Bloomberg reported on September 10, citing people familiar with the plans, that Microsoft intends to expand its global data-center capacity from roughly 12 gigawatts today to more than 38 gigawatts by 2032. The figure, more than three times the current footprint, is the clearest public measure yet of how much power the company expects the AI buildout to consume.
The plan also resets the mix. About 2 gigawatts of Microsoft’s existing capacity is dedicated to AI chips, and the target is to lift AI’s share of the 38 gigawatts to about a third. That shift, from a sliver to roughly 12 or 13 gigawatts of AI-dedicated compute, is where the money and the risk will concentrate.
The expansion is a response to a shortage Microsoft has rarely admitted in public. The company has said it turned away some cloud and AI contracts for lack of compute, restricted the paid tier of its Xbox cloud streaming service, and its GitHub unit has routed developer traffic to Amazon. For a company that sells computing as a service, running out of it is a striking confession.
The spending behind the plan keeps rising. Microsoft spent $145 billion on capital expenditures last fiscal year, expects roughly $175 billion in calendar 2026, and guided to about $50 billion for the first quarter of fiscal 2027. As of June 30, it had signed data-center leases worth more than $329 billion that had not yet been activated.
The scale translates into a number that is hard to ignore. At an industry benchmark of about $50 billion per gigawatt of data-center capacity, an increment of 26 gigawatts implies cumulative spending above $1 trillion. That is an order of magnitude beyond anything Microsoft has spent before, and it presumes AI demand keeps growing fast enough to justify building years ahead of need.
Microsoft’s own leadership has described the constraint in plain terms. The shortage is no longer chips, the company has told investors, but the physical shells, power, and cooling to plug them into. Chips can be ordered in volume; land, grid connections, and electricity cannot be bought at the same speed, and they are now the part of the buildout Microsoft is racing to secure.
Power is the binding resource. Data centers of this scale strain regional grids, and Microsoft has already signed long-term power deals to feed them, including an agreement with Constellation Energy to restart a reactor at Three Mile Island in Pennsylvania. The 38-gigawatt plan extends that logic: securing generation capacity is now as important to the cloud business as securing customers.
The plan is inseparable from the OpenAI relationship. Microsoft is the primary cloud provider behind OpenAI’s models, and the demand it is projecting is a bet that ChatGPT and the rest of the AI software stack will keep consuming compute at an accelerating rate.
Azure is the second-largest cloud after Amazon Web Services, and its growth has been capped by the same capacity ceiling Microsoft is now spending to lift. The company has framed the shortage as the reason some enterprise deals slipped, a concession that capacity, not demand, is the current limit on revenue.
Building 26 gigawatts in six years means clearing the permitting and grid-interconnection process in dozens of markets at once. Those timelines sit outside Microsoft’s control, and analysts said execution risk on power delivery is the real uncertainty in the plan, more than any question about whether customers want the compute.
The move also aligns Microsoft with the rest of the hyperscaler group. Google, Amazon, and Meta have each disclosed rising capital plans, and all four are bidding for the same scarce inputs: power contracts, transformers, and the labor to build and run the facilities. The 38-gigawatt figure is Microsoft’s bid to stay at the front of that queue.
Microsoft’s data-center ambitions date to the launch of Azure in 2010, but the AI era has changed the economics. Where early cloud capacity served thousands of customers with general workloads, the new facilities are built around dense AI clusters, and a single AI deployment can consume the power that once served an entire enterprise region.
Analysts said the number matters because it turns a vague AI thesis into a measurable commitment. A company that promises to triple capacity is telling investors, suppliers, and utilities to plan around a specific scale, and the commitments become hard to unwind once the leases and power agreements are signed.
The risk cuts both ways. If AI demand grows as Microsoft expects, the capacity will fill and the spending will look disciplined. If demand cools, the company will be left paying for power and buildings it cannot use, and the $329 billion in unactivated leases would become a liability rather than a sign of conviction.
The six-year horizon is the honest part. Microsoft is not promising the capacity next quarter; it is telling the market that the AI buildout is a decade-scale capital program rather than a cyclical one. The open question is whether the demand curve will agree with the plan long enough for a trillion-dollar bet to pay off.
