Microsoft EVP of Cloud + AI Scott Guthrie says the goal is to be able to 10X the AI training every 18-24 months. This likely includes maintaining the 2-4X pace of chip improvement.
Satya Nadella gave Dylan Patel (founder of SemiAnalysis) and Dwarkesh an exclusive first-look at their brand-new Fairwater 2 datacenter.
Microsoft is building multiple Fairwaters, each of which has hundreds of thousands of GB200s & GB300s. Between all these interconnected buildings, they’ll have over 2 GW of total capacity. Even a single one of these Fairwater buildings is more powerful than any other AI datacenter that currently exists.
Microsoft is preparing for AGI across all layers of the stack.
Epoch AI view of Microsoft Fairwater – Construction status – Building 1 cooling and backup generators all look complete from above but there is still construction visible around the building. Building 2 is well underway with a foundation and a complete substation. Large carpark and concrete area north of Building 2. Building 3 has a concrete foundation and a substation underway. Microsoft Fairwater data centers have a two-story structure, with GPU racks networked vertically as well as horizontally. Each of these GPU buildings is paired with a one-story building dedicated to data storage and processing using CPUs. Epoch AI estimated compute capacity based on disclosures from Microsoft, and permit applications showing the layout of server racks and mechanical equipment.
Microsoft CEO Satya Nadella and EVP of Cloud + AI Scott Guthrie showcase the newly opened Fairwater 2 data center in Atlanta, Georgia—billed as the world’s most powerful AI facility.
Microsoft has brought its data center in Atlanta, Georgia, online. The facility is the company’s second ‘Fairwater’ site, its new data center design that primarily uses a closed-loop liquid cooling system. The data center is connected to the first Fairwater site in Wisconsin with an AI Wide Area Network using dedicated fiber optic cables.
AI companies are competing to build and power the largest AI datacenters. xAI built the 350 MW Colossus datacenter in Memphis, which they plan to expand to 1.5 GW. XAI is building colossus 2 which has about 1-2 gigawatts of power.
XAI Colossus 2 could have about 1.4 Gigawatts of power by Feb 2026.
EpochAI Estimate of XAI Colossus 2.
OpenAI’s 240 MW Abilene Stargate datacenter is planned to reach 1.2 GW. At full scale, these single datacenters will rival the most power-hungry facilities in the world today, such as the 1.2 GW Maaden or the 1.6 GW Bahrain aluminium smelters.
Earth based training clusters in single or multiple sites are projected to reach 10 GW by around 2030. This is larger than the capacity of the US’s largest power plant, the Grand Colulee Dam, and nearly matches the total installed power capacity for all NVIDIA AI chips at the end of 2024.
Five AI datacenters at a scale of 1GW or more to come online in 2026. The question will be the energy and facilities will be built but how much of the chips will be installed and running?
Fairwater is a 10x increase in training capacity compared to what powered GPT-5, with network optics rivaling the entirety of Azure’s infrastructure from 2.5 years prior and supporting around 5 million connections.
The design enables aggregation of FLOPs across regional sites for massive training jobs via model and data parallelism, while also supporting data generation and inference.
Fairwater 4 is under construction and is nearby. Fairwater 4 has a high-speed petabit networking linking it to the Atlanta site, and an IWAN (Intelligent Wide Area Network) connecting to multiple Fairwater facilities in the Milwaukee/Wisconsin area for superpod-scale training across a campus and beyond.
This is the Wisconson area facility.
The tour reveals empty server cells, the roar of cooling systems, and design choices using NVIDIA GB200 GPUs with flexibility for future chips like Rubin Ultra—emphasizing iterative scaling over static builds.
The facility is critical for a software company powered by hardware bets on AI’s future.
Fairwater 2 is in Atlanta. It is currently the most powerful in the world. Fairwater 4 is under construction nearby. Multiple other Fairwaters are being built in Milwaukee/Wisconsin and will be connected via IWAN.
They want 10x training capacity increase every 18-24 months. This is 10x the training from the OpenAI GPT-5 training setup. It is designed for pods and superpods with model parallelism and data parallelism across the campus and to Wisconsin sites.
The networking optics are equivalent to all of Azure 2.5 years ago. There are ~5 million network connections. High bandwidth connects between sites in a region and between two regions. There is a petabit network for Fairwater 4 linkage. IWAN for aggregating training jobs across Wisconsin data centers.
Workloads are primarily for large training jobs but versatile for data generation and inference. It is not limited to one use case.
They are using NVIDIA GB200 GPUs. They are flexible for future chips like Rubin Ultra with varying power densities and cooling needs. Cells (modular units) shown empty, with racks to be filled. The exact racks per cell is not undisclosed. There is a two-story layout implied for latency reduction.
This is a big bet on future scaling. They emphasize scaling in time with iterative builds. They do not want to scale once and get stuck. They is a coupling between model architecture and physical layout.
Microsoft’s Fairwater data centers represent a new generation of AI-optimized superfactories. They are purpose-built for training and deploying frontier AI models at unprecedented scale. Unlike traditional data centers, they feature dense GPU clusters in a single flat network, advanced liquid cooling, and planet-scale connectivity to form elastic, distributed supercomputers. The initiative stems from surging AI demand, particularly from partners like OpenAI, and aligns with Microsoft’s $100B+ annual capex on AI infrastructure.
Expansions and Future Builds
Fairwater 4 is under construction adjacent to Fairwater 2 in Atlanta. It shares the same petabit network for seamless linkage, enabling superpod aggregation across the two sites.
Exact size are GPU count are undisclosed. It is designed as an identical Fairwater twin for modular scaling. Expected completion 2026-2027 to support immediate workload spillover.
Wisconsin/Milwaukee Expansions will have $7.3B total investment. There is a $3B initial investment for Fairwater 1 and $4 billion announced Sep 2025 for a second identical Data Center nearby. This includes a recent $38M land purchase (135 acres) in Racine County (Nov 10, 2025) for additional hyperscale sites.
Those costs do not include the chips as each set of 1 million B200s or later chips would be about $50 billion.
There are multiple identical Fairwater builds underway, forming a regional cluster connected via IWAN for cross-campus training. Second DC to break ground soon and targets 2027 operation.
There are 5+ more Fairwater sites under construction nationwide. They are potentially in other Azure regions. This is oart of a planet-scale AI superfactory integrating with 400+ global Azure DCs via AI WAN for distributed training across 70 regions. The total Microsfot capex for AI infra is $34 billion+ in 2025 alone. There is an emphasis on supply chain resilience and multimodal AI support.
Designs prioritize flexibility for next-gen chips like the NVIDIA Rubin Ultra in 2026+.
AI Training Timelines and Usage
Fairwater sites are optimized for trillion-parameter frontier models, supporting pre-training, fine-tuning, RL, and synthetic data gen with over 90% GPU utilization. The transcript’s 10x-from-GPT-5 reference aligns with GPT-6. They are expected to leverage the full superfactory for distributed runs across Wisconsin-Atlanta (700+ miles apart) in near real-time.
GPT 6
OpenAI (Microsoft’s key partner) began initial cluster prep in mid-2025. The full training start is estimated at Q1 2026 on Fairwater infra. There should be a with release by mid-to-late 2026. Acceleration rumors stem from competitive pressure but no firm date confirmed. Sam Altman hinted at amazing new model in 2026 with multimodal/memory features. Wisconsin site online in 2026 would enable this. Atlanta already handling early workloads like Copilot/GPT-image-audio variants.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
