Skip to content
Beyond Prompt AI Studio

Strategy & research

Azure's growth number is not a demand number – how to read an earnings report as a capacity forecast

July 31, 2026 · 12 min read · Beyond Prompt AI Studio

CapacityAzureMicrosoftMetrics

On the evening of 29 July 2026, Microsoft published its results for the fourth quarter of the fiscal year: 90 billion US dollars in revenue, Azure up 43 percent, the stock jumping roughly nine percent overnight. Headlines celebrated it as proof that the enormous AI investments are finally paying off. This analysis reads the same report differently. In the same earnings call, CFO Amy Hood explicitly confirmed that demand continues to exceed available supply. If that is true, the celebrated growth number is not a statement about how much customers wanted to buy – it is a statement about how much Microsoft managed to deliver that quarter. That is not semantics. It changes what the number can legitimately be used for – and, read correctly, it becomes one of the most useful metrics currently public for your own capacity planning.

Key points at a glance

  • CFO Amy Hood stated explicitly on the earnings call that demand exceeds available supply. That makes Azure's 43 percent growth rate not a demand number but a delivery number – it shows what Microsoft managed to ship, not what customers wanted to buy.
  • The guidance of roughly 45 percent for the current quarter is accordingly not a demand forecast but a commissioning commitment. That is exactly what makes it usable as a rough capacity forecast for your own planning.
  • Reported investment and actual construction are diverging: an accounting change (useful life of buildings extended from 15 to 25 years) makes the capex figure look more moderate. It explicitly applies to buildings, not GPUs – and the same quarter saw 31 new data centers brought online, 88 for the full fiscal year.
  • The real bottleneck has shifted from capital to time-to-live: Nadella cited the near-50-percent cut in the time from GPU delivery to production as its own success metric on the call. An operational bottleneck cannot be accelerated with more money, only with logistics.
  • Microsoft's own internal AI demand – over 30 million paid Microsoft 365 Copilot seats, 50 million GitHub Copilot users, 100,000 Foundry customers with doubled revenue – is itself growing fast and is demonstrably prioritised during shortages. It is the yardstick against which your own wait time is measured.
  • For your own planning, a simple rhythm follows: check every quarterly report from the major hyperscalers for three sentences – the stated ratio of demand to supply, the number of newly commissioned sites, and any comment on the provider's own Copilot usage. Those three sentences say more about your own wait time than any availability marketing claim.

What was reported on 29 July

Microsoft's fourth-quarter numbers read unambiguously positive at first glance: 90.01 billion US dollars in revenue, up 18 percent year over year. Azure grew 43 percent, crossing 100 billion in annual revenue for the first time in the company's history. The AI revenue run-rate climbed to 37 billion. For the current quarter, Microsoft guided Azure growth of roughly 45 percent at constant currency – acceleration, not a plateau. The stock gained about nine percent overnight, adding roughly 260 billion dollars in market value in a single evening.

These numbers are accurate and impressive. But they answer a different question than the headlines assume. The headlines read 43 percent growth as proof of strong demand. In the same call, CFO Amy Hood said something that directly contradicts that reading: demand, she stated, continues to exceed available supply. CEO Satya Nadella added that the company brought 31 new data centers online across five continents that quarter, 88 for the full fiscal year – and cut the time from delivery of new GPUs to their production use in the largest regions by nearly 50 percent.

Why 43 percent is a delivery number, not a demand number

The distinction sounds academic at first, but it is not. In a market where supply and demand settle freely, a growth rate roughly reflects both – how much customers wanted and how much the provider was willing to supply meet at the price. Hood's statement describes a different market: one where supply is the binding constraint. In that case, the growth rate is essentially measuring one thing – how quickly Microsoft could bring new capacity online. Demand could just as easily have been 43, 60 or 80 percent; the number simply does not tell you. All it tells you is that demand was at least as high as what got delivered.

This reframing has an immediately useful consequence. If the growth rate is a delivery number, then the guidance for the coming quarter – roughly 45 percent – is likewise not a demand forecast but a statement about how much additional capacity Microsoft expects to bring online over the next three months. That is one of the few reasonably reliable, publicly available figures for how quickly the general availability situation at one of the largest providers is actually changing. For your own planning it is more useful than any availability marketing claim, because it is an audited number backed by securities law, not a PR phrase.

The investment figure that conflates two different things

A second detail from the same report deserves closer scrutiny, because part of the coverage got it wrong. Starting fiscal year 2027, Microsoft is extending the estimated useful life of data centers and office buildings from 15 to 25 years. That lowers annual depreciation, and because more future data-center leases will now be classified as operating rather than finance leases, the reported capex figure also falls without actual construction slowing down. Important for interpretation: this change explicitly applies to buildings and infrastructure, not to the compute chips themselves – by the company's own account, roughly two-thirds of recent capex went to short-lived assets, primarily CPUs and GPUs, which follow a much faster replacement cycle than buildings.

The point here is not to accuse Microsoft of accounting spin – the change is disclosed and justified in a way that is standard for the industry. The point is methodological: a reported investment figure from a quarterly report is not a reliable measure of how much physical infrastructure is actually being built. Anyone trying to derive a capacity forecast from it should instead rely on operational metrics that are harder to dress up – such as the number of newly commissioned data centers. And that number points the same direction as the growth rate: more, not less.

