cloud
Microsoft built a systems software platform – Windows Server and its zillions of add-ons and extensions – long before it built a cloud and became a system supplier to millions of customers around the world. With tens of millions of Windows Server customers around the world, it was nearly a given that the company could build a cloud that could rival the likes of Amazon Web Services.
And despite some early difficulties, Microsoft has done just that as well as contributing to the font of knowledge about the design of systems, datacenters, and myriad system management, data services, and frameworks, and middleware software that helps it run all kinds of applications at hyperscale.
But it was not a foregone conclusion that Microsoft would be a player in the GenAI era, which is the main reason why the company so tightly coupled itself to OpenAI then the chattybot moment hit in November 2021 with the API release of the GPT-3 foundation model. Microsoft had some success with its Turing LLM and then a hybrid Megatron-Turing model created in conjunction with Nvidia, but a little more than a year before that chattybot moment, back in September 2020 Microsoft had inked a deal with OpenAI for an exclusive license to OpenAI models (including access to source code) in exchange for big investments and what we presume were access to GPU clusters on its Azure cloud at a reasonably discounted price.
As we have previously reported, Microsoft and OpenAI are still working together, but that tight coupling has been unwinding for the past year, and now Microsoft will no longer have exclusive rights to OpenAI GPT models. OpenAI has already been commissioning tens of billions of dollars to get its own hardware, working with Nvidia, AMD, and Cerebras Systems directly as well as a bunch of other clouds and neoclouds while also honoring its substantial commitments to rent compute engine capacity on Azure.
All of this means that Microsoft can relax a little, add other models to its Azure stack and work to bring more and more AI functions to its vast application and system software base as well as extend its Azure platform to be a place where companies who can’t get access to GPUs and XPUs or who never want to own such expensive iron can train their GenAI models and run their inference workloads. And if OpenAI implodes, Microsoft might catch some flack when a big portion of its revenue backlog disappears, but it won’t be taken out by what used to be utter inter-dependence between itself and OpenAI when it comes to capitalizing on the GenAI boom.
Personally, I fully expect for Microsoft to get back into being a frontier model builder and compete with Google, OpenAI, Anthropic, and AWS (which has its own Nova models). Hyperscalers and cloud builders want to control their entire stack – as does Nvidia, which can afford to open source its Nemotron models because it is so rich on hardware. (But trust me, when you buy Nvidia GPUs, you are paying for all that free software.) Microsoft is rich enough to be an AI model builder, and it is even rich enough to open source some or all of the models if it thinks that will help make the models better or more ubiquitous. I am not advocating for that any more than I would for Microsoft to give away free Maia AI XPUs.
It is against all of that context that we turn to Microsoft’s most recent financial results and see how it is doing in its AI business journey, particularly on the Azure cloud.
The thing about the Azure cloud is that it has scale. Microsoft has more than 80 Azure regions around the globe, with more than 500 datacenters across those regions plus an additional 190 plus point of presence (PoP) locations and interconnectivity of over 800,000 kilometers of network fiber. The whole shebang probably burns somewhere on the order of 10 gigawatts of juice and growing fast.
On the call with Wall Street analysts going over the numbers for its third quarter for fiscal 2026, Satya Nadella, Microsoft’s chief executive officer and the architect of its Azure expansion and GenAI strategy, said that the company added 1 gigawatt of additional capacity in the March quarter, and reiterated that the company was on track to double its capacity (as expressed in gigawatts, not flops) in two years.
And they say that Moore’s Law is dead. . . .
Nadella added that the company’s Cobalt homegrown Arm server CPUs are now deployed in half of its regions and that the “Braga” Maia 200 homegrown XPU, [revealed
back at the end of January](https://www.nextplatform.com/ai/2026/01/28/microsoft-takes-on-other-clouds-with-braga-maia-200-ai-compute-engines/4092134), are live in its datacenters in Iowa and Arizona. In the quarter, Microsoft had $82.89 billion in sales, up 18.3 percent year in year and up 2 percent sequentially. Operating income rose by 20 percent to $38.4 billion and net income was up 23.1 percent to $31.78 billion, representing 38.3 percent of revenues. Considering how much hardware Microsoft builds for Azure, this is a pretty good margin.
Speaking of which, Microsoft spent $31.9 billion on capital expenses in the March quarter, and exited the thirteen weeks with $78.27 billion in the bank so it can keep spending like a slightly intoxicated sailor. (No the drunken ones, Oracle and OpenAI.) About two thirds of that covered the cost for CPUs and GPUs, which Microsoft calls “short-lived assets.” The rest was for datacenters that have a lifespan of fifteen years or more.
