AI
The Anthropic SpaceX Compute Deal: The 90-Day Clause Changes Everything
SpaceX's S-1 filing disclosed a $45 billion AI compute deal with Anthropic — and a 90-day termination clause that may matter more than the headline number.
The Deal Nobody Actually Read
The Anthropic SpaceX compute deal made headlines on May 20, 2026, when SpaceX filed its S-1 with the SEC. The Anthropic SpaceX compute deal is structured at $1.25 billion per month through May 2029, granting Anthropic access to SpaceX's Colossus 1 and Colossus 2 GPU clusters in Memphis.
That number — roughly $45 billion in total over the contract term — is what every major outlet reported. It is also where most coverage stopped.
Buried in the same disclosure is a clause that changes how the deal should be read: either party can terminate the contract with 90 days' notice. The 90-day exit clause does not look like the foundation of a multi-year strategic partnership. It looks like a short-term lease wearing a long-term contract's clothing.
For context, every prior compute deal of comparable scale has been structured as a multi-year commitment with limited or no early-termination flexibility. Microsoft's original OpenAI agreement included multi-year exclusivity provisions. AWS's anchor-tenant agreements with Anthropic ran on multi-year terms. Google's TPU lock-ins with its own DeepMind unit are effectively perpetual. That structure exists for a reason. The capital expenditure required to build a hyperscale GPU cluster runs into tens of billions of dollars, and the financial models underpinning those builds require predictable revenue.
A 90-day off-ramp throws that model out — and the market has not yet priced what that means.
The Numbers Everyone Quoted (And the One They Missed)
The widely reported figures are accurate. Anthropic will pay xAI $1.25 billion per month through May 2029, with a discounted rate for the first two months while xAI completes its ramp-up. All told, the deal could bring xAI over $40 billion in revenue. The transaction grants Anthropic access to approximately 220,000 state-of-the-art Nvidia GPUs and 300 megawatts of compute capacity across the Colossus facilities.
Here is the calculation no outlet ran: at $1.25 billion per month spread across 220,000 GPUs, Anthropic is paying roughly $5,680 per GPU per month, or approximately $68,000 per GPU per year, just for compute access.
NVIDIA's H100 hardware list price has historically been quoted around $25,000 to $30,000 per unit. The H200 and GB200 generations are priced higher but still well within five figures per unit on a hardware basis. The implication: Anthropic is paying roughly 2 to 3 times the underlying hardware cost per GPU per year — every year — to rent compute rather than own it.
That ratio is not unreasonable when you factor in power, cooling, networking, real estate, depreciation, and the cost of building and operating a hyperscale facility. Data center economics carry real operational overhead beyond the hardware itself. But it does explain the gravitational pull behind every frontier AI lab simultaneously trying to lock down its own infrastructure: at $68,000 per GPU per year in rental costs, the build-versus-buy math gets uncomfortable fast.
If a 220,000-GPU cluster can be rented for $15 billion per year, the same cluster can theoretically be owned outright for roughly $30 to $40 billion in upfront capex plus annual operating costs in the low billions. The break-even horizon is two to three years. Every AI lab with the balance sheet to absorb that capex is doing the same calculation.
What "Monetize Unused Capacity" Actually Means
The most revealing sentence in the entire S-1 disclosure is one that almost no outlet quoted in full. SpaceX wrote that the arrangement "allows us to monetize unused compute capacity in our infrastructure."
Two words in that sentence carry the weight: unused capacity.
xAI built Colossus 1 in Memphis to train Grok. The facility came online in late 2024 after a build that Musk's team treated as a competitive flex — completed in 122 days, which became part of xAI's marketing narrative around execution velocity. Colossus 2 followed. The capacity was built for xAI's own model training needs, not as a commercial cloud product.
Eighteen months later, SpaceX is telling the SEC that this capacity is unused — and the company is renting it to a rival lab to monetize it.
This admission matters in two directions. First, it implies xAI's own compute demand has not scaled to match its build-out. That tracks with the broader concern that xAI posted $818 million in revenue and a $2.47 billion operating loss in the first quarter of 2026 — a loss rate of approximately $3 for every $1 of revenue. A company burning that fast does not have an obvious path to filling 300 megawatts of compute internally. Renting to Anthropic is not a strategic pivot. It is monetizing an asset that was built ahead of demand.
Second, the disclosure tells competitors that the largest pure-play AI infrastructure builder in the United States has surplus capacity available on a spot basis. That is not what a tight compute market looks like.
There is a deeper concern embedded in the xAI loss rate. A counterparty losing $3 for every $1 of revenue is, by definition, dependent on continued external funding to operate. If xAI's next funding round is harder than expected, the Colossus facilities themselves could become collateral, restructured, or sold. Anthropic's $15 billion annual outflow is going to a company whose own underlying business is unprofitable by an order of magnitude. The 90-day clause protects Anthropic against this risk — but only if Anthropic is willing to use it.
The Microsoft Trade — Why the Anthropic SpaceX Compute Deal Isn't Standalone
The Anthropic SpaceX compute deal does not exist in isolation. In late 2025, Anthropic agreed to purchase $30 billion of Azure compute capacity from Microsoft, with additional capacity reaching up to 1 gigawatt, including access to Nvidia's Grace Blackwell and Vera Rubin systems. The deal was tied to a Microsoft investment of up to $5 billion and an Nvidia investment of up to $10 billion into Anthropic, pushing the company's valuation toward $350 billion.
That Microsoft deal alone is more than double the headline value of the SpaceX agreement. Add the SpaceX $45 billion contract on top, and Anthropic has committed to roughly $75 billion in compute spending across two providers, before factoring in its existing AWS relationship — which preceded both deals and remains active.
