Satya Nadella — How Microsoft is preparing for AGI

The interview opens with an exclusive tour of Microsoft's Fairwater 2 data center. Nadella describes a network of Fairwater facilities containing hundreds of thousands of GB200 and GB300 accelerators and more than 2 GW of combined capacity. He says a single building already exceeds any existing AI data center. Microsoft's stated target is to multiply training capacity tenfold every 18 to 24 months, which would make Fairwater 2 roughly ten times the scale used to train GPT-5. The network's five million connections and optical equipment are presented as evidence that this is a campus-scale system, not simply a larger server room.

The facility is designed for several kinds of workloads rather than one permanent training job. Nadella describes aggregating compute across campuses and regions, linking Fairwater 2 to Fairwater 4, other Wisconsin sites, and a one-petabit network. Training is one use, alongside data generation and inference. This flexibility matters because model architecture and physical layout are coupled: new chips such as Vera Rubin Ultra can bring different power densities and cooling demands. Nadella therefore favors scaling over time and avoiding a single fixed specification that could strand capital when the hardware generation changes.

Nadella is enthusiastic about AI's potential but describes the industry as being in its early innings. He frames AI as a “cognitive amplifier” or “guardian angel,” emphasizing human utility rather than treating AGI as a mystical category. That framing does not deny that AI may perform tasks once exclusive to people; instead, he places AI among earlier technologies that extended human capability. He expects AI to increase the scale of organizations and work, while cautioning that technological availability does not automatically produce economic growth.

The distinction between capability and adoption becomes central when the discussion turns to growth. Nadella argues that the economic effects of the Industrial Revolution took decades to appear because technology had to diffuse far enough to change work products and workflows. He expects this transition to be compressed into perhaps 20 to 25 years, but says organizations still face change-management work. His claim is not that diffusion will be slow; it is that the speed of technical progress alone does not settle when productivity, revenue, or new workflows will materialize.