Anthropic Will Pay SpaceX $120.5B, Until Either Side Walks

By Daniel Rao

Anthropic has just told prospective investors it could pay SpaceX up to $120.5 billion for artificial intelligence compute — and then, in the same filing, handed both companies a 90-day exit. That is the strangest sentence in a 318-page IPO prospectus full of strange sentences, and it is the one that matters most. A $120.5 billion commitment sounds like a marriage. Read the small print and it looks more like a very expensive rolling lease, renewable every quarter, between two companies that also compete with each other.

The figure comes from Anthropic's Form S-1, filed publicly with the US Securities and Exchange Commission on Monday, 26 October 2026. We checked our own coverage of that filing from the same day, Anthropic Filed Its IPO — With an $80.7B Loss on Page One, and the $120.5 billion number does not appear in it at all. The first-day story was, understandably, about the $80.7 billion first-half loss. This is the filing detail hiding underneath it, and it reframes what Anthropic actually is: not just a model company, but one of the world's largest buyers of rented computing power, and SpaceX one of its biggest landlords.

Photo: File photo — a Falcon 9 carrying the Starlink 6-86 mission lifts off from Kennedy Space Center on 1 December 2025. This is a picture of SpaceX's rocket business, not of a data centre; no freely licensed photograph of the Colossus compute facilities behind this deal was available. Credit: Gwendolyn Kurzen, US Space Force, via Wikimedia Commons, public domain.

What the filing actually says

Between May and October 2026, Anthropic says it "entered into and amended several agreements" with SpaceX.AI LLC — the wholly owned SpaceX subsidiary behind the company's AI compute business, formerly X.AI LLC — "to access compute powered by NVIDIA GPUs". Under those agreements, it will spend "an aggregate of up to $120.5 billion over terms extending into 2029".

Once all the contracted capacity has been delivered, which Anthropic expects in December 2026, the bill comes to roughly $3.1 billion a month. That is about $37 billion a year, every year, for as long as the arrangement holds at full tilt — more than double Anthropic's entire first-half revenue of $16.29 billion, and a useful measure of just how lopsided the economics of frontier AI have become. Fees are different while capacity is still ramping up, the filing notes, which is the polite way of saying the meter is already running but has not yet hit full speed.

The number is also a sharp escalation from where this relationship started. In SpaceX's own IPO prospectus in June, the original May agreements were described as $1.25 billion a month through May 2029, covering about 325,000 NVIDIA GPUs. Reuters reporting on the then-confidential Anthropic draft had put the SpaceX deal at up to $84.5 billion through 2029. The public filing's $120.5 billion is roughly $36 billion more than that earlier reported figure. Something grew, a lot, between the confidential draft and the public document — more capacity, amended terms, or both. The filing does not break down how much Anthropic has paid SpaceX so far, which facilities host the capacity, or how the amended agreements divide between the separate contracts. Those are the obvious questions for the roadshow, and investors should ask them out loud.

The exit clause is the story

Here is the part that turns a headline into a real story. The $120.5 billion is a ceiling, not a promise. Under the separate agreements, the initial non-cancelable periods end in January 2027 and March 2027. After that, either Anthropic or SpaceX can cancel with 90 days' written notice.

Think about what that means in practice. Anthropic's biggest single compute bill — $3.1 billion a month at full delivery — rests on contracts that, within months, either side can walk away from in a quarter. That is not a criticism of the deal; it may be exactly why the deal was possible at this size. A ten-figure monthly commitment with no exit would be reckless for a company that has not yet proved it can sustain its current revenue pace. A ten-figure monthly commitment with a quarterly exit is a bet that can be resized if the market turns, if a cheaper supplier appears, or if Anthropic's own model economics change.

But the clause cuts both ways, and the sharper edge belongs to SpaceX. Anthropic's Claude models compete directly with Grok, SpaceX's own chatbot. Renting GPUs to a rival is a profitable business only until you need those GPUs yourself. SpaceX has said its contracts carry early outs precisely so it can reclaim capacity for its own models. If Grok's training demands spike, or if a higher-paying customer appears, Anthropic could find its landlord has other plans. A tenant with a 90-day notice period is not a tenant with security of tenure, no matter how large the cheque.

