Satya Nadella – How Microsoft Is Preparing for AGI (www.dwarkesh.com)

Microsoft gave an exclusive look at Fairwater 2, a next‑generation AI datacenter designed to aggregate enormous training capacity for future large models. The campus will host multiple Fairwater buildings—each packed with hundreds of thousands of GB200/GB300 accelerator cards, NVLink fabrics, "almost as much" network optics as all of Azure two years ago, and links across sites via a petabit AI‑WAN—yielding over 2 GW of power and more raw FLOPs than any existing AI datacenter. Microsoft says Fairwater represents roughly a 10× step up versus the compute used to train GPT‑5 and is architected for super‑pod scale training (model + data parallelism), high‑bandwidth data gen, and inference across regions. For the AI/ML community, Fairwater signals hyperscaler preparation for AGI‑scale workloads and the next phase of infrastructure-driven model progress: denser accelerators, extremely high‑bandwidth interconnects, and flexible site aggregation to avoid being locked into a single chip spec (Satya cited upcoming chips like Vera Rubin Ultra with different power/cooling needs). Technically, the build underscores how network and power (petabit links, millions of optical connections, GW‑class power) become first‑order constraints for scaling. Business implications are equally large: enormous capex, new consumption‑ and subscription‑based pricing meters, evolving margins around "token" economics, and organizational change needed to translate capability into broad economic productivity.