Inside Microsoft's Money Machine: Who's Really Paying For The AI Bet

Microsoft's stock has turned into a running referendum on artificial intelligence. Every Azure growth number, every Copilot seat count, every capex figure gets read as a verdict on whether the AI trade still holds up. But split the company into the three pieces it actually reports to the SEC, and a quieter story shows up in the numbers. 

One part of Microsoft is still writing more than half the company's profit. A second part is spending at a pace that hasn't been matched by what it earns back yet.

Let's break down what Productivity and Business Processes, Intelligent Cloud, and More Personal Computing each actually earn, at what margin, and how much of the AI story every one of them is genuinely carrying.

Microsoft Q3 FY2026 Business Segments: Revenue, Profit and Margins

Microsoft's fiscal year ends June 30, so its most recently completed quarter, the one with fully audited numbers, is the quarter ended March 31, 2026 (Q3 FY2026), reported on April 29. That is the cleanest snapshot available of how the business actually splits, and it looks like this:

Source: Microsoft Investor Relations, Q3 FY2026 segment results (10-Q filing)

1. Productivity and Business Processes houses Microsoft 365 (the paid version of Office), LinkedIn, and Dynamics 365. 

2. Intelligent Cloud houses Azure, server products, GitHub, and Nuance's healthcare cloud tools. 

3. More Personal Computing is Windows, Surface, Xbox, and Search advertising.

Here's the detail that gets lost in the Azure headlines: revenue between the first two segments is now nearly a dead heat, $35.0 billion against $34.7 billion. But the profit gap between them is still wide. Productivity and Business Processes generated about 52% more operating income than Intelligent Cloud on almost identical revenue. That gap, and why it exists, is really the whole story.

The Landlord: Productivity And Business Processes

Think of this segment as Microsoft's landlord. It collects steady, recurring rent from a subscriber base that measures in the hundreds of millions, and because software distribution is cheap once it's built, that rent shows up as profit at a rate few businesses ever touch: 59.9% operating margin this past quarter, up from 53.2% back in fiscal 2023.

The AI wrinkle inside this segment is Microsoft 365 Copilot, the $30-per-user-per-month add-on that layers AI assistance into Word, Excel, Outlook, and Teams. Paid Copilot seats hit 20 million in the March quarter, up from 15 million just three months earlier, a jump CEO Satya Nadella called an acceleration on the earnings call. Accenture alone runs 740,000 seats, described by Microsoft as its largest single Copilot deployment, and the number of customers buying more than 50,000 seats has quadrupled over the past year.

Twenty million seats sounds big until you set it against the base it's selling into. Microsoft said in January 2026 that paid commercial Microsoft 365 seats had crossed 450 million. Even allowing for growth in that base since, Copilot's attach rate is still in the single digits, somewhere around 4 to 5%. If every one of those 20 million seats paid full list price, that would work out to roughly $7.2 billion a year (20 million times $360 a year). Real-world enterprise discounting means actual Copilot revenue is almost certainly lower than that, and Microsoft has not disclosed the real figure separately. Worth verifying the exact discounting assumption before quoting this number as fact.

The Builder: Intelligent Cloud

Intelligent Cloud is the segment carrying the AI infrastructure story, and it behaves like a construction site: expensive, fast-moving, and not yet fully leased out.

Azure and other cloud services revenue grew 40% year over year in Q3 FY2026 (39% in constant currency, which just strips out the effect of a stronger or weaker dollar). But look one line down the income statement and the picture changes. Microsoft Cloud gross margin, which blends Azure with the cloud portions of Microsoft 365, LinkedIn, and Dynamics, slipped to 66% in the March quarter, down from 67% three months before that and 69% a year earlier. Management has attributed the slide directly to rising depreciation on AI data centre hardware and growing GitHub Copilot usage, only partly offset by efficiency gains elsewhere in the cloud.

The scale of the building project explains why. Microsoft's disclosed capital spending (capex plus finance leases) has gone from roughly $28 billion in fiscal 2023 to around $65 billion in fiscal 2025, based on the company's own SEC filings. The first three quarters of fiscal 2026 alone already add up to about $104 billion. On the April 29 call, CFO Amy Hood said the company expects calendar 2026 capital spending of roughly $190 billion, with about $25 billion of that increase tied to higher component and memory prices rather than added capacity.

Against all that spending, Microsoft's own disclosed number for what it's earning back is this: "Our AI business surpassed an annual revenue run rate of $37 billion, up 123% year-over-year," the company said in its April 2026 earnings release. That is real and it is growing fast. It is also a management-defined metric whose exact composition Microsoft has not fully spelled out in its filings, so treat it as a company-disclosed estimate rather than an audited line item.

Then there is Commercial Remaining Performance Obligation, or RPO, which reached $627 billion in Q3 FY2026, up 99% year over year. RPO is simply the money customers have already contracted to spend with Microsoft but haven't been billed for yet, similar to a banquet hall that's fully booked out three years in advance. The owner can point to a packed reservation register, but the till stays empty until each wedding actually happens. A large chunk of that reservation book belongs to one customer. OpenAI represents a substantial portion of Microsoft’s contracted backlog. At the end of Q2 FY2026, Microsoft said approximately 45% of its commercial RPO balance was associated with OpenAI. In Q3, commercial RPO grew 99% including OpenAI but 26% when OpenAI was excluded. 

