Terafab: Inside Musk’s Plan to Make His Own Chips, and Send Most of Them Into Orbit

On March 21, 2026, at the former Seaholm Power Plant in Austin, Elon Musk announced that his companies would stop waiting for the world’s chipmakers. “We need the chips,” he said, “so we’re going to build the Terafab.” He called it “the most epic chip building exercise in history by far.”¹

The name is a deliberate step up from the Gigafactory, Tesla’s battery plant, and the ambition matches it. Musk says Terafab will produce one terawatt of computing power a year, and that all the chip factories now on Earth make only about 2 percent of what his projects need.¹

The terawatt ambition

Modern chipmaking is a relay race run across continents. A chip may be designed in California, made on lithography machines from the Netherlands in a fab in Taiwan, paired with memory from South Korea and packaged somewhere else again. Terafab proposes to run every leg in one place: chip design, lithography, fabrication, memory, advanced packaging and testing, all under one roof in Texas.¹

The numbers are vast. At launch, Musk described a target of 2-nanometer chips, starting at 100,000 wafers a month and scaling to a million, and an output of 100 billion to 200 billion custom AI and memory chips a year.¹ In August, Tesla and SpaceX confirmed the site: Grimes County, about an hour northwest of Houston, on a reservoir that once cooled a coal plant. They put the first phase at about $16.8 billion, promised at least 3,000 jobs, and said the finished complex would exceed 100 million square feet, which they call the largest chip plant on the planet.² A research fab, the project’s precursor, broke ground at Tesla’s Austin Gigafactory in April.²

The money, though, depends on which document you read. SpaceX’s own IPO filing proposed $55 billion for an initial phase, and as much as $119 billion if later phases go ahead. The same filing called Terafab a “general framework,” warned there was “no assurance” it would meet its objectives on time “or at all,” and reportedly listed no binding commitments between the two companies.^2,3

Skeptics have plenty to point to. Electrek, which covers Tesla closely, noted that the company has no semiconductor manufacturing experience, that leading-edge fabrication is far harder than building cars or batteries, and that a single 2-nanometer fab in the United States costs roughly $28 billion and takes more than three years to build. At full scale, it calculated, Terafab would match about 70 percent of TSMC’s current output.¹ The partners are not starting from zero: in April, Intel said it would join the project and help “refactor” chip fabrication, though it has not said exactly how.⁴

The orbital strategy

The most startling detail is where most of the chips are meant to go. Musk said 80 percent of Terafab’s compute would be for AI satellites in orbit and only 20 percent for use on the ground, in Tesla’s cars and Optimus robots.¹ The space-bound chips have their own name, D3, and are custom-designed for orbital AI satellites.¹ In orbit they will have to survive radiation that degrades ordinary electronics.

His argument rests on physics. Sunlight in space is roughly five times stronger than at the ground, he said, and he predicted that compute in orbit could be cheaper than on Earth within two to three years.¹ The physics cuts both ways. A data center’s hardest problem is heat, and in a vacuum heat cannot be carried away by air or water; it can leave only as radiation, which demands large radiators. Most credible experts, Electrek reported, see little connection between the space-compute plan and near-term business reality.¹

Daniel Kahneman warned about the planning fallacy: our habit of forecasting from the best-case scenario of our own project rather than the track record of projects like it. The relevant track record here includes Tesla’s own. Electrek pointed out that the 4680 battery program, announced with similar fanfare in 2020, has reached only a small fraction of its original volume goal.¹ A bold forecast is not evidence. It is a claim waiting for evidence.

Sovereign industrial power

Whatever its odds, Terafab’s intent is clear, and it sits squarely on ground that has belonged to nations. Advanced chipmaking is the narrowest chokepoint in the modern economy, which is why Washington passed the CHIPS and Science Act in 2022 and has spent years subsidizing efforts to bring leading-edge fabs back to American soil.⁵ Governments treat the capacity to make chips the way they once treated steel or oil: as a foundation of national power.

Governments treat the capacity to make chips as a foundation of national power. Terafab would put a foundation of that size inside one founder’s group of companies.

Terafab would place that capacity inside a single group of companies, already closely bound together: SpaceX absorbed xAI earlier this year, and Tesla and SpaceX are partners in the fab.² If it works, the same group would make the chips, launch the satellites that run them, operate the network that connects them, and build the robots and cars that use them on the ground. In the Pierce Newsroom’s look at Starshield and DOGE, the question was who gets to say no to a company that controls critical infrastructure. Terafab raises the same question one layer further down, at the level of the silicon itself.

For now, Terafab is a site, a research fab, a set of filings and a very large promise. Walter Isaacson’s biography describes a man who sets impossible deadlines and sometimes, years late, meets them. The rest of the industry, and the governments that depend on it, would be wise to plan for both outcomes.