Three observations follow from that distribution.
One: No single compute provider has lock-in over Anthropic. The company has structured its infrastructure to be portable across at least three hyperscalers — AWS, Microsoft Azure, and SpaceX's Colossus network. If any one provider raised prices, degraded service, or experienced an outage, the others could theoretically absorb the workload. That portability is itself a strategic asset.
Two: The 90-day clause in the SpaceX deal is not an outlier. It is the natural endpoint of a strategy that treats compute as a fungible input, not a strategic partnership. Anthropic does not need SpaceX specifically. It needs compute. The 90-day clause makes explicit what the multi-provider structure already implied: this is procurement, not partnership.
Three: SpaceX's own S-1 contains a sentence that confirms the pattern is generalizing across the industry. The filing states the company "expects to enter into additional similar services contracts." Translation: more 90-day, multi-billion-dollar compute rentals are coming, both for Anthropic and for other AI labs. The neocloud business model — building data centers and selling capacity to whichever AI lab needs it that quarter — is being normalized in real time, and Anthropic is its largest customer to date.
For frontier AI labs, the implication is that compute scarcity, the dominant narrative of 2024 and most of 2025, may be transitioning into compute oversupply faster than the headlines suggest.
The 90-Day Clause Is the Real Story
In commercial real estate, a tenant who can leave on 90 days' notice is not an anchor tenant. They are a short-term lessee, and the landlord cannot use their lease to secure financing for the building. Banks underwriting a commercial mortgage want to see ten-year leases from credit-worthy tenants. Anything shorter forces the landlord to fund the project through equity or higher-cost debt.
The same principle applies to compute infrastructure. The financial model for a $20 to $30 billion data center build assumes predictable, multi-year revenue from anchor customers. When the largest contract on the books can be canceled within a fiscal quarter, the certainty premium disappears. Lenders price that uncertainty into the cost of capital.
This is where the 90-day clause becomes more than a contractual footnote. It directly affects SpaceX's ability to finance Colossus 3 and any subsequent data center expansions. If the bond market and SpaceX's institutional lenders cannot count on Anthropic's $15 billion in annual revenue, the company has to either fund future builds with equity (diluting shareholders) or accept higher interest rates on debt. Both outcomes pressure the IPO valuation.
SpaceX is heading into a roughly $1.75 trillion IPO valuation with the Anthropic contract as one of its proof points for AI infrastructure revenue. The market will need to decide how much credit to give a $15 billion annual revenue line that the counterparty can walk away from with one quarter's notice. Sophisticated institutional investors will discount it. Retail investors reading the press release may not.
A useful comparison: when Netflix signed its original content licensing deals with HBO for distribution rights, those were multi-year commitments. When Spotify locked in deals with the major record labels, those were multi-year. Long-form contracts exist because both sides accept reduced optionality in exchange for predictability. The recipient of the cash flow can finance against it. The buyer of the service can plan capacity against it.
The 90-day Anthropic SpaceX compute deal runs the other direction. Both sides preserved their ability to exit. Both sides preserved their ability to renegotiate. Both sides preserved their ability to redirect that compute or that capital somewhere else if circumstances change. The flexibility is mutual, and that is precisely the problem for anyone trying to value the contract as a long-term asset.
That is not a moat. That is a spot trade.
What This Means for the IPO Wave
Three AI-adjacent IPOs are now in motion in a six-month window. SpaceX filed its public S-1 on May 20, 2026, targeting a June 2026 Nasdaq listing at a roughly $1.75 trillion valuation. OpenAI is filing a confidential S-1 with the SEC, targeting a September 2026 listing at approximately $852 billion to $1 trillion. Anthropic is targeting an October 2026 listing at a roughly $900 billion valuation. The combined market capitalization implied by these three filings exceeds $3.5 trillion.
The Anthropic SpaceX compute deal is the financial bridge connecting all three.
For SpaceX's IPO roadshow, the $15 billion annual revenue stream from Anthropic dresses up the company's growth profile beyond launch services. Launch revenue is real but capacity-constrained by physics. AI infrastructure revenue scales differently, and Wall Street will assign it a higher multiple. For Anthropic's eventual IPO, locked-in compute supply demonstrates the company can serve growing demand without execution risk. The deal was disclosed exactly when both sides needed it to support their public market narratives.
The 90-day clause is what allows both sides to commit to a $45 billion headline number without taking on $45 billion in actual long-term exposure. It is a contract optimized for the prospectus, not for compute strategy.
There is also a counterparty risk angle that Anthropic shareholders should weigh as the company approaches its own listing. Anthropic is sending $15 billion per year to a company whose AI business is losing $3 per $1 of revenue. If xAI restructures, gets acquired, or significantly cuts capacity, Anthropic's primary inference capacity for Claude could be disrupted on short notice. The 90-day clause cuts both ways — it protects Anthropic from being locked in, but it also means SpaceX is not locked in to providing the service. Either side leaving creates operational risk for the other.
For investors evaluating any of the three IPOs, the question is whether the Anthropic SpaceX compute deal represents durable revenue or a financing-friendly arrangement that gets renegotiated as soon as both companies are public. The S-1 leaves both readings open. The 90-day clause is the reason.
The deeper question — the one that should shape how the entire AI infrastructure category is valued — is whether compute itself has become a commodity input rather than a strategic moat. If the answer is yes, the IPO valuations across SpaceX, OpenAI, and Anthropic deserve sharper scrutiny than the headline AI narrative implies.