The September clue that now makes sense

None of this arrived out of nowhere. On 10 September 2026, SpaceX chief financial officer Bret Johnsen told the Goldman Sachs Communacopia and Technology Conference that the company had closed a hosting deal worth about $1.11 billion a month starting 1 December, without naming the customer. "Earlier this month, we closed another hosting deal," Johnsen said. "By the end of this year, with annualizing our December number, we're on track to hit $100 billion of ARR."

It is tempting to join those dots and declare the unnamed customer was Anthropic. Do not. SpaceX did not name the customer in September, and the S-1 does not say the September deal is one of these agreements. The monthly figures do not even match neatly: $1.11 billion is a long way from $3.1 billion, though ramp-up pricing, partial capacity and separate agreements could all explain part of the gap. What can be said, safely, is that September was the month SpaceX started talking about its compute business in the same breath as its $100 billion annualised revenue target — and the S-1 now shows why that confidence was not bluster. Somebody very large was signing very large cheques. The filing still does not prove it was the same somebody.

What the September disclosure does establish is the template. Johnsen described SpaceX's compute contracts as short initial commitments with a 90-day exit, a structure the company says still supports a payback period of less than a year on the hardware. Speaking alongside SpaceX's $40 billion GPU financing announcement on 26 October, Johnsen repeated the claim in plainer terms: "we're seeing less than one year payback periods on our SI infrastructure investments." That is the landlord's pitch in one sentence. Buy the chips, rent them out, get your money back inside twelve months, and keep the right to take them back. If the pitch holds, the economics are genuinely formidable. If compute prices fall — and they have a habit of falling — the payback maths gets a lot less friendly, and the 90-day exits start to look less like flexibility and more like fragility.

SpaceX, the landlord

Step back and the transformation is genuinely startling. SpaceX's third-quarter results, reported last week, showed revenue up 161% to $14.58 billion, with its AI unit — the business now called Super Intelligence, renting data-centre capacity to outside customers — generating about $7.6 billion of that total, overtaking Starlink. Capital spending hit $19.6 billion in the quarter, $16.4 billion of it on AI infrastructure. This is a company spending more in three months on compute than most countries spend on science, funded in part by renting that compute back to the very labs racing it.

The S-1 adds one more suggestive data point, with a warning label attached. SpaceX's third-quarter report said a single customer in the segment that includes its AI computing business accounted for 34.5% of its $14.58 billion in third-quarter revenue. That works out to just over $5 billion in the quarter. But SpaceX does not name that customer, and says the customer was not above 10% of revenue a year earlier. It would be journalistically shoddy to print that customer's name as Anthropic on the basis of arithmetic alone. The number tells you concentration risk is real — one customer is now more than a third of a $14.58 billion quarter — but it does not tell you who. Anyone who tells you otherwise is guessing, and guessing with other people's money.

Keep the $40 billion GPU financing separate, too. On the same day Anthropic's S-1 went public, SpaceX closed a roughly $40 billion master finance lease for NVIDIA Vera Rubin chips and related equipment, arranged through Apollo Global Management — the deal we covered in SpaceX Closed Its $40B GPU Bet, Then Musk Warned Neoclouds. That is SpaceX borrowing to buy hardware. The $120.5 billion is Anthropic potentially paying to use hardware. They are two sides of the same buildout, not the same transaction, and conflating them double-counts the boom.

Why Anthropic signed this at all

For Anthropic, the logic is brutal and simple: there is no frontier model without frontier compute, and there is no frontier compute available at this scale without signing deals that would have looked insane three years ago. Claude's revenue growth — $16.8 billion in the third quarter alone, up 45% sequentially on investor-meeting figures — gives the company the cash flow to contemplate a $3.1 billion monthly bill. It does not give it immunity. The S-1 is candid that most of its customers can cut back in any given month, while its own biggest bills are measured in years and quarters. That mismatch — monthly revenue optionality on one side, quarterly-locked compute bills on the other — is the central risk of the entire AI boom, and Anthropic has just published it in numbers.

