Microsoft Launches Fairwater Atlanta: A Leap Forward in AI Infrastructure

Atlanta, Georgia – Today marks a pivotal moment in the evolution of artificial intelligence as Microsoft officially unveils its next-generation Fairwater datacenter in Atlanta. This strategically located facility joins its sister site in Wisconsin, alongside existing AI supercomputers and the expansive Azure global datacenter network, forming what Microsoft calls the world’s first planet-scale AI superfactory. The launch signifies a dramatic increase in computing power, designed to meet the surging demand for AI and unlock new possibilities for innovation.

The Architecture of Infinite Scale: Inside the Azure AI Superfactory

The demand for AI compute is escalating at an unprecedented rate, pushing the boundaries of traditional datacenter design. Microsoft’s Fairwater initiative represents a fundamental reimagining of how AI infrastructure is built and operated. Unlike conventional cloud datacenters, Fairwater utilizes a single, flat network capable of integrating hundreds of thousands of cutting-edge NVIDIA GB200 and GB300 GPUs into a massive, unified supercomputer. This architectural shift is the culmination of decades of experience and lessons learned from supporting some of the most demanding AI training workloads globally.

Maximizing Compute Density: Overcoming the Laws of Physics

Modern AI infrastructure faces a critical constraint: the speed of light. Minimizing latency between accelerators, compute units, and storage is paramount. Fairwater is engineered to maximize compute density, reducing these distances and boosting overall system performance. This is achieved through a combination of innovative cooling solutions and a unique datacenter design.

One of the most significant advancements is the implementation of a facility-wide, closed-loop liquid cooling system. This system reuses liquid continuously, requiring minimal replenishment – only when water chemistry dictates, designed for over six years of operation with an initial fill equivalent to the annual water consumption of 20 homes. This approach dramatically improves efficiency and sustainability. Liquid cooling’s superior heat transfer capabilities enable rack and row-level power densities of approximately 140kW per rack and 1,360 kW per row, allowing for unprecedented packing of compute resources.

Further enhancing density, Fairwater employs a two-story datacenter building design. This allows for three-dimensional rack placement, minimizing cable lengths and, consequently, latency, bandwidth limitations, and costs. Every GPU within Fairwater is directly connected to every other, making this optimized layout crucial.

Powering the Future: High-Availability and Cost-Efficient Energy

Delivering the necessary power to this immense compute infrastructure requires a novel approach. The Atlanta site was selected for its resilient utility power supply, capable of achieving 4×9 availability at a 3×9 cost. This reliability allows Microsoft to streamline power redundancy measures, reducing costs for customers and accelerating time-to-market.

Microsoft is also collaborating with industry partners to develop advanced power-management solutions. These include software-driven workload balancing, hardware-enforced GPU power thresholds, and on-site energy storage to mitigate power fluctuations and ensure grid stability as AI demand grows.

Cutting-Edge Technology: Accelerators and Networking

At the heart of Fairwater lies a purpose-built infrastructure powered by NVIDIA Blackwell GPUs. Each datacenter operates as a single, coherent cluster, leveraging an advanced network architecture that surpasses the limitations of traditional Clos networks. This scalability, supporting hundreds of thousands of GPUs on a single flat network, is achieved through innovations in scale-up, scale-out, and networking protocols.

Each rack houses up to 72 Blackwell GPUs, interconnected via NVLink for ultra-low-latency communication. These accelerators offer the highest compute density available, supporting low-precision number formats like FP4 to maximize FLOPS and memory efficiency. Each rack boasts 1.8 TB of GPU-to-GPU bandwidth and over 14 TB of pooled memory.

Scale-out networking is achieved through a two-tier, ethernet-based backend network supporting massive cluster sizes with 800 Gbps GPU-to-GPU connectivity. Utilizing a broad ethernet ecosystem and SONiC (Software for Open Network in the Cloud) minimizes vendor lock-in and reduces costs. Optimizations in packet trimming, packet spray, and high-frequency telemetry, alongside advanced congestion control and load balancing, ensure ultra-reliable, low-latency performance.

Extending the Reach: The AI WAN and Planet-Scale Compute

Even with these advancements, the demands of large AI training jobs—now measured in trillions of parameters—are exceeding the capacity of single facilities. To address this, Microsoft has built a dedicated AI WAN optical network, extending Fairwater’s scale-up and scale-out capabilities. This network, bolstered by over 120,000 new fiber miles across the US, enhances AI network reach and reliability.

The AI WAN connects different generations of supercomputers, creating an AI superfactory that surpasses the limitations of any single site. This empowers AI developers to leverage the broader Azure AI datacenter network, segmenting traffic based on workload requirements across scale-up and scale-out networks, and utilizing the continent-spanning AI WAN. What challenges do you foresee in managing such a geographically distributed AI infrastructure?

This represents a significant departure from previous architectures where all traffic relied on the scale-out network. It provides customers with fit-for-purpose networking at a granular level, maximizing infrastructure flexibility and utilization.

Frequently Asked Questions About the Fairwater AI Superfactory