On May 20, 2026, SpaceX filed its S-1 with the SEC ahead of a Nasdaq debut targeted for June 12 under the ticker SPCX. The filing sets a roughly $1.75 trillion valuation on the company and lists a $75 billion primary raise — the largest IPO on record if it prices anywhere near the range, per The VC Corner's teardown and BigGo Finance's summary.

The headline number is the valuation. The interesting number, buried in the customer-concentration disclosure, is what Anthropic is paying for Colossus.

Three weeks ago The Loop wrote up Anthropic's all-of-Colossus-1 lease and noted that "financial terms were not disclosed." The S-1 disclosed them. The bill is $1.25 billion per month, payable by Anthropic to SpaceX, through May 2029, with a discounted ramp for May and June 2026 and a clause that lets either side terminate on 90 days' notice. TechCrunch, Teslarati, and Simon Willison's direct quote from the filing all agree on the structure.

Roughly $15 billion a year. $45 billion across the term, if it runs to expiry. SpaceX's own projection in the filing is "over $40 billion in revenue" from the contract over its life. The deal is the largest single line in the prospectus, and the entire concentrated-customer-risk section of the S-1 is, in effect, about Anthropic.

What $15 billion a year buys

The compute is the same kit the May 7 announcement described: Colossus 1 in Memphis, more than 300 megawatts and 220,000+ NVIDIA GPUs across H100, H200, and GB200, plus shared access to Colossus 2 next door. Anthropic gets dedicated capacity on Colossus 1 and overflow on Colossus 2 for the larger training runs. SpaceX retains the right to use the campus for its own training and inference.

The S-1 puts that last point in writing: "We have sufficient capacity to provide compute for our own AI models, including support of our training and inference demands, and to satisfy the obligations under these agreements." That sentence is doing two jobs at once. To Anthropic, it is a guarantee that the contract is honoured. To the SEC, it is a statement that selling all of Colossus 1 to a competitor does not, somehow, leave xAI short.

Both can be true. Both also imply that xAI built more than it ended up needing.

The xAI segment, by the numbers

The S-1 breaks SpaceX into three reporting segments. The first two are familiar. Starlink booked $11.4 billion in 2025 revenue with $4.4 billion of operating income — the cash engine. Launch booked $4.1 billion in revenue and $657 million of operating loss — a strategic line item, not a profit centre. The third segment is the new one: xAI, post the February merger that folded the AI lab into SpaceX, posted $3.2 billion in 2025 revenue against a $6.4 billion operating loss, per The VC Corner's read of the segment table.

A $6.4 billion loss on $3.2 billion of revenue is not the kind of math an IPO roadshow wants to lead with. xAI's 2025 capex was $12.7 billion. Q1 2026 alone was $7.7 billion, an annualised pace north of $30 billion. The Anthropic contract — $15 billion a year of recurring revenue, signed before the prospectus dropped — is the largest available counterweight to those numbers. It does not fix the segment economics. It changes the shape of the chart the underwriters show.

TechCrunch's framing is the polite version: xAI's Grok usage "dropped significantly" in recent months, freeing capacity now being sold to "one of its closest competitors." The blunt version is that xAI built the cluster on the assumption Grok demand would fill it, Grok demand did not fill it, and Anthropic — sitting on the 80x Q1 growth Dario Amodei announced at the developer conference — turned out to be the customer who would.

The S-1 calls this "our dual monetization strategy." The phrase is doing a lot of work. Dual monetization is what it is called when you build a frontier cluster for your own model, the model does not draw the expected traffic, and you rent the silicon to the lab whose model did.

The 90-day clause, read from the prospectus side

The Loop's May 7 piece led with Elon Musk's "no one set off my evil detector" post and the clause letting him reclaim the cluster if Anthropic "engages in actions that harm humanity." From the safety-discourse angle, that clause was the headline.

The S-1 prefers a different framing. The same clause is described as a mutual 90-day-notice termination right exercisable by either party, with no mention of harm or evil — standard cloud-lease language that any underwriter would want in a megacontract concentrated on one customer. The reason both descriptions can exist on the same contract is that the contract is large enough to be written about in two registers.

For the prospectus, the 90-day window is risk-disclosure boilerplate. For Musk's X post, the same window is a moral kill switch. The S-1 wins on legal weight; the X post wins on virality. The text of the contract is identical either way.

What the 90-day clause actually does, financially, is cap each side's exposure to the other at one quarter of revenue or one quarter of compute. Anthropic can leave if a better deal emerges — by, say, Trainium delivery accelerating. SpaceX can leave if Grok 5 suddenly fills its training queue. Neither side is locked in for three years; they are committed to the next 90 days, on rolling renewal. The $45 billion total assumes nobody pulls the cord. Roadshows are built on those assumptions.

What this means

Three takeaways.

1. The Anthropic-xAI relationship is now a top-line item on a public-company prospectus. That cuts in two directions. It locks both sides into a public, audited disclosure cadence that the press will quote every quarter. It also gives Anthropic a counterparty whose IPO economics depend, in non-trivial part, on Claude's training demand staying high — a sharper alignment of interests than any vendor contract on the previous frontier round.
2. xAI's neocloud move is not a strategy, it is a remedy. Selling spare capacity to a competitor before your IPO is a thing companies do when the original plan — owning the demand for that capacity — did not arrive. SpaceX's "dual monetization" language is the polished version. The blunter read is that the merger that folded xAI into SpaceX in February is now visibly funded, at the segment level, by the lab xAI's models were built to replace. That is not stable forever. It is stable through May 2029, give or take 90 days.
3. The Anthropic compute bill is now a known unit of measure. Fifteen billion a year, on one contract, with one provider, for one stretch of the model life cycle. When the next big lab deal lands — and one will — the trade press will index it against this number. The Loop expects to see "a Colossus-scale lease" used as a unit of compute commitment within the quarter, the same way the industry started measuring valuations in "OpenAI rounds" in 2024.

The S-1 prices a deal The Loop covered three weeks ago. It also reveals that the customer who closed the largest single compute lease of the year is paying the parent of the model that lease was meant to train. The cluster is the same cluster. The check is now in writing.