Microsoft is planning to more than triple the amount of computing power it controls, and the number it is reportedly chasing tells you more about the state of the AI business than any keynote could. The target is 38 gigawatts of data center capacity by 2032, up from roughly 12 gigawatts today.

To put that in terms that don’t require a utility engineer to translate: 38GW of continuous draw would be larger than the peak electricity demand of New York State. One company. One workload category. Six years.

The Numbers Behind the Plan

The figures come from people familiar with Microsoft’s internal planning, and the company has declined to discuss its buildout roadmap publicly. What is already on the record, though, is the spending curve that makes a target like this plausible.

  • 2024 capital expenditure: $55.7 billion
  • 2025: $115.9 billion
  • 2026: $145.3 billion
  • 2027 (projected): roughly $175 billion

That is a near-tripling of annual capex in three years, and the trajectory points higher. Microsoft brought 88 data centers online during fiscal 2026, including 31 in the final quarter alone, which works out to roughly a gigawatt of new capacity landing every three months.

Owned, Leased, and Borrowed

The 38GW figure is not all steel that Microsoft will pour itself. It blends self-built, self-operated campuses with leased capacity from a now-substantial cohort of specialist AI hosting firms. Microsoft has signed capacity agreements with CoreWeave, Nscale, Lambda, IREN and Nebius, and by late 2025 had committed an estimated $60 billion to those kinds of deals.

This is the part of the AI buildout that gets underplayed. The hyperscalers are no longer just competing with the neoclouds, they are underwriting them. A company like Nscale or Lambda can raise debt against a Microsoft lease in a way it never could against a pipeline of startup customers. Microsoft gets capacity without the full balance-sheet hit, and the neocloud gets a credit rating by proxy.

Why Microsoft Is Doing This Now

The uncomfortable driver here is not ambition. It is that Microsoft has been losing business because it ran out of room.

Capacity constraints have pushed AI workloads toward rivals for the better part of two years, and the company has said as much in softer language on earnings calls, acknowledging demand it could not serve. When your constraint is physical, you cannot engineer your way out of it in a quarter. Transformers have lead times. Grid interconnects have queues measured in years. Turbines and switchgear are backordered. A gigawatt of compute is a construction project, a power purchase agreement and a permitting fight stacked on top of each other.

Planning to 2032 is Microsoft admitting that the bottleneck has moved from chips to electricity and concrete, and that you have to start now to have anything in 2029.

A Notable Change in Posture

There is a small but telling detail buried in all this. As of March 2026, Microsoft stopped requiring non-disclosure agreements on its data center projects. For a company that spent a decade treating site selection as a state secret, that is a real shift, and it reflects how hard it has become to build quietly. Local opposition to data centers over water use, noise, tax abatements and grid impact has turned into a genuine obstacle in multiple U.S. states. You cannot run a 26-gigawatt expansion campaign while refusing to tell a county commission what you are building.

The Part Nobody Has Solved

Capacity targets are the easy half. Power is the hard half, and the arithmetic is unforgiving.

Thirty-eight gigawatts of data center load needs generation, transmission and, ideally, something close to round-the-clock availability. Microsoft has signed nuclear restart deals, fusion offtake agreements that most energy analysts treat as aspirational, and a long list of solar and wind contracts. None of that adds up to 26 new gigawatts on a 2032 schedule without significant new gas capacity, grid upgrades that utilities have not committed to, and a regulatory environment that moves faster than it historically does.

There is also the demand-side question that gets asked less politely every quarter. Microsoft is building for a world where inference demand keeps compounding. If model efficiency improves faster than usage grows, or if enterprise AI adoption plateaus below the projections baked into these plans, a lot of depreciating silicon sits in buildings that cost nine figures each.

The Competitive Context

Microsoft is not the outlier. Every major cloud provider has pulled its multi-year capacity plan forward, and the combined hyperscaler capex for 2026 is on track to exceed the annual GDP of a mid-sized country. Oracle just reported $28.5 billion of capex in a single quarter. Amazon, Google and Meta are all running at or near record infrastructure spend.

What separates Microsoft’s plan is that it is explicitly about tripling, not topping up. The company appears to have concluded that the cost of being short on capacity is higher than the cost of being long on it.

What This Means

Read the 38GW target as a bet with a specific shape. Microsoft is wagering that AI compute demand in 2032 will be large enough, and durable enough, that the constraint worth optimizing for is physics rather than finance. It is spending $145 billion this year and likely $175 billion next year on that conviction.

If the bet lands, Microsoft owns the substrate that a meaningful chunk of the software industry runs on, and the companies that hesitated spend the 2030s renting from it. If it misses, this becomes the most expensive overbuild in the history of enterprise technology, and the write-downs will be studied for decades.

For now the signal that matters most is not the gigawatt figure itself. It is that Microsoft, a company with functionally unlimited access to capital, has been losing deals because it could not plug anything else in. That is the detail worth sitting with.