Microsoft Is Spending $190 Billion on a Bet That Compute Stays Scarce

Microsoft plans to spend about $190 billion this year building data centers. Put that in perspective: it is more than half of the $331.8 billion in revenue the company collected in all of its last fiscal year. In the final quarter alone it laid out $41 billion, roughly two-thirds of it on the chips that power AI and lose their value within a few years. This is corporate history scale. The question the market keeps circling is not whether AI demand is real — the data says it is. It is whether money spent at this pace turns into a business.

The skeptics' case is fair enough on its face. Big tech has announced more than $700 billion in combined 2026 capital expenditures, a jump that has strained cash flows and fed a running worry about overcapacity — that the industry is building more compute than anyone will ever rent. Every quarter the spend goes up, and the fear is that profit gets crushed under the depreciation before the demand ever shows up.

The evidence says the opposite so far. Azure grew 43% in the fiscal fourth quarter, beating the roughly 40% analysts expected, and crossed $100 billion in annual revenue for the first time. Microsoft brought in $214 billion of cloud revenue for the year, up 27%. Profit rose 31% to $133.7 billion even as the building continued. And the backlog that predicts future sales — remaining performance obligations — jumped 84% to $678 billion. That is not a company whose capacity is sitting idle; that is a company selling capacity faster than it can build it.

That last point is the one worth sitting with. The scarcest thing in AI right now is not intelligence or models, both of which are becoming abundant and cheap. It is the deployable computing capacity to run and serve them. Microsoft says demand is outstripping supply through at least the end of 2026, so much so that it is having to choose between powering its own AI assistant and renting computing out to customers. That is a good problem, and it is the signal that the build is not speculative: the constraint is supply, not demand. Notably, the company says nearly 90% of its cloud revenue now comes from customers outside frontier model companies — ordinary enterprises, not a handful of AI labs whose demand could vanish. That breadth is what makes the build look less like a bubble bet on one customer and more like durable infrastructure. In the fourth quarter alone it added 31 data centers across five continents and another gigawatt of capacity, doubling total capacity over two years — and it says all of it was monetized immediately.

Here is the wrinkle this year's numbers do not announce. Roughly two-thirds of the spending is going to short-lived assets, mostly the GPUs and CPUs that are depreciated fast. Starting in the new fiscal year, Microsoft extended the estimated useful life of its data centers and office buildings from 15 to 25 years and shifted future data-center leases onto its operating-lease line. Both moves pull depreciation and reported capital spending into the future — the company says the practical effect is that its reported capex for this year lands around $175 billion rather than the roughly $190 billion it is actually investing. The cash outlay is unchanged. So when you see Microsoft talk about "spending less" going forward, what moved was the accounting, not the checks it is writing.

That matters because of how the stock is priced. At roughly $497 a share and a $3.7 trillion market cap, Microsoft trades around 25 times forward earnings. Paying that multiple on a company whose profits are rising 31% is only reasonable if you believe the 40%-plus Azure growth and the backlog that backs it up keep converting into earnings for years. The mechanism of this bet is simple, and it is what the whole position rests on: Microsoft is front-loading a historic amount of cash to buy compute while compute is scarce, and monetizing it into the largest backlog in its history.

The honest caveat sits on the margin line, not the demand line. Microsoft's cloud gross margin was 65% in the quarter, down year over year, squeezed by the very AI infrastructure it is racing to build. The entire argument — the $190 billion, the ~25 times earnings, the headline that demand is surging — comes down to whether that margin stabilizes as the capacity comes online. Demand has already proven it. The conversion of all that spending into profit is what has not happened yet, and it is the number to watch, because it is the one the price already assumes.