TeraFab · Compute · Starship · 5 min read
One Unified Physical Machine
If you cannot land the booster, you cannot launch the server. If you cannot package the chip, you cannot power the robot.
The real bottleneck
The real bottleneck to civilizational-scale intelligence is not software. It is terrestrial power grids and thermal limits. Orbital compute is the only logical bypass. Yet the friction is extreme: to manufacture 100 GW of space-based compute each year, Starship must loft 1,000,000 tons of hardware annually. That cadence demands a flawless, rapid catcher-tower recovery loop.
The fab has to match the tower
To feed those satellites, the Grimes County TeraFab must reach 1 million wafer starts per month. That requires mastering disaggregated chiplets and advanced 3D packaging to escape the physical half-reticle limits of ASML’s High-NA EUV scanners. The output split is locked by physics: 80% space-grade, radiation-hardened D3 chips that keep ionizing radiation from bricking orbital servers. 20% AI5/AI6 chips that deliver the 50× efficiency gain allowing Optimus to run fully autonomous and off-grid.
Nothing here is a side project
If you cannot land the booster, you cannot launch the server. If you cannot package the chip, you cannot power the robot. It is one unified physical machine — catcher, wafer, radiator, biped. Specialist models are another 100× because the loss surface is smaller. A general model spends capacity on every language and every domain. A single-language, single-domain model spends it on the mechanism. Same joules, more of the thing you actually need. That is how orbital compute gets cheap. Not one giant mind. A factory of narrow ones.
Ask this paper
Grok answers from this essay only. Sign in so the call stays on your session — one question at a time.
Sign in to askNext
Mars’s Subsurface Sanctuary