Particle.news
Download on the App Store

Microsoft Fires Up Fairwater AI Superfactory, Linking Sites Into One Virtual Supercomputer

A continent-scale optical network binds hundreds of thousands of Nvidia Blackwell GPUs into a single system for frontier training.

Overview

  • Microsoft’s Atlanta Fairwater site is now live and interconnected with the Wisconsin facility over a dedicated low-latency AI WAN.
  • StartupHub.ai reports Microsoft added roughly 120,000 miles of new U.S. fiber to create the optical network that lets separate sites operate as one.
  • The platform integrates GB200 and GB300 GPUs on a single flat fabric with NVLink-connected NVL72 racks, pooled memory, and two-tier Ethernet at 800 Gbps using SONiC.
  • Fairwater maximizes density with two-story layouts, closed-loop liquid cooling designed for multi‑year service, and racks rated around 140 kW.
  • Microsoft adopts a grid-first power strategy at Atlanta, removing on-site generators and UPS in favor of high-availability utility power with software controls and on-site energy storage to smooth large AI workloads.