◆strategy / compute
Why we're building a terawatt-scale fab
AI compute demand has outgrown the fabs built for the last decade. Here's why Terafab is sized for the next one.
Every fab built in the last decade was sized for a demand curve that no longer exists. Frontier model training and inference now consume power and silicon at a scale that retrofitted, incrementally-expanded facilities were never designed to serve.
Terafab starts from a different premise: build one campus, sized for terawatt-class compute demand, from day one.
The gap we’re building for
Three things are true at once right now:
- Model training runs keep getting larger, not smaller.
- Inference volume for deployed models is compounding faster than training demand.
- Existing fab capacity is being expanded piecemeal, one module at a time.
That third point is the constraint. A campus designed for terawatt-scale output from the outset can be built with the power, water, and logistics infrastructure sized correctly the first time, rather than re-trenched for every subsequent expansion.
What “terawatt-scale” means in practice
It means power delivery, cooling, and cleanroom floor area sized for the compute demand we expect a decade out, not the demand of today. We’ll be sharing more on the specific engineering tradeoffs (power procurement, water reuse, floor plate design) in upcoming posts.
We’ll also be posting regular construction updates as the Grimes County site comes out of the ground. Follow along on the construction timeline.