and a national infrastructure project. Either party can walk with 90 days’ notice, which is unusually short for a multi-year compute agreement and reads as a hedge against the volatile economics of frontier AI.

The capacity itself comes from Colossus, the Memphis data-center campus where the GPUs physically sit. Anthropic begins with 300 megawatts at Colossus 1 and is already scaling into Colossus 2 with Nvidia GB200 silicon, the chip architecture that the largest labs are racing to deploy at scale. Three hundred megawatts is the kind of draw that registers on a regional power grid rather than a server-room dashboard; it is the floor at which a tenant becomes a customer of the local utility as much as of the landlord.

What makes the arrangement editorially strange, rather than merely large, is who is paying whom. Anthropic was seeded on a thesis of safety-first AI that explicitly positioned itself against Musk’s brand of accelerationism. The company spent the last two years building Claude as the careful alternative to xAI’s Grok, and yet the compute it needs to keep that argument alive now flows through the financial accounts of a Musk-led firm. Anthropic’s money does not buy Grok, but it does help finance the infrastructure that Musk’s broader portfolio runs on, including the same Colossus site that xAI helped popularize as the muscular shorthand for frontier-scale compute.

Behind the irony is a constraint that none of the frontier labs have solved: there is not enough compute in the world, at the right voltage and the right latency, to satisfy what the largest models want. Anthropic’s revenue line has begun to climb sharply enough that internal projections point to a first quarterly profit, the next quarter on track to be solidly black, and a possible public listing at a valuation discussed in trillion-dollar terms. None of that lands if the company cannot serve its own paying customers, and the SpaceX deal is the bluntest possible answer to that bottleneck.

The price is also the answer to a question SpaceX has been quietly working through: what is the company’s second business, once the rocket cadence saturates and Starlink reaches its capex peak. AI infrastructure-as-a-service, sold at the scale of a single $15-billion-a-year contract, is now visible in its IPO documents as that second business in motion. Musk has framed the Anthropic agreement publicly as proof that SpaceX can deliver compute-as-a-service at a scale the rest of the market cannot match yet, a positioning that doubles as marketing for the listing itself.

There are reasons to be cautious about the framing. Three hundred megawatts is significant, but it is still a fraction of the announced capacity the top three U.S. cloud providers have lined up for AI through the same window. The 90-day exit clause cuts both ways: it gives Anthropic an escape hatch if a cheaper provider emerges, and it gives SpaceX a path out if Anthropic’s growth stalls before the contract runs its course. The early-stage payments are reduced as the ramp begins, which means the $1.25-billion-a-month figure understates the real run-rate even more than a single line already suggests. And the model of paying a competing ecosystem for the picks and shovels of your own business has historical precedent, but the precedent is uncomfortable: the supplier eventually decides to compete with the customer, or the customer eventually concludes the supplier is too risky to depend on.

Geography matters here too. Colossus sits in Memphis, on a grid that the Tennessee Valley Authority is already stretching to keep up with new AI demand. The expansion to Colossus 2 adds load that has to come from somewhere, and the AI industry’s appetite for power is now a political issue in states across the southeastern U.S. Anthropic is, in effect, plugging Claude into a regional energy debate it did not start and cannot meaningfully shape.

For an international reader the more useful framing is probably this: the price of running a frontier model is now visible on the open market, and it is roughly $1.25 billion a month for capacity that is only a slice of what the largest labs consume. That number changes how you read every announcement about model releases, pricing tiers, and API rate limits. It also changes how you read every claim of an independent, safety-first AI lab; independence in this industry now means choosing which giant’s infrastructure you rent.

SpaceX’s S-1 filing is the document that put the number on the public record. The company plans to list on Nasdaq under the ticker SPCX, and the Anthropic disclosure is one of the load-bearing data points in that filing. The contract itself runs through May 2029, with reduced fees during the ramp in May and June and full pricing kicking in after the new facility is fully online. Anthropic, for its part, has signalled an IPO targeted for later this year. The next checkpoint is the Colossus 2 ramp, which Anthropic has said will accelerate through June, with the company expecting to draw materially more GB200 capacity as the site fills out.