Microsoft plans to more than triple data-center capacity to 38GW by 2032, after turning away customers
A capacity crunch reportedly forced Microsoft to decline cloud and AI clients and contributed to an 8-hour GitHub outage tied to server constraints. Meanwhile Microsoft is shifting lease terms to manage how the buildout hits its own books.
Microsoft plans to grow its global compute footprint from roughly 12 gigawatts today to more than 38 gigawatts by 2032, according to Bloomberg reporting carried on the Reuters wire on September 10, 2026, citing people familiar with the matter (Microsoft declined to comment). AI-dedicated capacity grows from about 2 gigawatts now to roughly a third of the 38-gigawatt total. Reported drivers include capacity shortages that forced Microsoft to turn away cloud and AI clients (one retailer reportedly went to Oracle) and an August outage tied to server constraints that took GitHub down for eight hours.
The lease terms change how to read the number
Microsoft is reportedly shifting new data-center leases from 15-year to 25-year terms, which lowers how much of the build cost is recognized as capex in any given year. The 38-gigawatt figure is physical buildout. The pace at which it shows up on Microsoft's financial statements is being managed.
Turning away paying customers is the signal
A hyperscaler declining business because it physically can't serve it is a more concrete data point on the AI capacity crunch than any capex guidance number. Microsoft's FY2027 Q1 capex guidance sits at $50 billion, with roughly $175 billion projected for calendar 2026. If you're a finance leader modeling AI infrastructure costs, weight the turned-away customers most heavily: they are evidence of constrained supply today, where capex guidance only projects demand.