There is a revealing detail buried inside Bloomberg’s report on Microsoft’s data-center ambitions. Before the company announced plans to triple its capacity, it was rationing what it already had. Capacity bottlenecks forced Microsoft to limit certain cloud subscriptions and turn away some high-profile clients due to insufficient infrastructure. A company that sells computing power was running out of computing power to sell. That is the context for everything that follows.
The plan calls for Microsoft’s network of facilities to surpass 38 gigawatts of total power capacity by 2032, up from approximately 12 gigawatts today. The roadmap incorporates both company-owned data centers and leased facilities. Neoclouds like CoreWeave, in other words, are a workaround, not the destination. Microsoft wants to own the infrastructure.
Only about 2 gigawatts of the current 12 is built around AI-specific silicon; that share is expected to reach roughly a third of the 38-gigawatt total. That means the majority of the buildout is general-purpose cloud compute, which matters because it signals this is not a bet on a single AI product cycle. It is a bet on the structural growth of cloud over the rest of the decade.
The financial commitment is not speculative. In its fiscal 2026 fourth-quarter earnings call, Microsoft said it expects capital expenditures to be over $50 billion for the first fiscal quarter of 2027 and that an accounting change brings its calendar-year 2026 expectation to approximately $175 billion. A 38-gigawatt target implies that pace will continue and probably accelerate through the rest of the decade. Microsoft is not building for a demand spike. It is building for a demand floor.
Where the Durable Businesses Sit
Readers focused on MSFT’s stock are looking one layer too shallow. The more interesting question is who builds 26 gigawatts of new capacity in six years, and what they need to do it. Power, cooling, and electrical infrastructure are the binding constraints, and the companies supplying them are booking orders that look less like a cycle and more like a decade-long contract.
GE Vernova is one clear beneficiary. In its second-quarter 2026 release, the company said Power orders rose 134% organically, and it reported that gas equipment backlog and slot reservation agreements grew from 100 to 116 gigawatts. The company is now taking reservations for 2031 deliveries, highlighting unusually long lead times for new generation equipment. That is not a backlog. That is a revenue schedule.
Eaton is converting the same demand into cash faster. In its second-quarter 2026 release, Eaton reported sales of about $8.5 billion, up about 21%, with adjusted EPS of $3.15. Management also highlighted data centers as a major growth driver across its electrical businesses. Eaton designs and manufactures heavy-duty electrical infrastructure, including transformers, switchgear, and uninterruptible power systems, essential for powering and protecting data centers. Every gigawatt Microsoft commissions runs through equipment like this.
Vertiv’s position is narrower but arguably more defensible. In its fourth-quarter 2025 release, Vertiv said organic orders grew 252% year over year, pushing backlog to $15.0 billion at a book-to-bill near 2.9x. AI data center infrastructure presents a specific switching-cost structure. Liquid cooling architecture decisions made for a hyperscaler’s buildout cycle are not reversible on a 12-month timeline: the rack layout, thermal management integration, and facility design are all structured around the chosen cooling technology. Winning the initial design decision locks in years of service revenue.
The Honest Risk
None of this comes without real uncertainty. Microsoft has said it is extending the useful life of data center and office buildings to 25 years from 15, lowering annual depreciation. Meanwhile, in March 2025, Bloomberg reported that TD Cowen analysts said Microsoft had walked away from new US and European data center projects totaling about 2 gigawatts. Plans of this scale get revised. At Vertiv specifically, the valuation risk is real: if the AI capital expenditure cycle hiccups, the multiple is what gets hit first.
Still, the direction of travel is not in doubt. The U.S. Department of Energy, citing Lawrence Berkeley National Laboratory, has said data centers used about 4.4% of U.S. electricity in 2023 and could rise to roughly 6.7% to 12% by 2028. Microsoft turning away customers and then committing to 38 gigawatts is a management team that has seen the demand firsthand and decided to build the asset that meets it. The electricity, turbines, and transformers required to execute that plan represent some of the most durable business visibility in the market today.
