Microsoft is preparing one of the largest infrastructure buildouts in corporate history, with an internal plan to expand data-center capacity to more than 38 gigawatts by 2032 — more than triple the roughly 12 GW it operates today — Bloomberg reported on Thursday, citing people familiar with the plans.
Fairwater at the center, neoclouds at the edges
The 38 GW target dwarfs the peak electricity demand of New York State. Only about 2 GW of Microsoft's current base runs on AI-specific chips; by 2032 roughly a third of the fleet is expected to be AI-dedicated silicon, with the company already operating large Fairwater campuses packed with hundreds of thousands of GPUs. Reaching the target implies a sustained pace of just over 1 GW brought online each quarter for the rest of the decade.
To hit that number without owning every megawatt, Microsoft is leaning heavily on so-called neoclouds. Bloomberg names multi-billion-dollar leases and contracts with Nscale, CoreWeave, Lambda, Iren and Nebius, with capacity coming online across the United States, Europe and Nordics data-center hubs. Some leases are being stretched from 15 years to 25 years to smooth reported annual costs.
Capex heading for $175 billion
The build-out sits on a rapidly expanding capital budget. Microsoft's capex has climbed from $55.7 billion in 2024 to about $145 billion in 2026, and Bloomberg's sources put fiscal 2027 near $175 billion on an adjusted basis, with roughly $50 billion guided for the first quarter alone. In fiscal 2026 the company brought 88 sites online, including 31 in a single quarter.
The scale is also a response to demand Microsoft has been unable to serve. The company has turned away some enterprise AI workloads, restricted Xbox cloud streaming for paying subscribers and watched GitHub route developer traffic to Amazon after Azure ran out of capacity.
Winners and losers in the compute pipeline
For AI model builders, the plan is both a demand signal and a bottleneck warning. Microsoft is one of the main venues where new model companies buy training and inference. If the company is already declining work at 12 GW, the next six years of capacity will help decide which vendors get reliable clusters — and which are pushed to smaller regional clouds or delayed launches. Related coverage: Oracle's $638B RPO surge, Nvidia's Australian 2 GW push and Cerebras's 165 MW Nordic site.
Reporting based on coverage from Bloomberg, Data Center Dynamics and TechStartups.