There is also a quieter strategic point. By spreading its compute across SpaceX, Google, Amazon and Microsoft, Anthropic avoids being wholly hostage to any one supplier — but it also makes itself a customer of its competitors in every direction. Google sells Claude's rival Gemini. Amazon and Microsoft sell their own models and host OpenAI's. SpaceX builds Grok. In this market, your landlord is very often your rival, your investor is very often your distributor, and the only thing that is not negotiable is the electricity bill. Anthropic's answer, apparently, is to rent from everybody and bet that scale covers the awkwardness.

Our judgment: a ceiling, not a verdict

So what should a reader actually take from the $120.5 billion? Not that Anthropic will definitely pay it — the exit clause makes that claim false on its face. Not that SpaceX has banked it — revenue is recognised as capacity is delivered and used, not when a ceiling is printed in a prospectus. And not that the AI buildout is either a bubble or a sure thing, because this filing supports both readings and settles neither.

What it does prove, colossally and usefully, is that the frontier AI business has become a real-estate business with a research lab attached. The scarcest asset is no longer the cleverest model architecture; it is power-backed, permitted space full of NVIDIA chips, available this quarter, at a price someone will actually pay. SpaceX owns a growing share of that space. Anthropic needs a growing share of that space. The $120.5 billion is the price tag they have put on needing each other — and the 90-day clause is the price tag they have put on not trusting each other. Both numbers are honest. Only one of them made the headlines.

The rest of the bill, in one paragraph

SpaceX is not even Anthropic's largest commitment in the filing. Anthropic has agreed to spend at least about $114.5 billion with Google and $110 billion with Amazon, and at least $44.9 billion with Microsoft, alongside five-year Broadcom equipment leases carrying about $161.2 billion in rent. Those figures sit in the same "most significant arrangements" section of the S-1, and unlike the SpaceX deal most of them are minimums that apply regardless of usage. Read together, they describe a company that has rented its future several times over, from several different landlords, before it has finished proving that future exists. That is either the boldest infrastructure bet in corporate history or the largest collection of exit clauses ever assembled. The next twelve months — the non-cancelable periods expiring, the December capacity delivery, the IPO roadshow — will tell us which.

What to watch next

Three dates matter. December 2026, when Anthropic expects all contracted SpaceX capacity to be delivered and the $3.1 billion monthly rate to kick in. January and March 2027, when the non-cancelable periods expire and the 90-day clock starts running in both directions. And the IPO itself, when Anthropic will have to persuade public investors that a company paying one supplier up to $120.5 billion, cancellable quarterly, deserves a valuation in the trillions. Watch whether the roadshow leans on the ceiling or the exit clause. That choice will tell you what management actually believes this deal is.

Related reading: our earlier reporting on Anthropic's capacity deals beyond the hyperscalers, Anthropic's Secret Bitcoin-Miner Deal Caps a $530B Sprint — note that the figure in that headline comes from earlier reporting, not from this S-1 filing, and the two should not be added together.

Sources: Anthropic Form S-1 (filed 26 October 2026), as reported by Barron's, CNBC (via coverage of the filing), Gizmodo (citing Reuters' earlier reporting on the confidential draft), TeslaNorth and AIStockWire; SpaceX third-quarter 2026 results and September 2026 Goldman Sachs Communacopia remarks by chief financial officer Bret Johnsen, as reported by Drive Tesla Canada, CoinCentral, AIStockWire and TipRanks/TheFly; SpaceX June 2026 IPO prospectus figures via AIStockWire; TechPulse USA's own 26 October S-1 coverage, checked for this article.

Daniel Rao is a TechPulse USA journalist covering AI infrastructure, data centres and the companies racing to power them.