Anthropic to pay xAI $1.25B per month under 90-day exit clause

Anthropic locked itself into paying xAI $1.25 billion per month — but only for 90 days at a time.

The SpaceX S-1 filing, submitted to the SEC on May 20, 2026, reveals that Anthropic's record compute deal with xAI includes a 90-day cancellation clause: either party can exit the $1.25 billion-per-month commitment with 90 days' notice, according to The Information and the S-1 itself. Anthropic will spend roughly $15 billion per year for access to xAI's Colossus data centers in Memphis — TechCrunch confirmed the counterparty is xAI, not SpaceX proper — a sum that exceeds SpaceX's roughly $18 billion in annual revenue, making a single customer on track to represent roughly 83 percent of SpaceX's total revenue, Axios reported based on the filing. The 90-day clause is the detail that changes the picture.

On the surface, $15 billion per year is an extraordinary commitment. But a 90-day walkaway right is not the contract structure of a company that has fully committed to its compute provider. It is the contract structure of a company that wants optionality — that wants to be able to exit if xAI's buildout slips, if a data center faces power-grid or regulatory problems, or if a better option appears. The guarantee is three years of dedicated infrastructure and power. The exit door is 90 days wide.

That door swings both ways. xAI can walk away too, which means the world's most closely watched AI safety company — the one that sells Claude to enterprises and advises the US government on AI policy — has written its training compute roadmap into a contract that the world's most prominent Elon Musk-affiliated company can terminate on 90 days' notice. If xAI triggers a dispute, misses a capacity milestone, or simply decides it needs the compute elsewhere, Anthropic's model schedule becomes a hostage to someone else's manufacturing timeline.

Anthropic is renting rather than building — a deliberate choice to move faster and avoid locking capital into infrastructure that may be outdated in three years. The strategic rationale is coherent. But renting at this scale, with a 90-day exit clause on both sides, means Anthropic's entire compute roadmap is a line item in xAI's production schedule with 90 days of protection. If a data center buildout slips, a chip supply bottleneck tightens, or an xAI facility faces a regulatory or power-grid setback, Anthropic has 90 days to find an alternative before its training runs stop.

The deal structure also raises questions about how SpaceX's IPO investors should value the Anthropic commitment. The S-1 lists the Anthropic compute contract among its material agreements — but a 90-day termination clause means the 83 percent revenue concentration figure could change quickly. Investors will need to read the risk factors section closely for what disclosure, if any, SpaceX provides about termination rights, exit clauses, or conditions that could trigger the cancellation provision on either side.

This is the real AI arms race — not over who produces the smartest model, but over who controls the physical floor underneath AI: the power, the chips, the data center floor space. Nvidia reported $81.6 billion in revenue in the first quarter of its fiscal year ending January 2027, up 85 percent year over year, with $58.3 billion in net income and 75 percent gross margins, with CFO Colette Kress projecting the Vera CPU line will reach $20 billion in annualized revenue, according to Nvidia's investor relations disclosure. The numbers are extraordinary. But they are also testimony to the same underlying reality: the AI boom is a power boom. Economic value is migrating toward the companies that control kilowatts and compute floor space, not toward the companies that write the most elegant training code.

The Trump administration walked back a planned AI executive order on May 21, 2026, after David Sacks — an investor with direct ties to multiple AI ventures — raised concerns inside the White House, Politico reported. The episode illustrates how entangled AI infrastructure policy has become with the companies building it. A data center is no longer just a building with servers. It is a national-interest object, which means the companies that control data centers have leverage that has nothing to do with their models.

On the All-In podcast, David Friedberg and Gavin Baker raised a possibility that has moved from fringe to plausible in the past year: that a CCP-funded campaign to discredit AI development and delay data center construction in the United States is not a hypothetical but an active strategy, the podcast summary noted, and that evidence is accumulating that such efforts are underway. Whether or not that framing holds, the underlying structural point is real. The AI industry is building its core infrastructure in a small number of locations, on a small number of grids, dependent on a small number of suppliers for the most advanced chips and the most advanced compute capacity. Concentration is a feature when the goal is speed. It is a liability when the goal is resilience.

What to watch: whether SpaceX prices the Anthropic deal as a structural asset — proof that the world's most consequential AI company needs compute badly enough to commit more than the rest of the market combined — or as a single-customer risk that a 90-day exit clause makes suddenly real. The answer will tell us whether the market treats AI infrastructure concentration as a moat or a vulnerability.