SpaceX’s $74 Billion Bet on a Gas Plant

Drive east from Brownsville, Texas, past the last subdivisions and the birding centers, and the road ends at a flat expanse of tidal mud where a company builds the largest rockets in the world. SpaceX calls the place Starbase. On Thursday it said it will spend $24 billion there on computing, and it will build, own and operate a gas-fired power plant to run it.

The announcement appeared on the company’s website without a press conference. A second facility, in Arizona, carries a $50 billion price tag. Together the two projects come to $74 billion, a figure that is larger than the annual capital budget of most utilities and comparable to a single year of data-center spending at a major cloud provider.

The Arizona project is the bigger and the less explained. SpaceX did not name a site, a county or a power source. A person familiar with the planning said the company has been negotiating land and utility arrangements in the state for months, and that the site selection was driven by access to transmission rather than by tax treatment. Arizona has become a magnet for data centers because land is cheap and the state has allowed utilities to build generation ahead of demand, a policy other states have been slower to adopt.

Power is the constraint that decides these projects now, not chips. Interconnection queues in most of the country run three to seven years, and the gas turbines that data centers want are sold out into the end of the decade. An executive at a power developer that has worked with hyperscalers said the only way to compress a schedule is to bring your own generation, accept that you will own it, and start the permitting before you know what you are cooling.

That is what the Starbase plant is. A gas plant that a company builds for itself skips the queue that a utility would join, though it does not skip the permits. Texas requires air permits for turbines, and the state’s environmental regulator has been processing a growing stack of them from data-center developers. The emissions from a plant of the size implied by $24 billion of computing would be material, and the review will not be quiet.

Water is the second constraint, and the one that produces the most local resistance. Evaporative cooling consumes millions of gallons a day; closed-loop systems use far less but cost more to build and more to run. In Arizona, where groundwater rules differ by basin and where subdivisions already fight over allocation, a facility of this scale would arrive as a political event before it arrived as an engineering one. Voters in several states put data-center measures on ballots this month, and the results suggested that the industry’s assumption of welcome is no longer safe.

Then there is the question of what the computing is for. SpaceX described the buildout as supporting its artificial-intelligence business without saying who the customers are or what the workloads look like. The company’s satellites generate a great deal of data, its rockets are designed with simulation, and its sister companies in the broader constellation of Elon Musk’s ventures have an obvious appetite for training capacity. Whether any of that capacity is sold to outside customers has not been disclosed.

The financing has not been either. SpaceX has revenue from launches and from Starlink subscriptions, and Starlink has become the company’s cash engine. Seventy-four billion dollars is a different order of commitment. A person close to the company said the projects will be funded over several years from operating cash and from outside capital, and that no single financing has been arranged. Analysts said the structure will probably resemble project finance, with the computing assets separated from the launch business so that lenders can be repaid from contracts rather than from the parent.

That separation matters for a company that has raised money in the past against its equity story rather than its cash flows. Launch is a business of fixed prices and government customers. Computing is a business of long contracts, fast depreciation and customers who renegotiate when a newer chip arrives. Lenders treat those differently, and a project-finance structure would force SpaceX to describe demand in a way it has never had to.

The comparison that will be made is with the hyperscalers. Microsoft, Amazon and Google have each committed sums of this size, but they did it from positions of established cloud revenue and with utilities and partners carrying part of the construction risk. SpaceX is proposing to carry it alone, on land it controls, with generation it owns.

There is an argument that this is the cheaper path. A company that controls its own power, its own site and its own construction schedule avoids the markups and delays that come from renting capacity from three parties who each need a return. The counterargument is that a rocket company is now in the utility business, and utilities are regulated, litigated and slow in ways that aerospace is not.

The permits will arrive before the answers do. A gas plant takes two to three years to permit and build; a data shell takes eighteen months; the accelerators inside it will be obsolete before the loans are half repaid. Executives who have watched this cycle said the winners will be the companies that secured electricity first and worried about utilization later.

At Starbase, the launch pad and the power plant will share a horizon. One of them will be finished on schedule.

  • Related Posts

    • November 14, 2026
    • 51 views
    Anthropic Packages Claude for the Factory Floor

    In a machine shop that supplies aerospace brackets, the quoting process has looked the same for two decades. An engineer opens a spreadsheet, pulls costs for material and machine time,…

    • November 14, 2026
    • 50 views
    Chipmakers and Cloud Providers Form an Alliance Around SGLang

    At the SGLang summit on Nov. 12, the slide that mattered was the one with eight logos on it. LMSYS, the research group behind the Chatbot Arena model leaderboard, used…