The Shopkeeper: More Personal Computing

The smallest and least discussed segment is Windows, Surface, Xbox, and Search advertising. Revenue actually dipped 1% year over year in the March quarter as PC and device sales softened, though Search advertising partly offset that. Its 27.8% operating margin is the lowest of the three segments. This is the shopkeeper of the business: modest, doesn't need anyone's help, and funds itself. It isn't part of either the AI growth story or the AI spending story in any meaningful way, and that's exactly why it barely gets mentioned on earnings calls.

Which Microsoft Business Segment Generates the Most Profit?

Here is the part that doesn't show up in any single quarter's headline, only when you line up several years side by side.

Source: Microsoft 10-K (FY2023-FY2025) and Q3 FY2026 10-Q, author's calculations

Read that table slowly. Intelligent Cloud's share of company revenue has climbed from 34.4% to 41.8% in three years, closing in fast on Productivity and Business Processes. But its share of company profit has barely moved, from 32.1% to 35.8%. 

Meanwhile Productivity and Business Processes have quietly kept generating well over half of Microsoft's total operating income every single year in this table, even as its share of revenue has slipped.

That gap is the cross-subsidy. Intelligent Cloud is growing revenue faster than profit because its cost base is absorbing the depreciation, power, and networking bill from the AI buildout. Productivity and Business Processes doesn't carry that same weight; its costs are mostly software distribution, support, and sales, which don't scale up nearly as fast as its subscriber revenue does. Microsoft doesn't wall off cash by segment. It runs one shared treasury. Which means the free cash flow paying for GPUs and data centres today is coming out of a pool that Productivity and Business Processes fills up disproportionately.

Call it Microsoft's landlord-builder split. The landlord collects rent every month, at a margin north of 55%, and hands a large share of that rent over to fund a builder who's still finishing the tower next door. It's a pattern familiar to a lot of Indian family businesses too. A textile trading house that's been quietly profitable for two generations often ends up bankrolling the family's newer, riskier venture, maybe a real estate project or a factory expansion, out of the shop's own cash flow. At the next family wedding, everyone wants to talk about the shiny new project. Nobody toasts the accountant keeping the old shop's ledger in the black, even though that ledger is what's actually paying the contractor.

Microsoft AI Revenue vs Capex: Is Spending Outpacing Earnings?

Put the two halves of this story next to each other and the mismatch is hard to miss.

Microsoft's AI business, by its own disclosure, runs at about $37 billion a year. Trailing twelve-month company revenue is roughly $318 billion. That puts the disclosed AI business at close to 12% of what the whole company earns, by my own rough calculation, comparing an annualised run rate against trailing revenue, so treat it as directional rather than exact.

Against that $37 billion, Microsoft has guided to roughly $190 billion of capital spending in calendar 2026, again comparing a calendar-year figure to a fiscal trailing number, so the periods don't line up perfectly. Even loosely, that is more than five times the disclosed AI revenue run rate, and works out to somewhere near 60% of trailing annual revenue being ploughed into capex in a single year, a ratio almost unheard of for a software company.

Copilot's attach rate is still in the single digits of the paid Office base. RPO grew 99% on the surface, but a large piece of that growth traces back to one customer's multi-year commitment rather than broad-based enterprise demand.

There's a further wrinkle sitting outside all three segments entirely. Microsoft holds roughly a 27% equity stake in OpenAI, valued at about $135 billion following October 2025's restructuring, with $11.8 billion of a $13 billion funding commitment actually paid in as of March 2026. Gains or losses on that stake flow through Microsoft's overall net income, not through any operating segment, and over the first nine months of fiscal 2026 that produced a net gain of about $5.9 billion, mostly a one-time accounting gain tied to the restructuring rather than a recurring profit. 

In late April 2026, the two companies also reworked their commercial terms: OpenAI keeps paying Microsoft a 20% revenue share through 2030, now capped at a fixed total, while Microsoft gave up cloud exclusivity and its own reciprocal revenue-share obligation. None of this shows up in Intelligent Cloud's operating margin. It shows up in headline net income and EPS instead, which is one more reason a single quarter's bottom line is a noisy way to judge how the AI bet is actually performing.

None of this proves the AI bet is a bad one. Infrastructure businesses take years to season, and Microsoft has the balance sheet and the patient internal funding (that landlord segment) to wait this out longer than almost anyone else in the industry. But it does mean the gap between what's being spent and what's currently being earned back is the honest starting point for thinking about this story, more useful than any single quarter's Azure growth print.

Our Take

The market is pricing Intelligent Cloud as if it already is Microsoft. The audited numbers say something more boring: Productivity and Business Processes still writes more than half the cheque, year after year, almost regardless of what's happening in Azure. That isn't a flaw in the AI story. It's the reason Microsoft can afford to run this experiment at a scale almost no other company could match, funded by a subscription business patient enough not to need the AI bet to work on any particular timeline.

Whether that patience gets rewarded depends on something nobody, including Microsoft, can honestly answer yet: whether Azure's AI consumption revenue can outrun the depreciation bill on all those data centres before the landlord's generosity starts running into its own limits. Microsoft's next quarterly results land on July 29th, 2026. Whatever that print shows, the structure underneath it, one segment funding another, isn't something a single quarter changes either way.