Bloomberg: Microsoft's AI Capacity Crunch Forced It to Turn Away Cloud Customers — Now Targets 38GW by 2032

Takeaway: Microsoft's compute shortage has gone from a "known constraint" narrative to a hard, sourced number — the company is targeting more than 38 gigawatts of data-center capacity by 2032, over 3x today's ~12GW, after capacity limits already cost it real customers and caused a GitHub outage. This is the clearest confirmation yet that the hyperscaler under-capacity story (not a demand-destruction story) is real, and it reinforces rather than changes the desk's existing CapEx thesis.

What happened: Bloomberg reported (syndicated via Yahoo Finance at 11:09 AM EDT, with Benzinga's write-through at 11:42 AM EDT, both Sept 11) that Microsoft plans to expand its data-center footprint from roughly 12 gigawatts currently to more than 38 gigawatts by 2032, citing people familiar with the plans. AI-specific capacity today is about 2GW and is projected to be roughly one-third of the 2032 total. The reporting details concrete fallout from the current shortage: Microsoft has turned away some cloud and AI customers for lack of capacity; GitHub suffered an eight-hour outage in August tied to server constraints; Xbox cloud-gaming usage has been capacity-limited; and Temu moved a significant cloud contract to Oracle because Microsoft couldn't serve the region it needed. A Microsoft spokesperson told Bloomberg the reported capacity figures were "inaccurate" but declined to offer alternative numbers — a non-denial that leaves the qualitative picture (severe under-capacity, aggressive build response) intact. Separately, Microsoft's most recent fiscal-year capex ran $145B, and Q3 (reported July 29) showed revenue of $90.01B (+17.7% YoY) and EPS of $4.74, beating the $4.24 consensus by $0.50 — the balance-sheet capacity to fund a 38GW buildout is not in question; execution speed is.

Mag7 impact: This is bullish, on balance, for the AI-infrastructure supply chain and mixed-to-bullish for MSFT itself. For NVDA, a hyperscaler publicly admitting it turned away business for lack of GPU/server capacity is about as clean a demand-durability signal as the tape gets — it dovetails with the desk's existing read that NVDA's IV30 (25.0%) sitting ~19 points under 43.8% realized vol is underpricing a name still riding a locked order book (CapexIndex $947.7B combined, $319.5B AI-attributed; TSMC revenue +53.3% YoY). For MSFT (px 492.44, trend +0.34%, RSI 55.0, RVOL 0.73x — flat/quiet into this print), the read is two-sided: confirmed demand overflow is bullish for the growth story, but a public admission that under-investment already cost it a named customer (Temu, to Oracle) and caused a GitHub outage is a real execution/reputational ding, and a tripling of long-dated capex commitments lands squarely in a market already discounting high-multiple CapEx spenders as rates bite (10Y ~4.83-4.96%, breakevens firming). For GOOGL and AMZN, the read-through is competitive: both are net beneficiaries of workloads Microsoft can't currently host, and this headline underlines why the desk's existing view — GOOGL's CapEx spend ($185.0B, the largest in the group) isn't yet being credited with a return — deserves a second look if displaced-demand wins start showing up in Azure-alternative growth commentary.

What to watch next: Whether Microsoft attaches a capex guidance update or timeline detail to this 38GW figure ahead of its Oct 27 print (EPS est. $4.82); any confirmation or pushback from Microsoft beyond the blanket "inaccurate" denial; incremental hyperscaler-capacity commentary from GOOGL/AMZN earnings calls that could quantify workload migration; and whether NVDA or AMD supply commentary references this capacity math directly. Also worth tracking: this lands the same day as a hotter-than-expected core CPI print (Fed hike odds near 70-85% into next week's FOMC) — a long-duration capex story landing into a higher-for-longer rates backdrop is exactly the tension the desk flagged this morning between "physical demand is real" and "high-multiple spenders de-rate first."