Musk's Terafab: Mining Sunlight for a Terawatt AI Hashrate, Not Just Tokens

Elon Musk has blueprinted Terafab, a hyper-scale chip factory with a single, degen-level ambition: to pump out a terawatt of AI compute annually. That's a cool 50x the entire planet's current AI chip output, a moonshot (literally) cooked up by a joint venture of Tesla, SpaceX, and xAI, now all orbiting under Musk’s personal aerospace umbrella.

The pitch is refreshingly simple: to level up civilization, you need to leave the planet, because Earth is basically a low-energy yield farm. “We want to be a civilization that expands to the galaxy with spaceships that anyone can go anywhere they want at any time,” Musk announced in a SpaceX broadcast, casually framing a terawatt of compute as merely the first grind toward a Kardashev-scale future.

Terafab would cram the entire chip-making pipeline—from mask creation to final testing—into one massive Texas facility. Musk claims this vertical integration could shave weeks off the usual, fragmented supply-chain feedback loops, turning semiconductor development from a slow rug-pull into something resembling fast finality.

Two chip families are on the roadmap, because why stop at one? The first is for edge inference, powering Tesla’s Optimus robots and its autonomous fleet, including the upcoming Cybercab. Musk envisions robot production eventually hitting 1-10 billion units per year, a number that makes today’s car market look like a testnet. The second family is hardened for space: radiation-tolerant, high-temperature silicon designed to trim the mass of thermal radiators on orbiting platforms, because in space, no one can hear your chips overheat.

Why launch compute into orbit? Simple math: U.S. electricity generation is a puny 0.5 TW, making a terawatt on Earth a physical impossibility. The plan, therefore, is solar-powered AI satellites. A prototype “mini-sat” would deliver 100 kW, with later versions scaling to megawatts. To hit the 1 TW goal, you’d need to launch roughly ten million tons of hardware to orbit yearly, assuming 100 kW per ton—a logistics puzzle that would break any legacy system.

Current Starship V3 can lift about 100t per launch; V4 aims for 200t. SpaceX has logged over 500 successful booster landings, driving launch costs from the Shuttle-era $65,000/kg down to a mere $1-2k/kg today. The target is $100-200/kg once Starship is fully optimized, a price point Musk believes will make space-based AI cheaper than Earth-bound alternatives within two to three years, flipping the entire economic model on its head.

Reusable launch systems like Starship are thus “a critical piece of the puzzle” for moving payloads of this scale. Musk also tossed out long-term ideas like lunar manufacturing and mass-driver catapults to further slash orbital deployment costs, because when you're playing on a galactic level, you might as well think with portals.

The compute gap is laughably stark: global AI capacity is about 20 GW per year, while all existing fabs combined represent only ~2% of Terafab’s one-terawatt ambition. Legacy foundries are still essential, but their expansion ceiling is a hard cap compared to Musk's sky-is-not-the-limit approach. “We either build the Terafab or we don’t have the chips,” he concluded, framing the space-first strategy not as a choice, but as the only viable bridge to a post-scarcity future.