Microsoft Is Turning Away Cloud Customers. Its Answer Is 38 Gigawatts.

Microsoft Is Turning Away Cloud Customers. Its Answer Is 38 Gigawatts.

The sentence I did not expect to read this week: Microsoft — the second-biggest cloud on this planet — has been telling paying customers it cannot serve them. Not next month. Not in one unlucky region. Now, on real deals.

Bloomberg broke it late Thursday: Microsoft plans to grow its data center footprint from roughly 12 gigawatts today to more than 38 gigawatts by 2032, according to people familiar with the plans. That is a power draw exceeding what the entire state of New York uses at peak. It is the company's answer to a shortage that has already cost it enterprise customers, restricted a paid Xbox cloud service, and pushed developer traffic off Azure and onto Amazon.

Then Microsoft's own spokesperson told MT Newswires that reports speculating about its future buildout "do not correctly reflect the company's financial results and projections," and pointed everyone to its SEC filings.

Sit with that. The biggest capacity number Microsoft has ever been attached to is not in the filings. Read a number that deniable like an engineer, not an investor.

The Number, the Denial, and the Customer List

Start with what is not in dispute, because that part comes from Microsoft's own mouth.

On recent earnings calls, CFO Amy Hood said demand across workloads, customer segments and regions "continues to exceed available capacity." Then she went further: "We are, and have been, short now for many quarters. I thought we were going to catch up. We are not. Demand is increasing."

That is a CFO admitting the company misjudged its own demand curve — after pausing data center expansion, the decision now blamed for the crunch it is living through.

CEO Satya Nadella has been blunter. Microsoft has GPUs sitting in inventory it cannot plug in, and roughly $80 billion of Azure orders sit unfulfilled. The constraint is not chips. Hood confirmed it from the finance side: access to hardware has not been the problem — space and power have.

Now look at what the shortage cost. Enterprise clients that could not be served went elsewhere. Microsoft limited Xbox cloud streaming for paying subscribers. GitHub — a Microsoft-owned property — routed developer traffic to Amazon because Azure had no room for it. Azure's UK South region was declining capacity requests months ago.

The catch-up is enormous. Microsoft added a gigawatt of capacity in a single quarter and opened 31 data centers across five continents. Its commercial remaining performance obligation — contracted revenue not yet delivered — hit $678 billion, up 84 percent year over year, nearly twice its entire fiscal 2026 revenue.

The First Reading: Nobody Turns Away Paying Customers Unless Demand Is Real

The bullish version of this story is strong and I will not pretend otherwise. Microsoft has sold $678 billion of future cloud and AI services. Copilot passed 30 million paid seats, up from roughly 20 million a quarter earlier. Azure keeps beating expectations whenever new capacity comes online early. With that much contracted demand and still no ability to deliver it, the problem is not demand.

For anybody in this industry that is a good signal. I have watched cloud providers spend a decade inventing reasons customers should move to them, and here is a company with a queue. But hold one thought: if Microsoft turns away business, that demand did not disappear. It went somewhere. "Somewhere" is the part most people get wrong.

The Second Reading: Two-Thirds of That Tripling Is Not AI

Of Microsoft's roughly 12 gigawatts today, only about 2 gigawatts is AI-specific silicon — the Nvidia GPUs everybody argues about. The other 10 gigawatts is ordinary CPU compute: databases, storage, network fabric, enterprise applications. By 2032, Microsoft expects AI-specific capacity to be about a third of the 38 gigawatts. Call it 12 to 13 gigawatts.

Run the arithmetic. That is roughly a sixfold increase in AI compute — real and enormous — but only about two and a half times growth in general-purpose cloud capacity. The celebrated tripling is mostly commodity cloud. As one Microsoft infrastructure executive put it, GPUs by themselves do not make great AI infrastructure. You need massive CPU fleets next to them, or the GPUs have nothing to talk to.

Then look at how the money was reported, because that tells you more than the gigawatts. Microsoft's calendar-2026 capital expenditure guidance moved from about $190 billion to roughly $175 billion. Nothing got cancelled. Microsoft extended the estimated useful life of its data centers from 15 years to 25, which reclassifies a chunk of future leases from finance leases into operating leases — and operating leases drop off the reported capex line. Same buildings, same cash obligations, different bucket.

Meanwhile two-thirds of that capex is for short-lived assets, primarily GPUs and CPUs. You cannot depreciate a GPU for 25 years. The buildings might last that long; the silicon inside them is a five-year asset wearing a twenty-five-year building's depreciation schedule.