The Intelligent Cloud group, which includes the Windows Server platform as well as the Azure cloud, had $34.68 billion in sales, up 29.6 percent year on year and up 5.4 percent sequentially. This is on par with the growth that [Amazon
Web Services showed in the same period](https://www.nextplatform.com/cloud/2026/04/30/aws-will-be-an-oem-just-like-google-and-maybe-microsoft/5219100), but half that [seen
by Google Cloud in the March quarter](https://www.nextplatform.com/cloud/2026/05/02/google-is-a-full-stack-ai-player-and-is-playing-well/5219190). Operating income for Intelligent
Cloud was $13.75 billion, up 24 percent and representing 39.7 percent of sales.
Note: In the chart above, it is immediately obvious where Microsoft peeled out cloud-ish revenues that it had in its PC and Intelligent Cloud businesses and added them into its application software businesses. (Perhaps things were not as cloudy as it had been saying, which is odd for Seattle. . . . )
The More Personal Computing group is the smallest and least profitable of the groups, with $13.19 billion in sales (down 1.3 percent) and an operating income of $3.67 billion (up 4.1 percent and only 27.8 percent of revenues). The Productivity and Business Processes group is the most profitable, as you might expect, with just a tad over $35 billion in sales and just under $21 billion in operating profits. But the enterprise app business only grew by 16.9 percent in the quarter.
The business segment the company calls Microsoft Cloud, which means the Azure infrastructure plus any cloud subscription for any of the company’s software as opposed to perpetual licenses and tech support, had $54.55 billion in sales, up 28.5 percent, with a gross profit (not an operating profit) of $36 billion, up 22.9 percent.
Microsoft does not split out Azure as a separate division, which is annoying but also understandable in the hybrid cloud world that Microsoft, perhaps more than any other of the big cloud builders, helped create. My best estimate is that Azure revenue was just north of $23 billion, about two-thirds of the Intelligent Cloud sales, and operating income was $9.6 billion. Both grew at around 39 percent.
As you know, I like to extract how the core, real systems business – infrastructure plus operating systems and tech support and financing but not including databases, development tools, and application software – does for all of the datacenter companies I watch on your behalf. Here’s what it looks like:
There is admittedly a lot of witchcraft and assumptions in this data, which is a hybrid mix of on premises Windows Server and on cloud Azure capacity minus applications and data stores. I am not sure what is causing the downward drift other than possibly companies being tight with the purse strings on Windows Server platforms as they try to make room in their budgets for GenAI.
In the chat with Wall Street, Microsoft dropped a tidbit about the annualized run rate for its “AI business,” and we are not sure how the company is drawing lines here to come up with this number, particularly as it relates to Copilot and other GenAI model access from within its applications. But in Q3 F2026, the AI business had an ARR of $37 billion, up 123 percent year on year. That gives us two data points. The company also gave a datapoint on this ARR back in Q2 F2025, saying it was $13 billion, up by a factor of 2.75X year on year.
This chart above shows those four datapoints in solid red and fills in the gaps for the missing quarters to give you a probably shape of the AI business ARR growth curve over time.
What is immediately obvious is that Microsoft’s capex for any given trailing twelve months is considerably larger than the AI business ARR at any given time. You have to invest in the future if you hope to reach it.
Which brings us to the Microsoft revenue backlog, which stood at $627 billion as the March quarter ended, just shy of double what it was a year ago.
Microsoft did not address the OpenAI backlog, which was bolstered by a $250 billion deal signed by the two companies last summer. But Microsoft did say that the non-OpenAI backlog grew by 26 percent, which means it hit $361.4 billion, which grew by 26 percent year on year and 5.1 percent sequentially.
If you do the math, that means the OpenAI backlog is now $265.6 billion, down 5.6 percent sequentially. That math implies that OpenAI spent $15.65 billion on Azure capacity in Q3 F2026 – but I would be careful jumping to that conclusion. I have a tough time believing that OpenAI represented 55 percent of Intelligent Cloud revenues, and an even tougher time believing that OpenAI represented 68 percent of all Azure revenues. It is hard to believe this is true, even if the math is clean. What is clear is that to service that OpenAI commitment, Microsoft is going to have to invest in AI infrastructure at the crazy rates of its peers: AWS, Google, and Meta Platforms. Amy Hood, Microsoft’s chief financial officer, said that Microsoft raised its calendar 2026 capex spending plan to $190 billion, and just to show us how bad the price increases on DRAM, flash, wafers, and substrates are, she added that $25 billion of that was to cover cost increases in those areas. With capex spending expected to be around $40 billion in Q4 F2026, that means for the remaining two quarters of calendar 2026 combined Microsoft will shell out around $120 billion. She added that in calendar 2027, capex spending would “significantly increase.”
GenAI is clearly a rich company’s game.