AI Steady, Cloud Accelerating Gives Microsoft A Big Datacenter Boost

Wall Street has been looking for some good news, and Microsoft came through with its financial results for the third quarter of its fiscal 2025 as its cloud business – and to be specific, its non-AI cloud business – grew much more strongly than expected.

In the quarter ended in March, Microsoft posted sales of $70.1 billion, up 13.3 percent year on year and up a smidgen sequentially. Operating income, despite all kinds of pressures including the depreciation of a portion of its huge fleet of systems underpinning the Azure cloud, rose by 16 percent to $32 billion, which was up 4.7 percent sequentially.

If we are entering a recession, either Microsoft is not feeling it yet or it is benefitting from the fear or the actuality of it. Time will, as it always does, tattle.

In any event, the story a quarter ago was that Azure could not make up for a profit decline in Microsoft’s on premises datacenter business, which has Windows Server and its myriad platform extensions on top of it. And now, that on premises business is still under pressure but customers are moving enterprise workloads that have nothing to do with AI to the cloud as well as upgrading their on premises Windows Server estates.

The part of Microsoft that we care about most as the chronicler of platforms is the Intelligent Cloud group, which includes the Windows Server stack and Azure together. To be more precise, Intelligent Cloud encompasses Windows Server, SQL Server, Visual Studio, and other middleware and tools sold into the datacenter as well as compute, storage, and networking capacity on the Azure cloud, plus related platform PaaS and SaaS services Microsoft peddles into the datacenter and on Azure.

This Intelligent Cloud group had $26.75 billion in sales, up 20.8 percent year on year. Operating income for Intelligent Cloud rose by 16.6 percent to $11.1 billion, which represented 41.5 percent of revenues and which is on the low end of average but considerably more profitable than a decade ago despite the enormous funds Microsoft sinks into datacenters.

One of the reasons that Microsoft can indulge in cloud building is that its application software lines for PCs and servers are a little bit bigger and considerably more profitable. The Productivity and Business Processes group had sales of $29.94 billion in Q3, up 10.4 percent, and it had operating income of $17.38 billion, up 14.8 percent. The operating income for applications represented 58 percent of revenues, and is by far the most profitable thing Microsoft sells. (Windows for PCs probably had higher margins back in the day when Microsoft had a monopoly on desktop and laptop operating systems.)

Microsoft lumps all of its cloud-based software and services into a single category called Microsoft Cloud, and it is important to remember that this includes Azure but it is not solely Azure. It includes anything sold under a utility model from Microsoft using its infrastructure – so think of it as infrastructure, software, and platform as a service for both servers and PCs all rolled up into one pile.

In the third quarter of fiscal 2025, Microsoft Cloud revenues rose by 201. Percent to $42.45 billion and gross profits (not operating income) rose by 15.1 percent to $29.29 billion, which was 69 percent of revenues.

Microsoft does not provide revenue and operating income figures for the Azure cloud, and it also does not break out its full systems business spread across its Azure regions and the on premises datacenters of its millions of server-product customers worldwide. We do our best to try to reckon these two numbers every quarter, and we also fully realize that there is some serious guesstimating going on here.

In our model, we believe that Azure drove $16.58 billion in sales, up 33 percent from the year ago period. This represented 62 percent of Intelligent Cloud revenues. Our best guess is that Azure had $6.88 billion in operating income, which is 41.5 percent of revenues.

If you want to understand the core systems business that Microsoft has built, you have to add up the basic on premises software in the Windows Server stack – not including databases and development tools – and add it to the capacity sold on Azure to run applications, which does by necessity include the cost of renting the underlying hardware. (With on premises gear, Microsoft customers buy their own servers, storage, and switching from OEMs and sometimes ODMs.)

With our spreadsheet magic, we think the “real” Microsoft systems business, including only systems-level hardware and software and tech support for it, accounted for $18.74 billion, up 25.4 percent, with operating income of $7.77 billion, up 21 percent.

Everybody wants to know how the AI business is doing, and on the call with Wall Street analysts going over the numbers, Satya Nadella, chief executive officer at Microsoft, said this:

“Model capabilities are doubling in performance every six months, thanks to multiple, compounding scaling laws. We continue to optimize and drive efficiencies across every layer – from datacenter design, to hardware and silicon, to systems software, to model optimization – all towards lowering costs and increasing performance. You see this in our supply chain, where we have reduced dock-to-lead times for new GPUs by nearly 20 percent. Across our blended fleet, where we have increased AI performance by nearly 30 percent ISO power. And, our cost per token, which has more than halved.”

Later in the call, Nadella said that Microsoft had processed over 100 trillion tokens this quarter, up 5X compared to this time last year, and that it had a record-breaking March when it chewed on and generated 50 trillion tokens.

We wonder how many of those tokens were processed by OpenAI, by Microsoft itself, and its Azure customers. It is hard to guess how that might break down, but clearly OpenAI’s training plus its API customers is one big chunk and Microsoft’s Copilot users are another big chunk. Enterprise customers training their own models on Microsoft infrastructure is probably a mere sliver at this point.