Microsoft Has Sold More AI Compute Than Its Data Centers Can Deliver

Microsoft ended its fiscal year with $678 billion in cloud services it has been contracted to deliver but has not yet built the capacity to fulfill. That figure — "commercial remaining performance obligations," the label Microsoft puts on services it has booked but not yet handed over — grew 84% in a single year. Independent analysts put a number on the strain behind it: in the busiest regions, demand is running at roughly three times available capacity, and a cloud build that used to take six months now stretches past eighteen.

That gap, not the size of the network, is what Microsoft's data-center story is actually about.

It has sold more than it has built

Microsoft did not frame it that way on its July 29 call. It touted 88 new datacenters opened during the fiscal year, 31 of them in the final quarter alone, a fleet whose capacity has doubled within two years, Azure growth back above 40%, and Azure revenue crossing $100 billion for the first time. All of that is real. But it describes the supply side of a company whose demand side has run ahead of it: Azure grew 43% in the quarter precisely because new capacity finally let it, and management still describes demand as exceeding supply.

The tension matters because a hyperscaler books revenue on usage, not on signing. The backlog is the load-bearing number for the next several years of Microsoft's income statement, and a large share of it arrives through a single funnel — roughly 45% of the booked pipeline is tied to OpenAI. Microsoft is managing a contracted obligation it cannot yet physically meet, which is a very different portfolio position from owning idle buildings.

The bottleneck is no longer the chip

Here the story stops being about Microsoft and becomes about physics and permitting. The binding constraint moved off silicon some time ago. Now the hard limits are high-density power, grid transformers, liquid cooling, and concrete, with the tightest single point sitting at the interface where the electrical grid meets the data-center floor. Power availability, not chip supply, is the governing factor in AI deployment.

That re-frames the coverage of a sprawling build-out. The useful unit of measure is not the datacenter or the gigawatt; it is how fast a megawatt can be turned on and what the delivered compute costs. Microsoft can announce a 3,200-acre site in Wyoming and a wave of purpose-built facilities, but each project joins a construction queue that has already doubled its own delivery time. When that build rate is the constraint, a slipped schedule is systemic execution risk, not a hiccup.

The cash behind the megawatts

Capacity is not free, and the books are telling two stories at once. Fiscal-year capital spending jumped about 80%, from roughly $65 billion to $116 billion, with the final quarter alone at $41 billion — two-thirds of it short-lived assets like CPUs and GPUs, one-third long-lived buildings. That outlay is consuming the company's own cash generation: free cash flow came in around $67 billion versus about $72 billion a year earlier, while cash reserves were drawn down from roughly $95 billion to $77 billion to fund the build.

Then comes the part the headline numbers tend to gloss over. Microsoft stretched the depreciable life of its datacenters and buildings from 15 to 25 years and shifted future datacenter leases from finance to operating treatment. Both choices push cost recognition into the future — lower depreciation hits profit now, and lease charges move below the operating line. The stated hope of remaining free-cash-flow positive in the coming year is notable less as reassurance than as a tell that keeping this spend cash-neutral is a live problem.

What the megawatts are worth

The bullish case is legitimate, and it is the reason the stock has withstood all of this. When you enjoy a supply constraint, every unit of capacity you add is monetized almost immediately, at prices the demand side is willing to pay; the backlog that sits unredeemed is contracted future revenue, not a guess. If the demand is durable — and it is not purely a Microsoft artifact, with Google Cloud and AWS growing even faster off real enterprise workloads — the fleet is a wound-up spring that pays out over years.

The risk is that the payout only happens at the speed power and concrete allow, and that the reported economics are being partially flattered by accounting that pushes the true cost further out. Microsoft has built itself into a position where its best asset and its binding constraint are the same thing. The number that decides whether the bet pays is not how many datacenters or how many gigawatts — it is whether each megawatt of new capacity converts the backlog into cash at a margin that survives the accounting choices, and whether the few customers on the other side of that pipeline stay. Everything else in the data-center press release is a measurement of the constraint, dressed up as an achievement.