Microsoft Is Publishing the Receipt on AI Capex — and It Shows the Catch

The argument that has shaped AI stocks all year comes down to one fear: the hyperscalers are writing enormous checks to build AI infrastructure that may never turn into real revenue. HSBC models that "hyperscaler overspending" as its most likely scenario, putting a 37% probability on it — the world where the companies doing the biggest spending underperform. That is the prevailing view, and it collapses the entire debate into a single question. Does the money come back out as booked, growing revenue? Microsoft is the cleanest place to watch for that answer, because it is the hyperscaler that actually reports an AI business line. As of its fiscal third quarter of 2026, that AI business hit a $37 billion annual run-rate — meaning, at the recent pace, it would book $37 billion over a full year — up 123% year over year.

Metric	Period	Value (USD bn)	YoY growth
AI business annual run-rate	Q3 FY26	$37.0B	+123%
Commercial RPO	Q4 FY26 year-end	$678.0B	+84%
Copilot annualized run-rate	30M seats × $30/mo	$10.8B	n/a

A run-rate can sound like an abstraction, so it helps to see which parts of the business are doing the work. The headline number includes the Copilot subscriptions and, behind them, revenue from customers drawing on Azure for AI computing. The share that is already booked rather than hoped for shows up in what Microsoft calls commercial remaining performance obligations, or RPO — the value of contracts signed but not yet delivered as revenue. That book hit $678 billion at fiscal 2026 year-end, up 84% year over year, with a weighted average contract duration of 2.3 years. An 84% jump in committed future revenue is the concrete check on the fear that the buildout is happening with no enterprise buyer attached. Someone has signed.

The Copilot piece is where the ambition and the math both show. At 30 million paid seats and $30 per user per month, Copilot annualizes to about $10.8 billion. But 30 million seats is still only about 6.5% of the roughly 464 million commercial Microsoft 365 seats Microsoft can reach, and more than 90% of eligible users have not converted yet. Push that penetration to 10% — 46 million seats — and the annualized figure roughly becomes $16.6 billion, all before the broader Azure AI consumption line. That is the compounding machine, and the part that can disappoint is painted in the same numbers: it is a penetration story, not one that is done. So the buildout is converting into booked, growing revenue. Azure crossed $100 billion in annual revenue in fiscal 2026, growing 43% year over year in the fourth quarter, and Microsoft guided roughly 45% constant-currency growth for the first quarter ahead. Here is the catch, and it is the reason to hold the enthusiasm down a notch. The conversion is not yet margin-accretive. Microsoft Cloud gross margin slipped to 65% in the fiscal fourth quarter from 66% the quarter before, pressured by the AI infrastructure investment and by growing AI usage itself — the more customers draw down the compute, the more it costs to serve them.

That margin slip is the honest boundary on the whole story. The AI line is self-funding in the sense that it is real, booked revenue that keeps growing. It is not yet self-funding in the stronger sense — the one where the spend pays for itself. What those two statements separate is whether the AI trade sits in the "overspending" world HSBC worries about or the "euphoria" world it prices at just 8%. One more thing worth flagging so you do not over-measure with a false precision: there is no agreed denominator for the capex side of this trade. One research house puts 2026 hyperscaler AI capex above $300 billion, another at roughly $775 to $800 billion across five spenders, and the two are not reconciled on what counts as a lease. So do not try to build a clean "revenue per dollar of capex" ratio from these figures — the gap between them is a definitional one, and no such ratio is defensible. What you can watch cleanly is the variable the whole thesis turns on: whether Azure's growth holds as the capex keeps rising. If Azure keeps compounding near that 43%-plus pace while hyperscaler budgets climb, the monetization story is doing its job. If that growth decelerates well below 43% while the spending still accelerates, the self-funding thesis breaks, and the overspending scenario the market has been circling finally gets its receipt.