Microsoft's 38GW target is a lease-and-grid forecast

Microsoft wants 38GW of data center capacity by 2032, up from roughly 12GW in operation now, according to a Bloomberg report carried by Data Center Dynamics, and that is a plan to more than triple a footprint that already anchors much of the world's AI compute, resting on people the report describes as familiar with it. Microsoft declined to comment on its building plans to Bloomberg.

The headline gigawatts are the least useful part of the account. The pace of deployment the target implies and the split between capacity Microsoft builds itself and capacity it leases from providers like the neoclouds matter more, and the report supplies the pace but not the split, leaving the ratio that decides whether 2032 is a capex story or a credit story.

On the pace, the numbers are deflationary. Microsoft has brought about 1GW of capacity online in recent quarters and stood up 88 data centers during its fiscal 2026, 31 of those in the fourth quarter alone, per the report; the 26GW gap between today's footprint and the 2032 target works out, on the report's own figure, to an average quarterly deployment of just over 1GW. That is a slight ramp from the pace the company is already running, not a step change in what its construction organization can deliver. Among the halls already live are two of its Fairwater sites, the AI-dedicated buildings that house hundreds of thousands of GPUs.

Capex tells a steeper story. Total spending ran $55.7bn in 2024, then $115.9bn, and $145.3bn in 2026, with 2027 estimated at $175bn, while the deployment curve bends gently and the spending curve climbs by tens of billions a year—a gap that is what powered shells, GPUs, and the switchgear installed ahead of the compute now cost, and some of that money is buying capacity rather than building all of it. The report names the neoclouds Microsoft has signed large contracts with—CoreWeave, Nscale, Lambda, Iren, and Nebius—and cites an estimated $60 billion of spend as of November 2025; it also notes Microsoft stopped signing NDAs on its projects in March, which leaves the pipeline more visible than it was.

Twenty-six gigawatts of new load is a power portfolio before it is a real estate program. The sector's fights this year have run through interconnection positions, load classifications, and the conditions states attach to permits, and a five-year target of this size is a standing bid the queue will have to absorb.

Building or leasing, the capacity has to land somewhere with power, and the report is direct that this is where the friction sits. Public pushback and scrutiny are climbing alongside the hyperscale buildout: US states are increasingly requiring data center projects to pay their own way and cover the cost of upgrading electrical and water infrastructure, and some are enacting pauses on new projects. Grid access is now the asset, with generation and compute as derivatives of connection consent. TeraWulf's decision to carry the full grid costs of a 482MW Kentucky connection is the clearest illustration of the price, since the local co-ops become financial partners and the interconnection moves. Consumer anger over water arrives through a different door, and our reporting on drink brands turning water anxiety into a permitting risk suggests local politics now function as an input to the schedule rather than a nuisance after it.

The leasing side is where the 2032 figure does its heaviest work. Nscale is taking a $51bn contracted backlog to a US listing while it funds a 10GW buildout, and Lambda's $1B private placement is underwritten by Microsoft's lease payments rather than Lambda's merchant demand. Both financings assume a hyperscaler demand curve that keeps rising for the better part of a decade, and this report is the most explicit statement of that curve anyone has put on the record. That is a comfortable arrangement for the neoclouds while the leases are long and the tenant is investment grade, but it also leaves a meaningful share of the AI build's residual risk on balance sheets whose credit is, in the end, one customer's signature.

The market has been willing to let one named tenant price the entire digital stack—the real estate, the debt secured against the GPUs, and increasingly the equity of the firms that host them—and nothing in this report argues for repricing that. What would is a leased share that stops working as a bridge and becomes the plan. The report does not say how the 26GW divides, and that omission is worth pressing on, because a Microsoft that builds 38GW and one that leases a large share of it are different counterparties for the utilities, the states, and the lenders already extending credit against the leased version.

Neither is the 2032 target: the quarterly deployment rate will show whether the ramp is real, and state cost-recovery rules will show whether the power is there to deploy against. The 38GW figure will do its work in other people's fundraising long before either answer arrives.