The AI infrastructure race crossed a new threshold this week, and the numbers involved are large enough to be measured against entire US states rather than individual companies.

Microsoft revealed plans to more than triple its global data-center footprint to over 38 gigawatts by 2032 — a target that would exceed New York State's peak electricity use — while Oracle reported a record $664 billion order backlog that stunned even Wall Street analysts who'd already priced in aggressive AI growth.

Microsoft's capacity crunch, in its own numbers

Microsoft's plan, first reported by Bloomberg, calls for growing its data-center capacity from roughly 12 gigawatts today to more than 38 gigawatts by 2032 — more than triple its current footprint in a six-year window. Only about 2 gigawatts of that current capacity is dedicated specifically to AI workloads today; by 2032, AI is expected to represent roughly a third of the much larger 38-gigawatt total.

The company isn't building at this scale out of abstract ambition — it's already been forced to turn away business it can't currently serve. Temu reportedly moved a major cloud deal to Oracle after Microsoft couldn't provide enough capacity in the regions the retailer needed. GitHub suffered an eight-hour outage in August tied directly to server constraints, with some developer traffic reportedly rerouted to Amazon because Azure had no room. Microsoft has also placed limits on Xbox cloud-gaming usage for paying subscribers — a consumer-facing sign of a capacity shortage that started as an enterprise problem. The build-out will draw on a mix of owned campuses and leased capacity from neocloud providers including CoreWeave, Nscale, Lambda, Iren, and Nebius, and it sits on top of a capital budget that's climbed from $55.7 billion in 2024 to an estimated $145 billion in 2026.

Oracle's backlog is the clearest evidence demand hasn't slowed

If Microsoft's numbers show a company straining against a capacity ceiling, Oracle's latest earnings show what happens on the other side of that same shortage. Oracle's remaining performance obligations — essentially contracted revenue not yet recognized — surged to a record $664 billion, well above the roughly $640 billion Wall Street had modeled. Quarterly revenue rose 30% year-over-year to $19.35 billion, and Oracle Cloud Infrastructure revenue specifically climbed 121% to $7.4 billion. The stock jumped roughly 7% in after-hours trading and another 5.5% in premarket the next morning on the news.

Oracle's CFO said the company expects around half of that $664 billion backlog to convert into actual revenue over the next 36 months — a detail investors are watching closely, since a backlog is only as valuable as how reliably and quickly it turns into cash-generating sales. Notably, Oracle has also shifted part of its financing approach: rather than leveraging its own balance sheet to build capacity for customers, as it traditionally has, this quarter saw more customers prepaying and bringing their own chips to the arrangement — a sign, according to industry analysts, that customers are now willing to fund capacity themselves simply to secure access to compute that's in short supply everywhere else.

Why the same story is playing out at Microsoft and Oracle simultaneously

Both companies are responding to the identical underlying condition: AI compute demand that continues to outrun available supply, even after years of aggressive infrastructure spending. The five largest hyperscalers — Amazon, Google, Meta, Microsoft, and Oracle — were on track to spend more than $750 billion on capital expenditures in 2026 alone, roughly 67% more than the prior year, with an estimated 75% of that spending earmarked specifically for AI infrastructure, according to Morgan Stanley analysis. That's a level of physical capital deployment usually associated with national infrastructure projects, not corporate technology budgets.

Power, not chips, is increasingly described as the binding constraint on how fast this buildout can actually happen. Microsoft's 38-gigawatt target runs into real-world limits: grid interconnection queues in some regions stretch four to seven years, and moratoriums on new data-center construction have already emerged in markets including Texas and New York. Nuclear power has drawn particular attention as one of the few generation sources capable of providing the continuous, round-the-clock power AI training clusters require, since solar and wind can't reliably run workloads that operate 24 hours a day — though even a notable deal like Microsoft's agreement tied to the Three Mile Island site covers only 835 megawatts of the 38-gigawatt target, illustrating just how large the remaining gap is.

What this means if you're watching this space

• If you're evaluating exposure to the AI infrastructure theme, understand that capacity constraints — not customer demand — are now the limiting factor for the biggest players. Both Microsoft and Oracle's numbers this week point to demand comfortably outrunning supply, a dynamic that tends to favor whoever can actually deliver capacity fastest.
• If you're a business relying on Azure, AWS, or another hyperscaler for growing compute needs, plan for continued capacity friction in the near term. Microsoft's own admission that it's turning away business is a concrete signal that availability, not just pricing, is a live constraint for enterprise customers right now.
• If you're tracking where capital is actually flowing in this cycle, power and grid infrastructure deserve as much attention as chips. The gap between what a single nuclear power deal provides and what a 38-gigawatt target requires shows how much of this buildout's success now depends on energy infrastructure that isn't Microsoft's or Oracle's to build alone.

Frequently Asked Questions

Is Microsoft's 38-gigawatt target realistic given the grid and permitting constraints described? It's an ambitious six-year plan facing genuine external obstacles — grid interconnection queues running four to seven years in some regions and construction moratoriums in markets like Texas and New York are real constraints outside Microsoft's direct control, which is part of why the company is pursuing a mixed strategy of owned campuses and leased neocloud capacity rather than relying on a single build approach.

Does Oracle's $664 billion backlog mean that revenue is guaranteed? No — a backlog represents contracted future obligations, not recognized revenue. Oracle's own guidance expects only about half of that figure to convert into actual sales over the next 36 months, and how reliably that conversion happens is exactly what analysts are watching most closely going forward.

Why did GitHub and Xbox get affected by a cloud capacity shortage? Both run on Microsoft's Azure infrastructure, and the reported capacity strain that forced Microsoft to turn away some new cloud and AI business has also reportedly reached existing services — GitHub experienced an outage tied to server constraints, and Microsoft placed usage limits on Xbox cloud gaming, illustrating that capacity pressure isn't limited to new enterprise deals alone.

Sources & References

• Bloomberg, "Microsoft AI Focused Data Center Plan to Add 26 Gigawatts of Compute"
• Reuters (via Investing.com), "Microsoft plans 38 gigawatts of data center capacity by 2032"
• Yahoo Finance / TradingView, "Microsoft Targets 38GW Data Center Buildout as AI Capacity Crunch Bites"
• Tech Times, "Microsoft Data Center Expansion to 38 GW Triggered by Lost Clients and Capacity Crisis"
• NAI 500, "Oracle's AI Boom Sends Backlog to $664B"
• TechStartups, "Top Tech News Today, September 11, 2026"
• Michael Parekh, "'Business as Usual' in AI. Microsoft, Oracle & SpaceX"