Microsoft Plans to Triple Its Data Center Fleet to More Than 38 Gigawatts by 2032

Microsoft's global data center fleet draws roughly 12 gigawatts of power today. By 2032, the company intends for that number to exceed 38 gigawatts — an addition of some 26 gigawatts in six years, and a footprint that would pull more power at peak than the entire state of New York.

That roadmap, reported by Bloomberg on September 10 from internal projections described by people familiar with the planning, is the most concrete picture yet of how far the world's largest buyer of AI infrastructure intends to go. Reuters put the target at about 38 gigawatts, more than triple today's footprint. The figure covers capacity Microsoft owns and capacity it leases directly, but excludes compute rented from neoclouds such as CoreWeave and Nebius — meaning its effective reach in 2032 would be larger still.

Composition matters as much as the total. Only about 2 of Microsoft's current 12 gigawatts runs AI-specific silicon: Nvidia and AMD accelerators, plus Microsoft's own Maia parts. By 2032, roughly a third of the 38 gigawatts is earmarked for that hardware — on the order of 12 gigawatts, a sixfold increase. The balance stays conventional cloud: storage, databases, the Azure services that still produce most of the revenue.

Microsoft is already accelerating. It brought 1 gigawatt online in the December 2025 quarter alone, after adding 2 gigawatts across the whole of fiscal 2025. Capital expenditure climbed in step — $34.9 billion in the September quarter, a record $37.5 billion in December, roughly $145 billion across fiscal 2026 — with guidance of about $175 billion for calendar 2026 and $50 billion for fiscal Q1 2027. Two-thirds of the December spend went to short-lived assets: GPUs and CPUs, not concrete.

Chief executive Satya Nadella frames the buildout as an efficiency problem, not a land grab. On the December-quarter call he said Microsoft is optimizing for tokens per watt per dollar by “increasing utilization and decreasing TCO using silicon systems and software” — which let him claim a 50 percent throughput gain on OpenAI inference workloads for Copilot without adding a watt. On chip sourcing he was blunter: “We want a fleet at any given point in time to have access to the best TCO. And it's not a one-generation game. I think a lot of folks just talk about who's ahead.”

Twenty-six gigawatts meets a grid that is saying no

Microsoft does not control the variable that matters most. People familiar with the plan cautioned Bloomberg that the roadmap could shift, because server farms take years to build — and the last two months showed how fast the ground moves. In July, New York halted new data center approvals for up to a year while it writes development rules. In August, Texas Governor Greg Abbott ordered an audit of every project in the ERCOT interconnection queue, freezing roughly 49.8 gigawatts of proposals. The queue had swelled to about 474 gigawatts of requests, of which “approximately 90 percent of the new power requests are data centers” — more than five times ERCOT's peak demand record. Any project failing the audit, he wrote, “must be denied.”

Bloomberg NEF calculated that the Texas pause alone put about 20 percent of the US data center pipeline at risk of delay, with revenue exposure reaching $8 billion by the first quarter of 2027. “Delays in energization over this period could thus put billions of dollars of data center leasing revenue at risk,” the firm wrote. BNEF's chip-based model pegs US data center demand at up to 207 gigawatts by 2033 — the reason interconnection, not capital, is now the binding constraint.

Microsoft has tried to get ahead of the politics, pledging alongside its December results to “pay our way to ensure our data centers don't increase your electricity prices,” to replenish more water than it consumes, and to stop accepting local tax abatements after two decades of collecting them. Whether regulators and ratepayers believe it will determine how many of the 26 gigawatts ever get energized.

Does the demand justify the concrete

The financial case rests on backlog. Microsoft reported remaining performance obligations of $625 billion, average duration 2.5 years — a contracted pipeline larger than most peers carry. The concentration risk is specific: roughly 45 percent of that RPO is OpenAI alone, after the restructuring the two companies negotiated in October 2025. CFO Amy Hood pushes back on reading it that way, telling analysts the remaining 55 percent, about $350 billion, is “related to the breadth of our portfolio, a breadth of customers, across solutions, across Azure, across industries, across geographies,” adding: “frankly, I think we have super high confidence in it.”

Wall Street is less settled. On the same call, a Morgan Stanley analyst noted the stock fell despite strong results, and diagnosed why: “capex is growing faster than we expected, and maybe Azure is growing a little bit slower than we expected.” Azure grew 39 percent that quarter, down from 40 percent. Hood's answer — that Microsoft is making long-term decisions across every layer of the stack — is true, and also the answer that cannot be falsified for several years.

Microsoft is not alone: Amazon and Meta have each signaled more than 10 gigawatts by the end of 2026, and the four largest hyperscalers had executed more than 133 gigawatts of energy supply agreements by the first quarter of this year. Microsoft's distinction is the horizon — 2032 is far enough out that almost nothing about it is binding.

Three things to watch. Microsoft's fiscal first-quarter report will show whether the $50 billion capex guide holds and whether the AI share of the fleet moves off 2 gigawatts. New York's siting rules and ERCOT's Batch Zero study will reveal how much of the American pipeline survives 2027 policymaking. And the ramp of Maia 200, the 3nm inference chip announced in January, will show whether Microsoft can bend its own cost curve — because at 38 gigawatts, a few percentage points of tokens per watt is worth more than another data center.