Key Highlights

  • Nameplate compute was 1.7 gigawatts at quarter end, up from 1.4 gigawatts in the second quarter.
  • About 2.2 gigawatts is expected by the end of October and 2.5 to 3 gigawatts by year end.
  • The chief executive is highly confident of 10 incremental gigawatts in 2027, or about 13 in total.
  • SpaceX is deploying about 20,000 GB300 units a month.
  • Fourth-quarter capital expenditures are guided to $35 billion to $40 billion.

How Fast Is SpaceX Building Compute, and Is a 13-Gigawatt Plan Credible?

SpaceX (NASDAQ:SPCX) ended the third quarter with 1.7 gigawatts of nameplate compute, up from 1.4 gigawatts in the second quarter, and expects about 2.2 gigawatts by the end of October. It targets 2.5 gigawatts to 3 gigawatts by year end, which would be 47% to 76% above the third-quarter level, as calculated.

The chief executive said the company is deploying about 20,000 GB300 units a month, with 2027 deployment focused on Vera Rubin hardware. He said he is highly confident of adding 10 gigawatts in 2027, for about 13 in total, and described 20 gigawatts of incremental power and cooling as difficult but possible.

Why Power and Cooling, Not Chips, Set the Limit

Management described power, cooling and electrical equipment as the main constraints. SpaceX plans to build power and cooling capacity well above what the chips need, so that infrastructure does not hold back deployment.

The chief executive said the company is using its rocket-engine casting capability to make turbine blades and vanes, a part he said runs 12 to 18 months behind the rest of a turbine. He also cited solar, wind and modular inference clusters built with power and cooling at the factory.

The chief executive said he spends about four days a week in Memphis bringing the Colossus II cluster online, and that a training cluster of 440,000 Vera Rubin units is planned for the end of the second quarter.

Short Payback and Spot Pricing Underpin the Economics

The chief financial officer said new deployments pay back in under a year, with pricing slightly above the high end of the $30 to $50 per watt range cited a quarter earlier. The chief executive said the company favors short-term rentals at near spot rates and is seeing contracts renew at higher dollars per watt, and that it aims to split compute roughly evenly between internal use and third parties.

He said a compute shortage is likely to persist for a long time. In a question from JPMorgan on funding, management pointed to its investment-grade rating, several pools of debt capital and possible partnerships with vendors and customers.

He said the company could rent its compute several times over if it agreed to every request, and noted that older H100 chips are renting for more than when they were new.

Outlook: Year-End Capacity and Capital Spending Show Whether the Plan Holds

The chief executive said not everything will go right, which is why 10 gigawatts is the figure he is confident in. Capacity at the end of the fourth quarter, against the 2.5 gigawatt to 3 gigawatt target, and capital expenditures of $35 billion to $40 billion are the next measures.