Inside the world’s most powerful AI datacenter
This week, we’ve unveiled a series of specially designed datacentres and infrastructure initiatives worldwide, all aimed at facilitating the global embrace of state-of-the-art AI workloads and cloud services.
Today in Wisconsin, we launched Fairwater, our latest US AI datacentre, which is the largest and most advanced AI factory we’ve created to date. Alongside Fairwater, there are several similar datacentres underway across the US.
In Narvik, Norway, Microsoft revealed a collaboration with nScale and Aker JV to establish a new hyperscale AI datacentre.
Additionally, in Loughton, UK, we announced a partnership with nScale to create the UK’s largest supercomputer to bolster services within the region.
These AI datacentres are significant capital investments, amounting to tens of billions of dollars and incorporating hundreds of thousands of cutting-edge AI chips. They will smoothly integrate with our global Microsoft Cloud, consisting of over 400 datacentres in 70 locations worldwide. By innovating ways to connect these AI datacentres into a distributed network, we significantly enhance efficiency and computing power, making AI services more accessible worldwide.
So, what exactly is an AI datacentre?
The AI Datacentre: The Factory of the AI Era
An AI datacentre is a specially constructed facility tailored for AI training and for operating large-scale artificial intelligence models and applications. Microsoft’s AI datacentres power products like OpenAI, Microsoft AI, and Copilot, among various other leading AI workloads.
The Fairwater AI datacentre in Wisconsin represents an extraordinary achievement in engineering, sprawling over 315 acres with three enormous buildings that collectively offer 1.2 million square feet of space. Building this facility involved the installation of 46.6 miles of deep foundation piles, 26.5 million pounds of structural steel, 120 miles of medium-voltage underground cables, and 72.6 miles of mechanical piping.
Unlike regular cloud datacentres, which are designed to manage numerous smaller tasks like hosting websites and emails, this datacentre is intended to function like a gigantic AI supercomputer by interconnecting hundreds of thousands of advanced NVIDIA GPUs through a unified network. It actually delivers ten times the performance of today’s fastest supercomputer, enabling unprecedented levels of AI training and inference tasks.
The Role of Our AI Datacentres – Powering Advanced AI
Effective AI models depend on thousands of computers collaborating, driven by GPUs or specialised AI accelerators that handle extensive simultaneous computations. They’re connected through lightning-fast networks that allow instant sharing of results, all supported by vast storage systems holding data—such as text, images, or video—broken down into manageable tokens, which are the fundamental units of information that AI learns from. The aim is to keep these chips constantly engaged; if the data or network lags, everything grinds to a halt.
The AI training process is a cycle: the AI processes tokens sequentially, predicts the next one, verifies against the correct answers, and fine-tunes itself. This loop continues trillions of times until the system excels at its specific task. You can visualise it like a professional football team’s practice. Each GPU represents a player executing drills, the tokens are the plays being practiced, and the network acts as the coaching staff, coordinating activities and ensuring everyone stays in sync. The team repeats plays until they nail them perfectly. In the end, the AI model, like the team, masters its strategy, ready to perform in actual conditions.
AI Infrastructure at Scale
Purpose-built infrastructure is vital for powering AI effectively. To calculate the token algorithms at the trillion-parameter scale of leading AI models, the core of the datacentre comprises dedicated AI accelerators (like GPUs) alongside CPUs, memory, and storage on server boards. Each server hosts several GPU accelerators that are interconnected for high-bandwidth communication. All these servers are assembled into racks, with top-of-rack (ToR) switches ensuring low-latency networking between them. So, while from the outside it looks like many independent servers, in reality, it operates as one massive supercomputer where hundreds of thousands of accelerators can train a single model simultaneously.
This datacentre operates a vast interconnected cluster of NVIDIA GB200 servers and millions of compute cores with exabytes of storage, all designed for the most demanding AI workloads. Azure was the first cloud provider to roll out the NVIDIA GB200 server, along with rack and complete datacentre clusters. Each rack houses 72 NVIDIA Blackwell GPUs, interconnected in one NVLink domain that provides 1.8 terabytes of GPU-to-GPU bandwidth, giving every GPU access to 14 terabytes of shared memory. Instead of functioning as multiple separate chips, each rack behaves like a singular, colossal accelerator, capable of processing an astonishing 865,000 tokens per second—the highest throughput of any cloud platform today. The datacentres in Norway and the UK will also employ similar clusters and will benefit from NVIDIA’s upcoming AI chip design (GB300), which further increases pooled memory per rack.
One of the challenges of achieving supercomputing scale, particularly as AI training demands continue to escalate, is perfecting the networking architecture. To ensure swift communication across various layers in a cloud setting, Microsoft extended performance beyond single racks. For the latest global deployments of NVIDIA GB200 and GB300, at the rack level, these GPUs communicate via NVLink and NVSwitch at terabytes per second, breaking down memory and bandwidth barriers. To connect across multiple racks into a pod, Azure employs both InfiniBand and Ethernet fabrics that allow 800 Gbps in a full fat tree non-blocking architectur, enabling each GPU to interact with every other GPU at full speed, without any delays. Additionally, within the datacentre, numerous pods of racks are linked to minimise hop counts, allowing tens of thousands of GPUs to operate as one large-scale supercomputer.
