Microsoft plans to more than triple its data center capacity, a buildout that would take the company’s global network past 38 gigawatts by 2032, up from about 12 gigawatts now, Bloomberg reported, citing people familiar with the plans. The scale is hard to overstate: 38 gigawatts would exceed the peak electricity demand of New York state. The company is racing to fix a shortage that has forced it to turn away some AI and cloud customers, a problem few chief executives would have predicted three years ago, and one that has reshaped how the entire industry thinks about electricity, land and water.
Why Microsoft Is Building
The shortage is real and quantified. Microsoft has told investors in recent quarters that compute constraints capped growth in its Azure cloud business, and OpenAI, its flagship AI partner, publicly complained about capacity shortfalls during their contract renegotiations earlier this year. Every gigawatt of AI training capacity is spoken for before it is built, and enterprise customers looking to deploy their own models face waiting lists measured in quarters, not weeks. The shortage has even shown up in Microsoft’s earnings guidance as forgone revenue, an admission few cloud providers made before the AI boom.
The plan implies adding roughly 26 gigawatts over about six years, an average of more than four gigawatts a year. For comparison, a single large nuclear reactor produces about one gigawatt. Microsoft has signed power agreements across multiple continents, restarted a reactor at Three Mile Island through a deal with Constellation Energy, and invested in geothermal and solar projects to feed the new sites. Grid interconnection queues, transformer shortages and turbine backlogs remain the binding constraints, not chips, and the company has started building its own substations to bypass utility delays where it can.
The Money Involved
Microsoft’s capital expenditures have ballooned along with the plans. The company spent more than $80 billion on data centers in its last fiscal year, and analysts expect well over $100 billion in the current one. Rivals are spending at similar rates: Amazon, Alphabet and Meta have each raised their capex guidance repeatedly, and Meta alone told investors its compute needs keep expanding beyond its own forecasts quarter after quarter. Across the four US hyperscalers, combined annual capex is now on the order of $400 billion, a sum that would have sounded absurd for the sector in 2022.
The spending has drawn scrutiny. Senator Bernie Sanders introduced a bill in March proposing a federal moratorium on new data center construction, citing energy and water consumption, and local fights over grid costs and water use have delayed projects in several US states. Arizona is the clearest example: TSMC’s Phoenix fabs draw 4.75 million gallons of water a day, and the state’s semiconductor boom keeps colliding with Colorado River allocations. NIST, meanwhile, opened public comment this summer on a draft security standard for AI data centers, with comments due September 25, a sign that Washington is engaging with the infrastructure rather than just the models.
Who Else Is Scaling
The arms race is global. China’s DeepSeek is planning a one-gigawatt AI campus in Ulanqab, Inner Mongolia, with at least a dozen other firms holding local approvals in the same city, part of a national push to domestic compute capacity. Anthropic has signed more than $135 billion in computing commitments this year alone, including a $35 billion deal with Nvidia-backed Lambda and a $100 billion, decade-long commitment to Amazon Web Services, with additional gigawatt deals at Google and Microsoft. Nvidia, which sells the accelerators all these facilities require, agreed this week to buy Hugging Face for $12.9 billion, betting that open model distribution belongs inside the company selling the compute to run it.
Power economics are shifting the design of the facilities themselves. Rack densities in AI halls now exceed 100 kilowatts, pushing operators toward 800-volt direct current distribution, and Flex’s $4.4 billion acquisition of EPC Power this month was framed explicitly around that architecture. The vendors building the electrical chain, from switchgear to solid-state transformers, are consolidating as fast as the data centers are multiplying, and lead times for grid transformers still stretch past two years in some markets.
The Risks in the Plan
Tripling capacity is a bet that AI demand keeps compounding. If model efficiency improves faster than usage grows, or if enterprise adoption disappoints, some of that capacity could sit underused, and the depreciation schedules on data center equipment are short. Microsoft’s own finance team has acknowledged the risk in investor calls, framing the buildout as committed spend against contracted demand rather than speculative capacity. The distinction matters for how the market reads the company’s earnings: capacity that serves signed contracts is different from capacity built on hope, and analysts now ask hyperscalers about contracted utilization the way they once asked telecoms about lit fiber.
There is also the question of where the power comes from. Utilities in Virginia, Texas and Ireland have warned that data center growth is outpacing generation additions, and some jurisdictions have paused new grid connections outright. Ireland’s grid operator has effectively frozen new connections in the Dublin area through the end of the decade. Microsoft’s plan assumes those bottlenecks clear, through on-site generation, better storage and faster permitting. For now, the company is building as fast as the electrical grid will let it, and customers are still being turned away. That is the clearest signal of demand anyone has, and it is why the buildout continues even as skeptics multiply.