Inside the world’s most powerful AI datacenter
This week we have introduced a wave of purpose-built datacenters and infrastructure investments we are making around the world to support the global adoption of cutting-edge AI workloads and cloud services. Today in Wisconsin we introduced Fairwater, our newest US AI datacenter, the largest and most sophisticated AI factory we’ve built yet. In addition to our Fairwater datacenter in Wisconsin, we also have multiple identical Fairwater datacenters under construction in other locations across the US. In Narvik, Norway, Microsoft announced plans with nScale and Aker JV to develop a new hyperscale AI datacenter. In Loughton, UK, we announced a partnership with nScale to build the UK’s largest supercomputer to support services in the UK. These AI datacenters are significant capital projects, representing tens of billions of dollars of investments and hundreds of thousands of cutting-edge AI chips, and will seamlessly connect with our global Microsoft Cloud of over 400 datacenters in 70 regions around the world. Through innovation that can enable us to link these AI datacenters in a distributed network, we multiply the efficiency and compute in an exponential way to further democratize access to AI services globally. So what is an AI datacenter? The AI datacenter: the new factory of the AI era Aerial view of Microsoft’s new AI datacenter campus in Mt Pleasant, Wisconsin. An AI datacenter is a unique, purpose-built facility designed specifically for AI training as well as running large-scale artificial intelligence models and applications. Microsoft’s AI datacenters power OpenAI, Microsoft AI, our Copilot capabilities and many more leading AI workloads. The new Fairwater AI datacenter in Wisconsin stands as a remarkable feat of engineering, covering 315 acres and housing three massive buildings with a combined 1.2 million square feet under roofs. Constructing this facility required 46.6 miles of deep foundation piles, 26.5 million pounds of structural steel, 120 miles of medium-voltage underground cable and 72.6 miles of mechanical piping. Unlike typical cloud datacenters, which are optimized to run many smaller, independent workloads such as hosting websites, email or business applications, this datacenter is built to work as one massive AI supercomputer using a single flat networking interconnecting hundreds of thousands of the latest NVIDIA GPUs. In fact, it will deliver 10X the performance of the world’s fastest supercomputer today, enabling AI training and inference workloads at a level never before seen. The role of our AI datacenters – powering frontier AI Effective AI models rely on thousands of computers working together, powered by GPUs, or specialized AI accelerators, to process massive concurrent mathematical computations. They’re interconnected with extremely fast networks so they can share results instantly, and all of this is supported by enormous storage systems that hold the data (like text, images or video) broken down into tokens, the small units of information the AI learns from. The goal is to keep these chips busy all the time, because if the data or the network can’t keep up, everything slows down. The AI training itself is a cycle: the AI processes tokens in sequence, makes predictions about the next one, checks them against the right answers and adjusts itself. This repeats trillions of times until the system gets better at whatever it’s being trained to do. Think of it like a professional football team’s practice. Each GPU is a player running a drill, the tokens are the plays being executed step by step, and the network is the coaching staff, shouting instructions and keeping everyone in sync. The team repeats plays over and over, correcting mistakes until they can execute them perfectly. By the end, the AI model, like the team, has mastered its strategy and is ready to perform under real game conditions. AI infrastructure at frontier scale Purpose-built infrastructure is critical to being able to power AI efficiently. To compute the token math at this trillion-parameter scale of leading AI models, the core of the AI datacenter is made up of dedicated AI accelerators (such as GPUs) mounted on server boards alongside CPUs, memory and storage. A single server hosts multiple GPU accelerators, connected for high-bandwidth communication. These servers are then installed into a rack, with top-of-rack (ToR) switches providing low-latency networking between them. Every rack in the datacenter is interconnected, creating a tightly coupled cluster. From the outside, this architecture looks like many independent servers, but at scale it functions as a single supercomputer where hundreds of thousands of accelerators can train a single model in parallel. This datacenter runs a single, massive cluster of interconnected NVIDIA GB200 servers and millions of compute cores and exabytes of storage, all engineered for the most demanding AI workloads. Azure was the first cloud provider to bring online the NVIDIA GB200 server, rack and full datacenter clusters. Each rack packs 72 NVIDIA Blackwell GPUs, tied together in a single NVLink domain that delivers 1.8 terabytes of GPU-to-GPU bandwidth and gives every GPU access to 14 terabytes of pooled memory. Rather than behaving like dozens of separate chips, the rack operates as a single, giant accelerator, capable of processing an astonishing 865,000 tokens per second, the highest throughput of any cloud platform available today. The Norway and UK AI datacenters will use similar clusters, and take advantage of NVIDIAs next AI chip design (GB300) which offers even more pooled memory per rack. The challenge in establishing supercomputing scale, particularly as AI training requirements continue to require breakthrough scales of computing, is getting the networking topology just right. To ensure low latency communication across multiple layers in a cloud environment, Microsoft needed to extend performance beyond a single rack. For the latest NVIDIA GB200 and GB300 deployments globally, at the rack level these GPUs communicate over NVLink and NVSwitch at terabytes per second, collapsing memory and bandwidth barriers. Then to connect across multiple racks into a pod, Azure uses both InfiniBand and Ethernet fabrics that deliver 800 Gbps, in a full fat tree non-blocking architecture to ensure that every GPU can talk to every other GPU at full line rate without congestion. And across
