SpaceX IPO filing puts Anthropic’s Colossus bill at $15 billion a year

Buried on page 13 of SpaceX’s recent prospectus is a number that reshapes the picture of the AI compute market: Anthropic is paying SpaceX $1.25 billion every month for capacity inside the Colossus 1 and Colossus 2 clusters. That figure works out to roughly $15 billion a year, a sum that exceeds the publicly reported annual revenue of every listed AI pure-play in the world.

For anyone using Claude, the number is a rare glimpse at the bill behind the chatbot. Students who debug code with it, lawyers who summarize briefs through it, hospitals piloting it for triage all sit on the same physical infrastructure, and that infrastructure now costs $1.25 billion a month to keep alive. The disclosure also clarifies how Anthropic is securing the GPU capacity needed to train and serve its next iteration of models. The financial document itself is public; the line item had not been reported at this scale until the prospectus reached the SEC’s filing window.

Until that filing dropped, the Anthropic-SpaceX relationship was described in earlier reporting as a compute supply deal anchored on Colossus 1, xAI’s Memphis-based GPU cluster in which SpaceX holds investment exposure. The new filing adds a second cluster, Colossus 2, and a price tag that runs to fifteen billion dollars over twelve months. For comparison, OpenAI’s annual revenue is in a similar order of magnitude on publicly reported estimates, and OpenAI counts more than three hundred million weekly users to recover those costs. Anthropic is committing to that compute footprint without an equivalent consumer install base.

The other unusual line in the filing is who Anthropic is paying. SpaceX, Elon Musk’s rocket company, holds investment exposure to xAI, the AI lab Musk founded explicitly to compete with Anthropic and OpenAI. Anthropic’s training infrastructure therefore passes through balance sheets connected to its most public detractor. Cloud computing has always involved competitors selling capacity to competitors, but the optics are unusual enough to attract notice from regulators in Brussels and Washington, where compute concentration is already a live question.

A monthly bill of this size points to a class of hardware that has existed at this scale for only about eighteen months. Colossus 1 was publicly described as a one-hundred-thousand-GPU cluster when xAI announced it; later reporting put the figure higher. Colossus 2 is unknown territory in any public sense. The prospectus does not state how many GPUs Anthropic actually rents, whether the $1.25 billion buys exclusive use, priority access, or a slice of total throughput. Without that detail the dollar figure is the only verifiable number. At Nvidia H200 list prices the same money would buy outright more than four hundred thousand GPUs across the year, which is one way to gauge how aggressively rental economics now favour the supplier of installed capacity over the buyer of new silicon.

Several things in the disclosure invite caution. The prospectus uses the financial-accounting language of binding capacity commitments, which is not necessarily the same as hardware sitting in racks today. A contract for capacity that comes online over the next two years would still appear as $1.25 billion a month on a forward-looking basis. Anthropic has not commented on the figure, and SpaceX’s filing does not name the counterparty in the body text of every reference. The deal could be smaller in immediate cash terms than the headline number implies, even if the total contractual exposure matches it.

Even with those caveats, the disclosure changes how every adjacent player has to plan. OpenAI’s Stargate project, Microsoft’s Azure pipeline, Meta’s Hyperion build-out, and Google’s TPU strategy all rest on the assumption that the company doing the training also controls the physical infrastructure doing the training. Anthropic has chosen a different shape: one very large external supplier, no equity tie. The arrangement makes the company nimble because it can scale by signing a contract rather than pouring concrete. It also leaves Anthropic exposed if the supplier raises prices, restricts capacity, or sells to a more lucrative bidder.

Anthropic’s services run worldwide, but the compute powering them sits predominantly in the United States by these numbers. European and Asian customers using Claude through Amazon Bedrock or Google Cloud’s Vertex AI still query a model whose training and serving rests on Tennessee soil. That has implications for jurisdictions with data-residency expectations, where the geography of where a model is actually computed has begun to matter as much as where it is offered. Enterprise contracts in regulated sectors already specify the physical location of inference; with $15 billion a year flowing to a single American facility, the next round of those contracts may include questions that Anthropic has not previously had to answer in public.

The S-1 filing is the first formal step toward a SpaceX IPO, which means the document will be amended in coming weeks as the company responds to SEC comments. The amendments may disclose the term length of the contract and the structure of the capacity commitment. Whether Anthropic is one customer among many or the cornerstone tenant that makes SpaceX’s compute business viable will determine how the next round of training-run budgets is set across the industry.