#AnthropicDiscloses$84.5BComputeDealWithSpaceX
Anthropic's confidential IPO prospectus, seen by Reuters, shows the company has committed as much as 84.5 billion dollars to SpaceX for Nvidia-based computing capacity through 2029. That is nearly double the roughly 45 billion dollars SpaceX disclosed in its own IPO filing in May, an increase of almost 88 percent. The structure is simple and brutal: about 1.25 billion dollars a month, roughly 15 billion dollars a year, for capacity at the Colossus 1 data center in Memphis, Tennessee, where more than 300 megawatts of power feeds more than 220,000 Nvidia GPUs. Both sides can walk away with 90 days of written notice, and the 84.5 billion figure is a ceiling rather than a guaranteed cheque.

Put that 15 billion dollars a year next to SpaceX's own numbers. SpaceX reported total revenue of 18.7 billion dollars for 2025. A single lease, to a single customer, at full ramp, equals roughly 80 percent of everything the entire company earned last year. That single fact explains why SpaceX, a rocket and satellite business seven weeks away from the largest listing in Wall Street history, suddenly reads like an infrastructure company with a launch pad attached. The AI infrastructure platform it built around xAI's Colossus clusters, now part of SpaceX itself, went from a drag on earnings to a profit engine: xAI lost 2.4 billion dollars in the three months to March 2026, up from 936 million a year earlier, after spending 12.7 billion dollars on AI infrastructure in 2025 and another 7.7 billion in the first quarter of 2026 alone, an annualised pace near 30 billion dollars. Empty GPUs are a liability. Leased GPUs at 1.25 billion dollars a month are a business.

Anthropic, meanwhile, is not guessing about demand. Its revenue went from 4.73 billion dollars in the first quarter of 2026 to 11.5 billion dollars in the second, growth of about 142 percent in a single quarter, meaning one quarter was larger than all of 2025 combined. Its annualised run-rate reached roughly 47 billion dollars by mid-May, up more than five times, or over 420 percent, from approximately 9 billion dollars at the end of 2025. More than 1,000 business customers now spend over one million dollars a year each, double the number from under two months earlier. When your product is capacity-constrained, compute is not an expense line, it is the product.

That is why the SpaceX deal is only one tile in a much larger mosaic. The same prospectus shows Anthropic expects to spend at least 518 billion dollars over the next decade with six infrastructure partners, and about 80 percent of that total sits in non-cancelable arrangements that must be paid regardless of usage. The breakdown is 111.1 billion dollars with Alphabet's Google, 110 billion with Amazon, and 31.4 billion with Microsoft. Those three partners alone account for roughly 48.7 percent of the total, and the SpaceX commitment represents about 16.3 percent of it. Around this core sit other commitments: up to 5 gigawatts with Amazon, including nearly 1 gigawatt of new capacity by the end of 2026, 5 gigawatts of next-generation Google and Broadcom TPUs starting in 2027, 30 billion dollars of Microsoft Azure capacity with Nvidia, and a 50 billion dollar American AI infrastructure program with Fluidstack.

Here is what each side actually buys. For Anthropic, the prize is speed and optionality. Building a gigawatt-scale campus takes years, while Colossus capacity arrives in weeks, the same month in some cases, and directly lifts the limits available to Claude Pro and Claude Max subscribers. The 90 day termination clause matters too: it lets Anthropic treat SpaceX as flexible surge capacity that can be trimmed once cheaper or more permanent hyperscaler capacity comes online. For SpaceX, the prize is validation plus cash flow. The company gets a marquee customer that de-risks the capital it poured into Memphis, a second paying tenant in Google, which reports say will pay 920 million dollars a month from October 2026 through June 2029 for capacity including at least 110,000 Nvidia chips, and a public market story in which compute revenue grows alongside Starlink and launch.

