Elon Musk presented the plan at a facility in Austin on November 7 and summarized it on X the next day: Tesla and SpaceX will build Terafab, which will technically be two fabs, each producing only one chip design.
One fab serves Tesla's vehicles and Optimus humanoid robots. The other serves SpaceX's space-based data centers.
Why build at all
His framing was blunt: "We either build the Terafab or we don't have the chips."
The argument is arithmetic on the demand side. He expects the two companies' compute needs to exceed total global chip output, and said existing suppliers, which he thanked by name (Samsung, TSMC, Micron), could not scale fast enough. Current global AI chip production, by his estimate, covers only a fraction of what his companies will need.
The target: one terawatt of compute capacity per year at full build-out. For scale, total US capacity today is about half a terawatt, and he expects global AI chip output next year to be roughly 35 to 45 gigawatts, which would make a single Terafab about 30 times that volume.
Space chips are a different engineering problem
The second fab's difficulty is not process node but environment. Musk described the requirements: resistance to high-energy ions and photons, tolerance for electron accumulation, and operation at higher temperatures.
"It's a hostile environment in space," he said. "You want to optimize it for space, and you also want to generally run it a little hotter than you would normally run a chip on Earth to minimize the radiator mass."
That last clause carries the real trade-off. There is no convection in vacuum, so heat leaves by radiation, and radiator area translates directly into launch mass. Running chips hotter buys savings on radiator mass at the cost of reliability, a compromise no terrestrial data center would make.
How hard is this, really
Musk gave no timeline, which matters. He has announced ambitious projects before, and some have slipped or vanished.
Three constraints dominate advanced fabs. Equipment: EUV lithography has one supplier and lead times measured in years. Yield: a new line typically takes one to two years to ramp. Talent: process engineers are scarce and cannot be hired into existence quickly.
The "one design per fab" approach does simplify things, and it fits Musk's habit of decomposing complex systems into repeatable units. But semiconductor manufacturing is not only design; it is materials, contamination control and equipment tuning, the unglamorous parts.
What it means for the supply chain
Near term, nothing changes. There is no timeline, no equipment order, no ramp plan, only stated intent.
Medium term, if it happens, the biggest impact lands on automotive and robotics silicon. Tesla currently depends on TSMC and Samsung; owning capacity would restore some pricing power and allow custom parts for Optimus, which needs chips in the millions.
Long term, this is the logical extension of Musk's vertical integration: rockets, satellites, cars, batteries, and now chips. Integration buys coordination and cost, but every link requires enormous capital, and a failure anywhere drags the whole structure.
The transferable lesson is not "build a fab." It is the engineering instinct behind it: narrow the complexity until the problem becomes repeatable.
Musk's Terafab: Two Fabs in Austin, One Terawatt of Compute a Year
← Previous: SpaceX Nears a $100B Run Rate, With Anthropic as Its Largest Customer
Next: →SpaceX Plans a $24B Gas Plant at Starbase to Power Its Own Compute →