SpaceX's Terafab: a $119B bet to vertically integrate AI silicon
A rocket company, an electric-car company, an AI lab and a chipmaker walk into a foundry. It sounds like the setup to a joke; it's actually one of the largest industrial bets ever proposed. The interesting question isn't whether Terafab gets built. It's why these people, why now, and what that tells the rest of us.
The bet, in numbers
Terafab is a planned semiconductor mega-fab jointly pursued by Tesla, xAI, SpaceX and Intel, announced by Elon Musk in March 2026. The stated goal is audacious: produce more than one terawatt, a trillion watts, of AI compute capacity per year. In May 2026, SpaceX put the initial investment at ~$55B and the all-phases total at up to $119B, sited in Grimes County, Texas, with a prototype fab near Tesla's Gigafactory in Austin (TechCrunch, CNBC).
The design philosophy is total vertical integration: chip design, fabrication and lithography, memory, advanced packaging and test, every stage under one roof, reportedly using Intel's 14A process at full scale, with Tesla leading the prototype and SpaceX the initial full-scale build (Wikipedia, TeslaNorth).
Why would a rocket company build a fab?
Read it against everything else happening in silicon. Advanced packaging is rationed for years. HBM is rationed. Leading-edge wafers are rationed. If your ambitions depend on a supply chain that someone else controls, and that someone has already reserved most of the capacity, then the only way to guarantee your own future is to own the bottleneck itself.
That's the SpaceX playbook applied to chips. They didn't accept the launch market's prices and queues; they built their own rockets. Terafab is the same instinct: when the scarce input throttles your destiny, stop renting it and start making it.
The constraint hiding inside "one terawatt"
Sit with that goal: a terawatt of compute capacity a year. Whatever else it means, it means an almost unimaginable appetite for electricity, both to manufacture the chips and, downstream, to run them. Every story in modern AI eventually collapses into the same sentence: the binding constraint is power. You can make the chips. Can you power what they become?
Even the most vertically integrated chipmaker in history doesn't escape that. The terawatt of silicon Terafab wants to ship still has to land in buildings, somewhere, that can feed and cool it. Owning the fab solves the supply of chips. It does nothing for the supply of cheap kilowatts.
You can't out-fab SpaceX, but you can apply the same lesson at the layer you can own. Liwa is vertical integration of the infrastructure tier: a secured long-term power agreement at $0.10/kWh, a liquid-cooled shell rated to 150 kW/rack, built and operated as one stack. Most operators will never make their own chips. But they can stop renting power and space from a market that's getting tighter, by locking it in now, at a founder rate, under their own brand.
Questions we're sitting with
- If the titans are integrating down into silicon, what's the equivalent integration move for an operator, owning power and facility?
- A terawatt of chips needs a continent of power. Does Terafab make the energy constraint better or far worse?
- When does "make it yourself" beat "rent it", and which layer of your stack is closest to that tipping point today?
Own the layer you can: power and space.
Lock liquid-cooled, 150 kW-ready capacity at $0.10/kWh on a 36-month founder rate, your hardware, your brand.
Sources
- TechCrunch, SpaceX may spend up to $119B on Terafab
- CNBC, Musk's SpaceX chip fab to cost up to $119B
- Wikipedia, Terafab
- TeslaNorth, SpaceX plans $119B chip factory
- Teslarati, Terafab announced
Terafab is an announced/planned project; figures are company statements and reporting as of May 2026 and may change as plans firm up.