SpaceX IPO Reveals Anthropic Pays xAI $1.25B a Month for Compute

SpaceX filed its IPO paperwork with the SEC on May 20, and buried in the S‑1 filing was a number that made the AI world do a double‑take: Anthropic has agreed to pay Elon Musk's xAI $1.25 billion per month for access to its data centers. That's $15 billion a year, running through May 2029, according to The Verge. The deal could total over $40 billion if it runs its full three‑year term, per.²

The arrangement was first disclosed in early May when CNBC reported that Anthropic and SpaceX had struck a compute deal — but the dollar figures were private until the SEC filing forced them into the open. Now we know exactly how much Anthropic is paying a company whose AI chatbot, Grok, directly competes with Anthropic's own Claude.

The deal is a product of raw compute desperation. Anthropic CEO Dario Amodei has publicly acknowledged what he called "difficulties with compute" in early May, CNBC reported. Demand for Claude and the Claude Code developer tool has surged, and the company needs capacity now — not in two years when new data centers might come online.

SpaceX, meanwhile, has a compute oversupply problem. After merging with xAI earlier this year, the combined company spent $12.7 billion on AI capital expenditures in 2025 and another $7.7 billion in Q1 2026 alone, according to DataCenterDynamics. But xAI's own AI division lost $2.4 billion in Q1 2026, and Grok downloads have been declining. Renting unused capacity to Anthropic turns a cost center into revenue.

As the IPO filing itself puts it: "This structure allows us to monetize unused compute capacity in our infrastructure, while still permitting reallocation of the capacity for our own internal initiatives if needed in the future," the filing states, DataCenterDynamics reported.

The deal covers 300 megawatts of compute capacity across the Colossus I and Colossus II data centers in Memphis, Tennessee. But the actual usable power is more complicated than Musk's public claims. While Musk has touted a 1‑gigawatt cluster, the IPO filing's fine print defines that as "nameplate compute draw" — the theoretical maximum if every GPU ran at full power simultaneously, which never happens — DataCenterDynamics noted from the IPO filing.

In reality, DataCenterDynamics reports that Colossus I provides about 130MW and Colossus II about 210MW — roughly 340MW total. Satellite imagery from January 2026 suggested cooling equipment capable of handling only about 350MW. Expansion plans call for 220,000 additional GB300 processors and over 400 more megawatts at Colossus II, but no timeline has been disclosed.

The data centers have drawn significant legal and environmental scrutiny. xAI faces multiple lawsuits over its use of natural gas turbines without proper permits, including alleged Clean Air Act violations. In April 2026, the NAACP filed suit accusing xAI of illegally running 27 methane gas turbines at Colossus II, DataCenterDynamics reported.

Despite the legal risk, the IPO filing reveals SpaceX/xAI has committed to $2.8 billion in turbine purchase deals, including a $2 billion contract signed April 30 that's expected to close this month. The environmental tension creates an unusual dynamic: Anthropic, which markets itself as a safety‑conscious AI company, is effectively funding fossil fuel infrastructure through its compute bills.

The IPO filing also unveils perhaps the most ambitious — and speculative — AI infrastructure plan yet: orbital data centers. SpaceX plans to deploy AI compute satellites using Starlink technology, with inter‑satellite lasers creating "coherent computing clusters across free space." The company has requested permission to launch up to 1 million space data center satellites, with a target launch date as soon as 2028.

The filing is unusually candid about the risks, DataCenterDynamics notes, describing the technologies as "novel and untested" and warning investors: "We expect to incur significant capital expenditures over a period of years before our AI products and services become profitable, which may never occur."

The deal's most telling detail is its flexibility. Either party can terminate with just 90 days' notice, The Verge reported. That's an unusually short leash for a $40 billion commitment — it signals that both sides expect the AI compute landscape to shift rapidly, and neither wants to be locked in if better options emerge or the relationship sours.

There's also a ramp‑up discount: Anthropic pays reduced fees in May and June 2026 as capacity comes online, according to AI Business. The full $1.25 billion monthly rate kicks in once the infrastructure is fully operational.

The compute crunch is real and it's expensive. When the company behind Claude — one of the best‑funded AI labs on the planet — is cutting $15 billion‑a‑year checks to a competitor just to keep the lights on, it means the GPU bottleneck is far from solved.

For builders, the implications are clear: API prices aren't coming down anytime soon, because the underlying infrastructure costs are astronomical. The 90‑day exit clause is the one silver lining — it means compute agreements are becoming more flexible, and capacity can shift as new options (Microsoft's Maia, Amazon's Trainium, Google's TPUs) mature. The AI infrastructure game is becoming a multi‑vendor world, and that's good news for anyone who doesn't want to be locked into a single provider's pricing.