The line sat deep in a securities filing, phrased the way financial lawyers prefer: one customer, identified only as “Customer B,” generated about 34.5% of revenue last quarter.
People familiar with the matter say Customer B is Anthropic. The division that booked the money is now called Super Intelligence, and what it sells is not a launch but a data center — accelerators humming in Memphis, Tenn., rented by the hour and billed by the month.
For a company whose reputation rests on reusable boosters and satellite broadband, the arrangement introduces a business with an entirely different rhythm. More than a third of quarterly revenue now arrives from a single tenant, under contracts that run to 2029 and carry a stated ceiling of $120.5 billion. Starting in December, the bill comes to roughly $3.1 billion a month.
The figures cut two ways. Anthropic, an artificial intelligence developer backed by Amazon and Google, will spend more on computing capacity in a month than most semiconductor companies book in a quarter. SpaceX, for its part, has converted spending on buildings, power and cooling into a revenue line that investors can put a multiple on.
Shares fell 2.9% in after-hours trading once the filing circulated. The reaction was less about the money than about the concentration. A customer that size is a customer with alternatives, and the paperwork gives Anthropic several.
According to the disclosure documents, Anthropic can cancel the relevant agreements after January and March 2027, given 90 days’ notice. Clauses of that kind are common in large compute contracts, where buyers want room to shift workloads as chip generations turn over. They also mean the revenue is not guaranteed for the full term. The capacity is reserved; the invoices are not.
Analysts who follow the company said the disclosure changes how the business should be read. One described the structure as a launch operation with lumpy government contracts, a broadband business with subscriber churn, and now a third leg that behaves like a data center landlord with one anchor tenant. He spoke on condition of anonymity because the filing is recent.
That third leg is the newest part of SpaceX and, on this evidence, a material slice of the top line. It is also the part with the shortest history. Launch contracts are signed years ahead with agencies and satellite operators. Broadband subscriptions renew monthly. A compute lease sits between the two: long enough to justify construction, short enough that the customer can leave.
The economics of such arrangements have become familiar. Building capacity takes years — land, turbines, transformers, permits, then the racks themselves. Selling spare capacity to one large buyer spreads the fixed cost of that construction and turns a cost center into revenue while the owner’s own workloads ramp up. Every large operator of data centers has drifted toward the same logic, and the ones that build fastest are the ones with a tenant already signed.
For Anthropic, the appeal is speed. The company has assembled capacity across several providers, a strategy meant to keep any one supplier from holding pricing power over it. Leasing from an operator that already has power contracted is faster than building, and faster still than waiting in an interconnection queue. The trade-off is a bill that arrives whether or not the models sell.
For SpaceX, the appeal is the reverse: a tenant that pays in dollars rather than tokens, and commits for years. Contracts of this shape function as infrastructure financing, with the customer’s credit standing standing in for a bank’s. What the filing does not disclose is the margin. SpaceX does not break out the cost of the underlying buildings or the depreciation applied to them, and without those figures the 34.5% describes scale rather than profitability.
The concentration itself is the item regulators and index committees tend to notice. Public companies are generally expected to flag customers above 10% of revenue, and SpaceX has done so. A single buyer at 34.5% is unusual outside defense contracting, where the buyer is a government and the relationship is measured in decades rather than quarters.
There is a symmetry that cuts the risk. Anthropic needs the capacity; SpaceX needs the revenue. Neither can walk without cost. The cancellation windows in early 2027 are the tell — they fall roughly when the current generation of accelerators will be superseded and when Anthropic expects to know whether its own demand forecasts were right. A tenant that leaves in January 2027 is a tenant that concluded it had bought too much.
Whether the arrangement becomes a template is a separate question. AI developers have now signed capacity agreements with cloud providers, chipmakers and a rocket company. Each deal moves a slice of the buildout onto a balance sheet that is not the developer’s, which works while capital is abundant and works less well when the tenants decide they have overbought. The revenue is real either way. The question is how long it lasts.
For the quarter just reported, the ledger reads plainly. One tenant, one third of revenue, one clause that lets it leave.
