Microsoft is revealing the architecture powering its next-generation AI capabilities, centered around what it calls the Azure AI superfactory. A key component is the unveiling of a new Fairwater site in Atlanta, Georgia, expanding on the initial Fairwater deployment in Wisconsin. Planet-Scale AI Superfactory This purpose-built datacenter connects to previous AI supercomputers and Azure's global infrastructure. The Fairwater sites use dense computing to efficiently meet AI compute demands, driving advancements in model intelligence. Microsoft has reinvented the design of AI datacenters and the systems running within them to make this possible. Fairwater Design and Architecture Fairwater departs from the traditional cloud datacenter model, using a flat network to integrate hundreds of thousands of NVIDIA GB200 and GB300 GPUs into a supercomputer. The design accommodates different AI workloads such as pre-training, fine-tuning, and reinforcement learning. A dedicated AI WAN backbone dynamically allocates diverse AI workloads. Key Technical Innovations:

Maximum Compute Density: Fairwater maximizes compute density to minimize latency.
Advanced Cooling: AI servers are connected to a closed-loop, facility-wide liquid cooling system designed for longevity and sustainability.
Two-Story Design: A two-story datacenter minimizes cable lengths and improves latency, bandwidth, reliability and cost.
High-Availability Power: The Atlanta site achieves 4x9 availability at 3x9 cost.
Power Management: Microsoft has developed power-management solutions to mitigate power oscillations.
Accelerators and Networking: Each Fairwater datacenter runs a cluster of NVIDIA Blackwell GPUs, scaling beyond traditional limits.

Networking Systems Each rack houses up to 72 NVIDIA Blackwell GPUs connected via NVLink. The ethernet-based backend network supports large cluster sizes with 800 Gbps GPU-to-GPU connectivity and avoids vendor lock-in using SONiC. Planet Scale Reach Microsoft’s dedicated AI WAN extends the scale of the Fairwater networks. The high-performance backbone directly connects different generations of supercomputers into an AI superfactory, offering flexible networking options based on workload needs. The new Fairwater site represents the next leap in Azure AI infrastructure, combining breakthrough innovations in computing, sustainability, and networking. This architecture provides a flexible infrastructure for modern AI workloads.