Microsoft spent $145.3 billion on capital expenses in fiscal 2026, roughly two-thirds of it on CPU and GPU clusters, and Azure broke through $100 billion in annual revenue for the first time. The AI business, for all the noise around it, still accounts for 12.4 percent of quarterly revenue. The conversation is dominated by AI. The income statement is not.

The company closed fiscal Q4 2026 with total revenue just over $90 billion, up 17.7 percent, operating income of $40.6 billion, up 18.3 percent, and net income of $35.77 billion, up 31.3 percent. Wall Street approved. Microsoft posted its best week ever and now carries a market capitalization of $3.44 trillion.

Where does the money actually come from?

Not from AI, mostly. The Productivity and Business Processes group, which sells application software for PCs and servers, booked $37.85 billion in Q4 sales, up 14.3 percent, with $21.9 billion in operating income — a 57.9 percent margin. The Intelligent Cloud segment, spanning Windows Server and Azure, took in $39.31 billion, up 31.6 percent, though heavy infrastructure costs held operating income to just under $16 billion, or 40.6 percent of revenue.

More Personal Computing, the PC software and devices group, declined 4.4 percent to $12.85 billion with a thin $2.75 billion in operating income. Services revenue drove the quarter: $72.77 billion, up 22.7 percent, against product sales of $17.23 billion, up a mere six-tenths of a point.

The broader Microsoft Cloud category — any PC or server software sold with utility pricing — generated $59.3 billion, up 27 percent, with gross profit of $38.55 billion.

How big is Azure, really?

By the analyst model at The Next Platform, Azure infrastructure revenue reached $31.91 billion in Q4, up 64 percent, with operating income of about $13.93 billion, up 68.8 percent. That is a 43.6 percent margin on what is, at bottom, a hardware business. OEM and ODM server makers can only dream of such economics.

Strip out the cloud utility and look at the "real" systems business — baseline Azure IaaS plus on-premises Windows Server — and the model puts it at just over $22 billion in sales, up three-tenths of a point, with $8.94 billion in operating income. That rivals IBM at its peak in X86 servers more than a decade ago. Few systems businesses in history have been larger.

What about the OpenAI backlog?

The backlog stands at $678 billion as Q4 closed, comprised mostly of Azure contracts stretching into the future. Microsoft did not break out OpenAI separately this quarter, but the company said the non-OpenAI portion grew 25 percent. Working backward through historical data, OpenAI still holds $255.1 billion of contracted capacity, leaving $422.9 billion from other customers.

The AI annualized run rate, estimated by The Next Platform since Microsoft did not provide one, sits at roughly $44.5 billion — a little more than double the year-ago figure. That implies actual Q4 AI revenue of $11.13 billion. Two years ago, the AI business was five times smaller.

Is the spending sustainable?

Microsoft says yes, and says it with restraint. For fiscal 2027 the company plans to stand pat at $190 billion in capital expenses, declining to match the spending escalation of its hyperscale peers. About a third of fiscal 2026 capex went to datacenter facilities and other long-lived assets; Q4 alone saw $41 billion spent, and Microsoft exited the quarter with $76.84 billion in cash and equivalents.

Copilot adoption adds a second, quieter signal: 30 million seats, up from 20 million only three months earlier. Azure cloud revenue keeps rising on the same track. In a market where only the rich can steer AI forward, because state-of-the-art models demand so much expensive hardware, Microsoft is big enough to control its own fate.

The verdict for semiconductor watchers: a $190 billion capex budget held flat is still an enormous procurement floor for CPU and GPU suppliers, and the $678 billion backlog suggests the silicon underneath it is already spoken for. The day AI dominates Microsoft's business will come, as it did for cloud computing — but not yet.