TMTB: Dwarkesh Podcast with MSFT CEO Satya Nadella Key Quotes

Nadella is always great to listen to in a setting like this so recommend the full listen here…

---

Scale, topology & the Fairwater architecture

“We’ve tried to 10× the training capacity every 18–24 months. This [Fairwater 2] would effectively be a 10× increase from what GPT-5 was trained with… the number of optics in this building is almost as much as all of Azure across all our data centers two and a half years ago… We’ll aggregate flops for large training jobs across sites. You’ll use it for training, then data-gen, then inference—it’s not one workload forever… Fairwater 4 will also be on that one-petabit network so we can link them at a very high rate; the AI WAN connects to Milwaukee where multiple Fairwaters are being built. You can literally see the model parallelism and data parallelism this campus is built for.”

Build pacing, fungibility & not getting stuck on one generation