Buried in the fine print of SpaceX’s registration statement, filed with the Securities and Exchange Commission on May 20, was a number larger than the company’s entire annual revenue. Anthropic, the AI lab behind the Claude models, has agreed to pay SpaceX nearly $45 billion over the next three years for computing power, according to the filing. The payments run about $1.25 billion a month through May 2029, and either party can walk away with 90 days’ notice.
The arrangement was announced earlier in the spring, but the dollar figure was kept private until the S-1 forced disclosure. At roughly $15 billion a year, the contract nearly matches SpaceX’s standalone revenue — a single customer paying, by itself, what the whole company earned last year. For Anthropic, the deal is the largest single compute commitment it has disclosed, and it shows how far the lab will go to secure the capacity its models need.
The hardware behind the contract
The computing power comes from SpaceX’s Colossus data centers, the clusters built to train xAI’s Grok models. Under the agreement, Anthropic gets access to about 300 megawatts of capacity and roughly 220,000 Nvidia GPUs, according to people familiar with the terms. The initial months are priced at a discount while the second Colossus site is being installed, with the full monthly rate applying once capacity ramps. Anthropic’s compute chief said the lab will be scaling up on Nvidia’s flagship GB200 chips at the second site through June.
For SpaceX, the deal monetizes capacity it was already building. The company has said it expects to sign similar agreements with other firms, converting its data-center expansion into a recurring revenue stream ahead of its public listing. The S-1, filed ahead of a planned offering, frames the Anthropic contract as a template for future compute sales.
A lab buying its way out of a bind
The context for Anthropic is capacity pressure. The lab’s revenue has been growing quickly — its annualized run rate has climbed past $30 billion, according to people familiar with the figures — and demand for Claude has repeatedly outpaced its supply of compute. Earlier this year, customers experienced rate limits and latency spikes as the lab’s infrastructure strained. The SpaceX contract is a fast answer to a problem that building new data centers cannot solve quickly: the capacity exists, and the price, however large, buys immediate relief.
The deal also fits a pattern. Anthropic has secured compute from every major supplier — a long-term agreement with Amazon’s cloud unit, a commitment with Google’s infrastructure arm, and a partnership with Microsoft and Nvidia — and it is in talks to add more. Each contract diversifies the lab’s dependence on any single vendor while betting that Claude’s growth will outpace the cost of the capacity.
A deal with an exit hatch
The 90-day termination clause is the detail worth watching. It works both ways: Anthropic can leave if SpaceX raises prices or reclaims capacity, and SpaceX can give notice if xAI needs the clusters back for its own models. That makes the $45 billion figure a maximum rather than a commitment — the contract is best understood as a floating arrangement with a very large ceiling.
There is also a strategic irony the industry has noted. Anthropic has positioned itself as the safety-first alternative in AI, and it is now paying its biggest competitor’s infrastructure arm for the right to run its models. The arrangement does not change who owns the capacity; it changes who pays for it. For investors in both companies, the contract shows who owns the scarce asset in AI: everyone else is renting it from someone.
The deal also reframes how the AI industry is financing itself. Anthropic’s compute spending now spans four major suppliers — Amazon’s cloud unit, Google’s infrastructure arm, the Microsoft-Nvidia partnership and now SpaceX — and each contract runs into the tens of billions. The lab is effectively renting its entire computing foundation, committing to payments that must be serviced whether or not its revenue grows at the pace its backers hope. The 90-day termination clauses spread across those deals give Anthropic flexibility, but they also mean its entire compute stack is, in principle, renegotiable on a quarter’s notice.
For the IPO market, the contract is a data point. SpaceX’s offering, expected later this year, will now include a revenue stream from the AI industry that was not part of the company’s story a year ago. Investors valuing the rocket company will have to weigh a business that is part launch provider, part data-center landlord — and one whose largest customer is a company that could, in theory, exercise its exit clause at any time. The S-1’s disclosure that more such agreements are expected suggests SpaceX intends the AI business to be a durable part of the public-company story, not a one-off.
SpaceX’s IPO filing has exposed the shape of the AI compute economy: Anthropic will pay the rocket company nearly $45 billion over three years for the chips to run Claude. The contract nearly doubles SpaceX’s revenue run rate, gives Anthropic the capacity it urgently needs, and shows how far the AI boom has moved the center of gravity — from model quality to the hardware underneath it. With a 90-day exit hatch on both sides, the arrangement can bend, but at these numbers, it defines the market.