Now the part that changes AI development itself. First, capital risk is shifting from the chip vendor to the model lab. When 80 percent of a 518 billion dollar commitment is non-cancelable, the AI lab is underwriting the infrastructure bubble it depends on, and earnings quality becomes the new battleground rather than benchmark scores. Second, the bottleneck has moved. It is no longer only about who can buy the best GPUs, it is about who can secure land, power, grid interconnects and turbines, which is exactly why an aerospace company with a Memphis campus became a serious AI landlord overnight. Third, the industry is going vertically integrated: Anthropic is moving from a cloud-only posture toward dedicated data centers, directly leased chips and its own buildout, which improves unit economics but also locks in decade-long obligations against a business that did not exist three years ago.

Fourth, and this is the genuinely new frontier, compute is leaving the planet. SpaceX filed with the FCC in January 2026 for an orbital data center constellation of up to one million satellites, arguing that once Starship flies reusably, launching one million tonnes of satellites a year at 100 kilowatts of compute per tonne would add 100 gigawatts of AI capacity annually, roughly a fifth of current United States electricity consumption, with minimal ongoing maintenance. It unveiled the AI1 orbital data center satellite on 9 June 2026 and has targeted 1 gigawatt of space compute by late 2027, alongside a fabrication plan in Texas worth an estimated 55 billion dollars. Whether or not orbital compute economics hold, and serious engineering skeptics say cooling and radiation in orbit remain unsolved, the deal flow now assumes space as a real answer to terrestrial power scarcity rather than a science fiction slide.

The market is already pricing both sides of this. SpaceX, trading as SPCX after pricing its June 2026 IPO at 135 dollars and raising roughly 75 billion dollars in the largest listing ever, closed its first day at 160.95 dollars, up 19.2 percent, at a market capitalisation above 2.1 trillion dollars. It has since swung between 105 and 226 dollars: up about 10 percent from its IPO price near 149 dollars in late September, yet roughly 34 percent below its high, with a 5 percent drop in a single session on 9 September when a 47.2 billion dollar insider unlock hit the tape, and a roughly 3 percent gain on the day the 84.5 billion dollar Anthropic disclosure landed. The chip complex around it has been just as volatile: Nvidia traded at 228.78 dollars, up 0.69 percent, after a 7.8 percent rise from about 212 dollars in mid-September, alongside Broadcom up 0.39 percent and TSMC down 0.32 percent. One mid-September week of AI bubble anxiety wiped 3.9 percent off Nvidia, about 200 billion dollars of market value, 7.3 percent off Broadcom, about 140 billion dollars, more than 10 percent off Micron and AMD, 6 percent off Amazon and 3 percent off Microsoft, even as the underlying deals kept getting bigger.

The honest counterweight is that scale cuts both ways. Anthropic's 2025 result included a net loss of about 42 billion dollars, of which roughly 34 billion dollars was a non-cash accounting charge tied to the rising estimated value of financing convertible into shares, and it held 20.28 billion dollars in cash, equivalents and short-term investments at the end of that year. Nearly a quarter of its revenue came from just two customers, many of its largest clients are not locked into long-term contracts, and its own prospectus warns that the technology it is scaling could be dangerous, a strange thing to read in a document written to raise up to 100 billion dollars at a valuation its backers hope exceeds 2 trillion dollars. Dependence on Amazon, Google and Microsoft is itself a listed risk, given that all three run competing models while also supplying its infrastructure and distribution.

None of that changes the direction of travel. The Musk versus Amodei rivalry was loud enough that Musk publicly called Anthropic "the most hypocritical company" in March 2026, and weeks later his company handed over the entire output of a data center to them. That is what a compute shortage does to ideology. The lesson for anyone watching this sector is that the AI story is no longer written in model releases. It is written in megawatts, monthly invoices of 1.25 billion dollars, 90 day exit clauses, and increasingly, in whether a satellite in orbit can be a cheaper place to think than a warehouse in Tennessee.
$SPCX