Microsoft's Fairwater 2 building beats every AI data center

Microsoft's Fairwater 2 site has passed 2 gigawatts, and one building there beats any AI data center now running, Satya Nadella told Dwarkesh Patel.

That 2-gigawatt figure is the real constraint behind the tight rate limits and slow inference that anyone building on frontier models runs into. The bottleneck is measured in gigawatts and construction schedules, not lines of code.

For builders, relief runs on a construction timeline. A better retry strategy doesn't fix a throttled API when inference capacity is tight. Only finished buildings like Fairwater 2 do, and that takes quarters.

Nadella didn't say when Fairwater 2 reaches full capacity, or how much of that 2 gigawatts is already serving live traffic versus still being commissioned. Patel's interview doesn't answer that either, and it matters if you're trying to plan around when the limits actually ease.

The next real relief for inference-constrained products shows up on a construction schedule, not in a model release. Watch for the next Fairwater-scale building coming online, since that's the number that resets the constraint.