Microsoft's reported 38GW target sets the pace for every AI startup that rents compute

Microsoft intends to grow its data centre capacity from roughly 12 gigawatts today to more than 38GW by 2032, according to a Bloomberg report published on Thursday, September 10, which cited people familiar with the plans. Of the current footprint, only about 2GW is built around AI-specific chips; by 2032 that share is expected to reach about a third of the total. Microsoft declined to comment, and the target is an internal projection rather than a public commitment. Bloomberg's sources cautioned that the numbers could change with customer demand.

Even as a projection, the figure matters well beyond Redmond. A company whose cloud business generated $59.3bn of revenue in a single quarter, and which buys chips, power and leased data centre space on a scale few rivals can match, is telling its suppliers how much it expects to need for the rest of the decade. For start-ups that build models, rent out GPUs or sell software that runs on Azure, the plan is in effect a forecast of the price and availability of their most important input.

What the numbers imply

The arithmetic is simple and demanding. Moving from 12GW to 38GW means adding about 26GW in roughly six years, or a little more than 4GW a year. Microsoft has recently been adding capacity at about 1GW a quarter, which puts the target within reach only if the current pace is sustained without interruption through 2032.

The AI share is the more striking part. A third of 38GW is close to 12.7GW, more than six times the roughly 2GW of AI-focused capacity reported today. In other words, the reported plan assumes that AI hardware will grow from a modest corner of Microsoft's estate into one of its largest components, while the conventional cloud business keeps expanding alongside it.

The figure also fits what management has already said in public. On its earnings call on July 29, chief executive Satya Nadella said Microsoft had added 31 new data centres across five continents in the June quarter and "another gigawatt of capacity", and that it remained "on track to roughly double our overall capacity in just two years." Chief financial officer Amy Hood told analysts that "customer demand continues to exceed available capacity."

The bill behind the plan

Capacity on this scale is paid for long before it earns anything. Microsoft reported capital expenditure of $41bn for the June quarter including finance leases, and Hood guided to more than $50bn for the first quarter of fiscal 2027. She said a change in lease accounting adjusts the expectation for calendar 2026 to approximately $175bn. In the fiscal year to June, the cash line for additions to property and equipment alone came to $115.9bn.

The company also changed the accounting that sits behind those figures. From the start of fiscal 2027, Microsoft is extending the estimated useful lives of its data centres and office buildings from 15 to 25 years. Longer lives spread depreciation over more years and soften the hit to reported earnings, a reasonable choice for concrete and power infrastructure, but one that also signals how long Microsoft expects these buildings to pay their way.

The revenue side is growing quickly, though not on the same scale. In the June quarter Microsoft reported revenue of $90.0bn, up 18%, Microsoft Cloud revenue of $59.3bn, up 27%, and growth of 43% in Azure and other cloud services. The question for shareholders is whether that growth can keep pace with a building programme sized for 2032.

Why the neoclouds are listening

Reports differ on one detail that matters a great deal to private markets. Some accounts say the 38GW figure includes capacity leased from third parties, including neoclouds such as CoreWeave, Nscale, Lambda, IREN and Nebius. Others describe the count as covering owned and leased facilities but excluding computing power rented from neoclouds. Either way, the specialist GPU clouds are exposed to Microsoft's choices. If the hyperscaler counts their capacity towards its target, it is promising them years of demand; if it does not, it is telling them that its own build-out is meant to stand on its own, with rented capacity as a supplement.

That distinction shapes the valuation case for a whole cohort of venture-backed and recently listed infrastructure companies. Many of them have raised debt against contracts with a handful of large customers, and Microsoft is often among the biggest. A credible long-term plan from Microsoft supports those contracts; a plan in which Microsoft builds most of what it needs itself would, over time, compress the premium that neoclouds earn for being available now.

Power, permits and the start-up layer

Gigawatts describe planned capacity, not a finished fleet, and the public reporting includes no region-by-region schedule or power procurement plan. That is where the risk lies. Some US states now require data centre projects to pay for upgrades to local power and water infrastructure, and some have paused new developments. Microsoft has reportedly stopped using non-disclosure agreements for its data centre projects, a sign that it expects community scrutiny to keep rising.

For start-ups further up the stack, the effects run in two directions. More capacity should eventually loosen the supply of GPUs and lower the cost of inference, the main operating expense for many AI application companies. In the near term, however, Microsoft's own statements suggest that demand will keep exceeding supply, which favours customers with long-term commitments and deep balance sheets over young companies buying capacity month to month.

Our view

The reported 38GW target is best read as a statement of intent about market structure rather than a construction schedule. Microsoft is signalling that it expects to own or control a growing share of the world's AI compute, and that it is willing to carry the cost on its balance sheet for decades. For founders, the practical lesson is that compute will remain a negotiated input, not a commodity, for some years yet: the companies that lock in capacity early, or design products that use less of it, will have an advantage that capital alone cannot buy. For investors in neoclouds, the key question is not whether demand exists but whose demand it is, and on what terms it can be withdrawn. Watch three things: whether Microsoft confirms the figure, how its lease commitments evolve in the next filings, and whether the pace of about 1GW a quarter accelerates. A plan this large leaves little room for a slowdown in AI spending, and the companies built around it should price that risk honestly.

Sources

• Baxtel: Microsoft plans 38GW of data center capacity by 2032
• Let's Data Science: Report: Microsoft targets more than 38 gigawatts of Azure data center capacity
• Remio: Microsoft data center expansion targets 38 GW but power is the real constraint
• Microsoft: FY26 Q4 press release
• Microsoft: Earnings call FY26 Q4

Filed under Big Tech. Written by InvestTimes Editorial under our editorial standards. Commentary and analysis for information only, not investment advice.