There is also a number most coverage skipped. As of June 30, 2026, Microsoft reported more than $329 billion of data center leases signed but not yet commenced — contracts activating between fiscal 2027 and fiscal 2033. That is not current debt, not deployed capacity, and not in the published capex number. It is a promise about a decade.

The Bottleneck Nobody Wants to Say Out Loud: A Gigawatt Announced Is Not a Gigawatt Plugged In

Nadella used the phrase everyone should tape to their monitor: warm shells. He does not need chips. He needs buildings with power connected and cooling running, ready to receive racks he already owns.

That gap is the whole story, and it has nothing to do with Microsoft's balance sheet. Of the roughly 12 gigawatts slated to come online in North America during 2026, only about a third has broken ground. Between one-third and one-half of expected 2026 builds are slipping or being cancelled, and estimates of 2026 AI capacity pushed to 2028 run as high as 30 to 50 percent. Every one of those is somebody's headline capacity figure that will not be energized on schedule.

Power tells the same story from the other side. The PJM capacity auction — the market serving the biggest data center corridor on earth — fell short of its reliability requirement in December 2025 by 6,623 megawatts, the first shortfall in that market's history, and cleared at the price ceiling of $333.44 per megawatt-day. In July 2026 it failed again, 6,831 megawatts short. Two consecutive failed auctions is not an event. It is a regime.

The equipment queue does not negotiate either. Large transformers ordered today arrive around 2028, interconnection agreements run on timelines measured in years, and the pipeline is stuffed with projects that will never be built — the industry calls them phantom, and the phantom rate has run as high as 72 percent.

For the record: 38 gigawatts by 2032 means adding roughly a gigawatt per quarter for 26 straight quarters. Microsoft has never done that. And the plan excludes capacity rented from neoclouds, so the real footprint is a blurrier number.

What This Actually Means for Independent Hosting Providers

First — the customer you will actually meet is not Microsoft's whale. The clients Microsoft is losing are not all frontier labs. A company that cannot get a database server in its home region does not care about GPU clusters. It cares about running its business next quarter — a customer an independent host can serve today, with hardware that does not need a substation.

Second — watch energized capacity, not announced capacity. A decade-long gigawatt plan is a press release. A substation being energized is a fact. Track lease commencements, interconnection approvals, and local permitting votes in your own markets. They move months before any capex headline does.

Third — price against their depreciation, not their marketing. Microsoft needs 25-year asset lives to make the math work on buildings housing five-year silicon. You can buy last-generation hardware, run it five years, and beat their price on any workload that is not model training.

Fourth — the $329 billion lease pipeline is your early-warning system. When signed leases commence late, get renegotiated, or quietly disappear, that is the first honest signal the demand picture changed — long before it shows up in a capex number.

Fifth — speed to power is a product. If you can energize a megawatt next quarter and Microsoft cannot energize a campus before 2029, that difference has a price. Sell it. The whole industry is bidding for the same electrons and the same transformers, and the smaller operator who already has both holds an asset the giants cannot manufacture on demand.

The Counter-Argument — "Give Them Two Years and They'll Catch Up"

The obvious objection: Microsoft has more money than most countries, so give them a couple of years and they will build their way out of it.

Hood already answered that. She expected Microsoft to catch up. Her words were "We are not." That was not one tough quarter — the company has been short for many quarters with demand still climbing, and the constraint is not money. Microsoft can fund $175 billion a year of construction. What it cannot buy is a grid connection that does not exist, a transformer nobody has built, or a county board that has not voted.

The bull case and the risk here are the same fact: the demand is real, and the physical world will not let them serve it at the pace they promised.

The Bottom Line

Microsoft is not in trouble. Microsoft is constrained — publicly, in writing, by a CFO who said out loud that she thought they would have caught up by now. The company has $678 billion of contracted work and cannot energize fast enough to deliver it. That is a good problem to have and a brutal one to fix.

But do not let the AI framing fool you. Two-thirds of that 38-gigawatt plan is commodity CPU cloud — the exact business small hosts compete in — and the customers getting squeezed out are ordinary businesses that need a server, not a supercomputer.

A backlog is a promise. An energized rack is revenue. Microsoft has a decade of promises and a transformer queue that says otherwise.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: Bloomberg ("Microsoft Plans Data Center Push to Triple Its Computing Power," September 10, 2026; Bloomberg Intelligence podcast with Anurag Rana, September 11, 2026), Reuters, MT Newswires, TechTimes, Microsoft fiscal 2026 results and earnings call commentary (Satya Nadella, CFO Amy Hood), DatacenterDynamics, DIGITIMES, Sightline Climate via Bloomberg, PJM Interconnection capacity auction results, Jefferies and Aterio capacity data via Axios, Lambda Finance operator capacity estimates.