Microsoft more than triples its gigawatt plan after Azure sent GitHub's agents to AWS

Only about 2 of Microsoft's current 12 gigawatts is built around AI-specific accelerators ^[4]. The 2032 plan takes AI silicon to roughly a third of the total, about 12.7 gigawatts ^[5], which Tech Times reads as a roughly six-fold rise in dedicated AI compute over six years ^[6]. Do the subtraction on the other two thirds and general-purpose capacity goes from about 10 gigawatts to about 25.3, up 2.5 times ^[2].

Who gets the capacity while it is short has already been settled. CFO Amy Hood told the July 2026 earnings call that new GPUs and CPUs go to first-party workloads first, with the remainder made available to Azure customers ^[8]. That ordering puts Copilot ahead of paying tenants. Temu had signed contracts with competing cloud providers when Microsoft could not meet its compute needs ^[2], and the gaming division capped cloud streaming time for Xbox subscribers ^[7].

Satya Nadella has been explicit about the binding input. "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," he said in a 2025 podcast interview ^[9]. Hood told investors on a 2025 earnings call: "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." ^[10]

The spending rate makes the six-year target arithmetically plausible. Capital expenditure was $145bn in fiscal 2026 ^[11], about $175bn on a calendar-year basis once finance leases are counted ^[12], and the guide for the first quarter of fiscal 2027 is roughly $50bn ^[13], which annualises to $200bn ^[3]. Set against the $145bn, the $329bn of leases signed and not yet commenced is about 2.3 years of capex ^[4]. Spread across the 26 gigawatts the plan adds ^[1], it comes to about $12.7bn a gigawatt ^[5], and that figure flatters the true cost, because owned campuses are in the plan as well and rented neocloud compute from the likes of CoreWeave is excluded from the 38 ^[15].

From fiscal 2027, the useful life Microsoft assigns to data centers and office buildings goes from 15 years to 25 ^[16]. Per dollar of asset, the annual depreciation charge falls from 6.7 cents to 4, a 40% reduction ^[6]. On a cohort the size of fiscal 2026's capex, that is $5.8bn a year instead of $9.7bn ^[7].

Tech Times wrote that "The 38-gigawatt target is not forward-looking ambition. It is a response to documented failure." ^[17] Half of that survives contact with the dates. The failures are specific and behind us: nine GitHub outages in May 2026 ^[3], a named customer buying elsewhere ^[2]. The 26 gigawatts arriving between now and 2032 answers none of them, and the leases running to fiscal 2033 price demand six years out ^[14].

Two readings compete. If first-party workloads keep absorbing the increment before Azure tenants see it, 26 gigawatts is a conservative number and the lease book grows again. If AI demand flattens around 2029, the leases commence on their contractual schedule anyway and the 25-year life is where the shortfall shows in reported earnings ^[16]. The case for a durable shortage would fail if Azure got back into balance without the build, and the route to that is the rented neocloud capacity the 38-gigawatt figure leaves out ^[15].