When arranged in a traditional datacentre layout, the physical distance between racks can introduce latency. To tackle this, the racks in the Wisconsin AI datacentre are set up in a two-storey configuration that connects racks not only side by side but also above and below.
This stacked approach distinguishes Azure’s offering. Microsoft Azure was not only the first cloud provider to launch GB200 at both rack and datacentre scales; we’re achieving this on a massive scale with customers today. By co-engineering the entire system with top-tier industry partners alongside our tailored solutions, Microsoft has created the world’s most powerful AI supercomputer, specifically designed for frontier models.
Addressing Environmental Impact: Closed Loop Liquid Cooling
Traditional air cooling systems struggle to manage the high density of modern AI hardware. Our datacentres employ advanced liquid cooling technologies—integrated pipes circulate chilled liquid directly into servers for efficient heat removal. This closed-loop system ensures no water is wasted; it only needs to be filled once and is then recycled continuously.
By designing datacentres specifically for AI, we have been able to incorporate liquid cooling systems directly into the facilities, allowing for greater rack density. Fairwater is supported by the second largest water-cooled chiller plant globally and will constantly circulate water within its closed-loop cooling setup. The warm water is routed to the cooling “fins” at each side of the datacentre, where 172 20-foot fans chill and recirculate the water back to the datacentre, thus maintaining optimal efficiency, even during peak loads.
More than 90% of our datacentre capacity employs this system, which only requires water once during construction, continually reusing it with no evaporation losses. The remaining 10% of traditional servers utilize outside air for cooling, switching to water only on the hottest days, significantly reducing water consumption compared to conventional datacentres.
We’re also leveraging liquid cooling to facilitate AI workloads in many existing datacentres; this process is supported by Heat Exchanger Units (HXUs) that also function with zero operational water use.
Storage and Compute: Designed for AI Efficiency
Modern datacentres contain exabytes of storage and millions of CPU cores. For the AI infrastructure cluster to function, an entirely separate datacentre infrastructure is required to manage the data used and produced by the AI cluster. To illustrate the scale—Wisconsin’s AI datacentre features storage systems that stretch the length of five football fields!
We’ve reengineered Azure storage to accommodate the most demanding AI workloads across these extensive datacentre setups, achieving true supercomputing scale. Each Azure Blob Storage account can handle over 2 million read/write transactions per second, and with millions of available accounts, we can elastically scale to meet virtually any data need.
This capability is underpinned by a fundamentally redesigned storage foundation that aggregates capacity and bandwidth across thousands of storage nodes and hundreds of thousands of drives. This allows for exabyte-scale storage, eliminating the necessity for manual sharding and simplifying operations for even the largest AI and analytics workloads.
Innovations like BlobFuse2 ensure high-throughput and low-latency access for GPU node-local training, guaranteeing that computing resources remain engaged, and that extensive AI training datasets are readily available when required. The multiprotocol compatibility allows seamless integration with diverse data pipelines, while deep integration with analytics engines and AI tools speeds up data preparation and deployment.
Automatic scaling dynamically allocates resources as demand increases, coupled with advanced security, resiliency, and cost-effective tiered storage. Azure’s storage platform is leading the way for next-generation workloads, offering the performance, scalability, and reliability necessary.
AI WAN: Connecting Multiple Datacentres into a Larger AI Supercomputer
These new AI datacentres are components of a worldwide network of Azure AI datacentres, interconnected through our Wide Area Network (WAN). This approach transcends a single structure, representing a distributed, flexible, and scalable system that operates as one powerful AI entity. Our AI WAN is engineered to enable large-scale distributed training across various, geographically dispersed Azure regions, allowing customers to access the power of a massive AI supercomputer.
This marks a pivotal shift in how we conceptualise AI supercomputers. Instead of being confined to a single location, we’re creating a distributed framework where computing, storage, and networking resources are pooled and orchestrated seamlessly across different datacentre regions. This leads to enhanced resilience, scalability, and adaptability for customers.
Bringing Everything Together
To tackle the pressing demands of expansive AI challenges, we had to overhaul every layer of our cloud infrastructure stack. This isn’t just about isolated innovations, but integrating numerous new methodologies across silicon, servers, networks, and datacentres, resulting in progress where software and hardware are aligned as a cohesive, purpose-built system.
Microsoft’s Wisconsin datacentre will play a pivotal role in the future of AI, driven by genuine technology, solid investment, and real community impact. As we connect this facility with other regional datacentres and harmonise each layer of our infrastructure into a unified system, we are paving the way for a new era of cloud-based intelligence, ready for the challenges ahead.
To explore more about Microsoft’s datacentre innovations, take a virtual tour at datacenters.microsoft.com.
Scott Guthrie oversees hyperscale cloud computing solutions and services, including Azure, Microsoft’s cloud platform, generative AI solutions, data platforms, and cybersecurity. These services assist organisations globally in addressing critical challenges while fostering long-term transformation.
