Microsoft는 2032년까지 38 GW 데이터 센터 용량을 목표로 설정(현재 약 12 GW 대비). 고객 손실과 용량 위기 대응.

Microsoft has lost enterprise clients it could not serve, restricted Xbox cloud streaming for paying subscribers, and watched GitHub route developer traffic to Amazon because Azure had no room for it. The company's answer, revealed by Bloomberg on September 10, is a six-year plan to grow its global data center network from roughly 12 gigawatts today to more than 38 gigawatts by 2032. The expansion would more than triple Microsoft's current capacity, making it the most ambitious corporate infrastructure build in the company's history—and it is already running into the same political and physical constraints that created the shortage in the first place.

The 38-gigawatt target is not forward-looking ambition. It is a response to documented failure.

Capacity bottlenecks inside Azure forced Microsoft to turn away paying customers at scale across 2025 and into 2026, with consequences that reached every business unit. Temu signed contracts with competing cloud providers after Microsoft could not meet its compute needs. GitHub, which Microsoft owns, suffered nine outages in May 2026 alone and was forced to route its AI agent traffic to AWS because Azure had insufficient capacity. Microsoft's gaming division imposed restrictions on cloud streaming time for Xbox subscribers when hardware strain reached consumer products.

CEO Satya Nadella explained the problem directly in a 2025 podcast interview: "If you can't do that, you may actually have a bunch of chips sitting in inventory that I can't plug in. In fact, that is my problem today. It's not a supply issue of chips. It's the fact that I don't have warm shells to plug into." The constraint was not primarily about GPU supply—it was about the physical shells that house them. CFO Amy Hood told investors on a 2025 earnings call that capacity ran short of demand: "We had hoped to be in balance by the end of Q4 but we did see some increased demand. So we are going to be a little short, a little tight as we exit the year."

The practical effect has been that Microsoft's own internal products—Copilot, Azure AI, first-party research—consume available capacity before third-party cloud customers can access it. Hood confirmed this explicitly on the July 2026 earnings call, describing a deliberate trade-off that routes new GPUs and CPUs to first-party workloads first, with remaining capacity made available to Azure customers.

Of Microsoft's current 12-gigawatt footprint, only about 2 gigawatts is built around AI-specific chips—accelerators purpose-built for the matrix arithmetic that large language models require. The 38-gigawatt target calls for AI-specific silicon to grow to approximately one-third of the total, or roughly 12.7 gigawatts, by 2032. That implies a roughly six-fold increase in dedicated AI compute capacity over six years, against a backdrop of demand that has already exceeded the company's ability to deliver it.

The plan covers both company-owned campuses and long-term leases, but excludes computing power rented from neocloud providers such as CoreWeave. Microsoft's capital expenditures reached $145 billion in fiscal year 2026. On an adjusted basis—incorporating finance leases—calendar year spending reached approximately $175 billion. The company is guiding to roughly $50 billion in capital expenditure for the first fiscal quarter of 2027 alone.

The lease obligation picture reveals a much larger commitment. Microsoft reported more than $329 billion in data center leases that had been signed but had not yet commenced as of June 30, 2026—contracts scheduled to activate from fiscal 2027 through fiscal 2033. That figure is not current debt, not already-deployed capacity, and not included in Microsoft's published capex number. It is the funded pipeline for the build-out.

An accounting change announced alongside the FY2026 fourth-quarter earnings deepens the picture further. Starting in fiscal 2027, Microsoft extended the estimated useful life of its data centers and office buildings from 15 to 25 years. Under US accounting rules, a lease is classified as a finance lease—counted in reported capital expenditure—or an operating lease, not counted, partly based on whether the lease term constitutes a major portion of the asset's economic life. A longer economic life means more leases fall below the threshold for finance classification, reclassifying them as operating leases and mechanically reducing reported capex from roughly $190 billion to roughly $175 billion, without changing the underlying cash obligations. Morningstar senior equity analyst Dan Romanoff noted that the change would provide a "minimal boost" to margins, reflecting legitimate accounting methodology that the company fully disclosed. Investors and Azure customers should nonetheless note that the headline number represents the smaller figure.

Microsoft's capacity shortage existed in the first place because the same forces now threatening the expansion created it: grid interconnection queues that run four to seven years in the densest US markets, transformer and switchgear lead times of three to five years, and a shortage of skilled construction workers. A data center building can be completed in twelve to eighteen months. A grid connection to power it can take five to seven years. That structural mismatch—not chip supply—is the primary reason Microsoft has powered facilities sitting idle.

The 38-gigawatt plan requires adding 26 gigawatts over six years in an environment where those queues have not cleared and transformer manufacturers have not expanded capacity. The US interconnection queue held more than 2,600 gigawatts of pending projects across the country, with 474 gigawatts in ERCOT's queue alone.

Political opposition has now become a parallel constraint. Texas Governor Greg Abbott halted data center grid approvals in early August 2026, directing the state's power grid operator to audit all projects seeking connection to the ERCOT grid before any could proceed. The directive stopped up to 1,800 projects. New York Governor Kathy Hochul imposed a statewide one-year moratorium on new hyperscale data center approvals in July 2026. Pennsylvania Governor Josh Shapiro imposed new restrictions around the same time. Public opposition has crossed party lines: a June 2026 survey by Echelon Insights found that 62 percent of voters opposed a new data center in their community—dropping only to 58 percent after respondents heard arguments about economic and technological benefits. Among Republicans, 54 percent opposed; among Democrats, 69 percent; among independents, 53 percent.

Abbott, who called Texas "the epicenter of AI development" as recently as November 2025 when he joined Google to announce a $40 billion campus investment, framed his reversal as a failure of the industry's own making. "Gaining the support of people in local communities is essential," he told ABC News. "Data centers have dug their own grave."

Microsoft's existing Wisconsin campus at Fairwater provided a preview of community friction: when cooling fans came online in April 2026, residents of the neighboring village of Sturtevant reported a persistent hum audible indoors with windows closed. Microsoft adjusted fan speeds and added sound-reduction components, but the governance gap—a neighboring community with no formal voice over a project that directly affects it—remains unaddressed as the company pursues a build-out roughly thirty times Fairwater's size over six years.

To secure baseload power that AI workloads require around the clock, Microsoft has signed a 20-year contract with Constellation Energy to buy 100 percent of the output of the revived Three Mile Island nuclear facility in Pennsylvania, scheduled to restart in 2028 at 835 megawatts under the new name Crane Clean Energy Center.

**Note:** The source text appears incomplete; the final sentence cuts off mid-word.