The number came out on a stage, the night before the listing, in answer to a question about customers.
At the DealBook summit on Nov. 18, Elon Musk said Anthropic had told SpaceX that the most likely case was another 10 gigawatts of computing capacity. The remark landed a day before the company priced the largest share sale on record, and it described a quantity that no contract covers.
Anthropic did not confirm the figure. A spokesperson for the company, which was in the middle of its offering, had no comment on the scale of any additional capacity.
What exists today is smaller and more specific. The signed agreement covers an entire hall of the Colossus 1 campus in Memphis, more than 300 megawatts of capacity and 220,000 Nvidia processors, on a three-year term at about $1.25 billion a month. That works out to roughly $15 billion a year for a bit more than a third of a gigawatt.
Multiply that rate by the number Musk used and the arithmetic stops behaving. Ten gigawatts at the same price per megawatt would run to something near $500 billion a year, which is more than the ceiling the company has disclosed for the entire relationship through 2029. Either the pricing falls sharply at volume, or the figure describes an intention rather than a deal.
The disclosed ceiling is the more reliable number. SpaceX’s prospectus puts the contract limit with SpaceXAI at $120.5 billion, running to 2029, and states that after an initial period in which neither side can walk away, either party may terminate on 90 days’ notice. A cap and a termination right of that kind make the agreement a framework rather than a mortgage on future revenue.
Ten gigawatts is a physical quantity before it is a financial one. A large nuclear reactor produces about one gigawatt. Ten of them would be a substantial share of the generating capacity built for data centers in the United States in any recent year, and the electricity has to come from somewhere before the chips can be switched on.
That is the constraint that has been binding the industry for two years. Accelerators can be bought with capital; power requires turbines, substations, transmission and an interconnection queue that in most American markets is measured in years. Companies that have moved fastest have done it by building their own generation on site, which brings permits, emissions rules and neighbors.
Memphis has tested that model. The campus has expanded at a speed that local utilities and regulators have struggled to match, and the load it draws is now one of the largest single additions to the region’s grid. Adding ten times that load at the same site is not a matter of ordering more machines.
For Anthropic the attraction is obvious. The company has just raised $107 billion in equity and debt combined and has told investors that compute is the constraint on revenue. A model developer that cannot get capacity cannot serve demand, and the buyers who put money into the listing were buying a growth rate that depends on it.
For SpaceX the arrangement converts capital into a contracted annuity. Building capacity for a single tenant with a three-year commitment and a 90-day exit is a different business from building it for yourself: the returns are lower, the visibility is shorter, and the financing has to be arranged against contracts that can end.
That is where the two stories meet. The $40 billion of chip financing arranged with Apollo the day before runs through a separate entity and depends on the durability of exactly these leases. A customer that can leave in 90 days is a weaker foundation for a lender than one locked in for a decade, and every incremental gigawatt is negotiated against that weakness.
Musk’s own incentives cut in several directions at once. He runs the supplier, he is a large shareholder in the company that just listed a competitor to his own models, and he has spent years arguing that compute is the only currency that matters. Announcing a customer’s ambitions from a stage is a way to advertise capacity that has not been built to buyers who have not signed.
The pattern is familiar to anyone who has watched him forecast. Production targets for vehicles have arrived late. Launch cadences have slipped. Satellite counts have been revised. The announcements have usually described a direction correctly and a date badly, and the market has learned to discount the schedule while accepting the trajectory.
None of that makes the 10 gigawatts false. It makes it a statement about intent, made by one party, about a conversation the other party will not discuss, one day before its shares began trading. Anthropic’s own filings will eventually describe what it has committed to, because a company of that size has to disclose obligations that large.
The question for investors in either company is the same one. Not whether the capacity gets built, but who carries the cost if it is built and the demand does not arrive on schedule. The tenant has a 90-day exit. The landlord has a loan.