The bottleneck has moved: from capital to time

The most revealing sentence of the entire call may be its most unassuming one. A CEO citing the reduction in time from GPU delivery to production use as his own success metric on an investor call is a statement about the nature of the bottleneck. A year ago, the dominant question was whether there was enough capital for enough chips. That is now largely solved – the investment figures speak for themselves. The dominant question now is how quickly purchased chips actually turn into usable capacity: power hookup, cooling, network connectivity, commissioning.

That distinction matters because the two types of bottleneck are solved differently. A capital bottleneck can be addressed – as last week's guarantee discussions show – through financing structures, however contentious those may be in a given case. An operational bottleneck in logistics and commissioning cannot be accelerated with more money, only with better processes, and those typically improve more slowly than capital can flow. So when even the world's largest cloud provider treats a near-50-percent cut in commissioning time as noteworthy, that is a signal of just how tight this operational chokepoint actually is.

Who is ahead of you in the queue – and how fast that queue is growing

In an earlier analysis, we showed that Microsoft prioritises its own Copilot products over customer workloads when capacity is scarce. This report supplies the current numbers behind that: more than 30 million paid Microsoft 365 Copilot seats, 50 million GitHub Copilot users, 100,000 Foundry platform customers with revenue more than doubled year over year. This internal demand is not static – it is itself growing fast, and according to Hood, efficiency gains in capacity are quickly monetised because the imbalance between demand and available resources persists.

For your own wait time that means: it depends not only on how fast Microsoft builds new capacity, but also on how fast the provider's own internal demand grows. If both grow at roughly the same pace, an external customer's relative position stays unchanged even as absolute capacity keeps rising. That is an uncomfortable but important insight: rising construction activity alone is not yet a reliable signal that the allocation situation is easing.

What this means in practice

The value of this analysis is not in distrusting Microsoft's numbers, but in using them for a purpose they were not designed for: as a rough but usable leading indicator for your own capacity planning. Three elements of a quarterly report are useful for this; the rest mostly are not.

  • The ratio of demand to supply, when management comments on it explicitly. A statement like demand continues to exceed supply is rare, precise and legally binding – companies do not phrase that lightly on an investor call.
  • Operational metrics over financial ones: number of newly commissioned data centers, statements about commissioning time, concrete regional figures. These are harder to dress up than investment totals and closer to physical reality.
  • Growth of the provider's own internal AI products (Copilot seats, internal platform usage). This shows how fast the competition for the same capacity is growing against an external customer.

These three points can be pulled from your main provider's investor-relations page or earnings-call transcript in a few minutes per quarter, whether that provider is Microsoft, Amazon or Google. This is not elaborate reporting, just a short, recurring check – and it does not replace contractually securing your own capacity, but it adds a sense of which direction the situation is likely to move over the coming months.

What explicitly does not follow

This analysis does not conclude that Microsoft is spinning the capacity situation or manipulating its figures – the accounting change is disclosed, justified in a standard way, and demonstrably does not touch the short-lived assets that determine actual compute capacity. Nor does it follow that switching cloud providers is the obvious response – the same supply shortage currently affects all major hyperscalers to varying degrees, and a switch does not solve the structural problem, at best it relocates it. What follows is simply a more disciplined way of reading a number that public debate almost consistently misreads.

Conclusion

A quarterly report like this one is routinely read as a success story, and at the company level it undoubtedly is one. For a company planning an AI project with exactly this provider, though, it is above all something else: the most precise publicly available signal so far that the supply constraints your own planning has to reckon with are not easing, only relocating – from a capital question to a logistics question, from a construction problem to an allocation problem. Anyone who from now on reads these reports for three sentences instead of one headline knows more about their own wait time than any availability guarantee in their own contract could tell them.

Frequently asked questions about Azure's growth numbers and capacity planning

Does 43 percent Azure growth mean the capacity shortage is easing?

No, if anything the opposite. The growth rate shows how much Microsoft managed to deliver that quarter, not how much customers demanded. The CFO confirmed on the same call that demand continues to exceed available supply. High growth alongside persistent scarcity means more capacity is coming online, but not necessarily enough to close the gap.

Why is less investment reported when more is actually being built?

Because of an accounting change: Microsoft is extending the estimated useful life of data centers and office buildings from 15 to 25 years, which reclassifies more future leases as operating rather than finance leases and lowers the reported investment figure. This change explicitly applies to buildings, not to the short-lived compute chips. The same quarter saw 31 new data centers commissioned – a sign that actual construction activity has not declined.

How can we use these numbers for our own planning?

Not as a precise forecast, but as a rough leading indicator. Three elements of a quarterly report are useful: explicit statements on the ratio of demand to supply, operational metrics such as newly commissioned data centers, and growth in the provider's own internal AI products. These three points can be pulled from your main provider's earnings-call transcript in a few minutes per quarter, and they complement – but do not replace – contractually securing your own capacity.

Should we switch providers because of the capacity shortage?

Not for that reason alone. The supply shortage currently affects all major hyperscalers to varying degrees, so a switch does not fundamentally solve the structural problem. It makes more sense to keep your own architecture switch-capable and monitor the three metrics above regularly to catch shifts in the availability situation early.

Do you want your AI capacity planning built on solid signals instead of gut feel?