Microsoft

A runtime for agents

Every other company in this series trains a model and sells access to it. On his April earnings call, Satya Nadella described something more durable: “Windows is becoming a runtime for agents, not just a desktop for applications.” Read past the line and it’s a platform claim on the scale of the PC and the cloud combined — a plan to own the layer underneath whichever model wins, so that the question consuming everyone else (whose AI is smartest?) becomes, for Microsoft, beside the point.

Microsoft is the company I think the market has most wrong, and I say that as someone who keeps re-checking the figure because it looks too clean: roughly $38 billion in operating profit in a single quarter (FY26 Q3), audited and filed, on $82.9B of revenue with Azure growing 40% (Microsoft, CNBC). The AI debate is loud. Microsoft’s profit is quiet, and it’s real.

While the market argues about whose model is smartest, Microsoft is quietly rebuilding the software layer everything else runs on — and collecting the toll regardless of who wins.

The number everyone misreads

So how can Microsoft brag about a “$37 billion AI business” when its flagship AI app is half-ignored? Because most of that $37B isn’t Copilot. Microsoft doesn’t itemize it, but the bulk is Azure AI — other companies (OpenAI very much included) renting Microsoft’s cloud to train and run their models. Copilot — 20M seats at ~$30/month — is only on the order of ~$7B of it.

One date stamp before we go further. The $37B is the AI run-rate Microsoft last disclosed at that level of detail — FY26 Q3, up 123% year over year. The FY26 Q4 print at the end of July reported roughly $90B of quarterly revenue and a full year on the order of $332B, but did not restate the AI run-rate at the same granularity. So when you see “$37–38B of AI business” quoted anywhere — including here — read it as a Q3 figure carried forward, not a July number.

A huge share of its “AI revenue” is OpenAI and other labs paying Microsoft for the compute to run their own products. That’s also why you should never read “$37B AI business” as “$37B of Copilot.”

Two facts, kept apart

Two facts get blurred constantly, so keep them apart. Microsoft’s profit is audited and enormous. The famous “$37B AI business” is a company-defined run-rate — useful, but not an auditor-blessed “AI revenue” line. Both true; different kinds of fact.

Bought, or used? Both answers are true

Because the press treats Copilot as enterprise AI’s great success, it’s worth holding the two honest readings of it side by side — because they don’t agree, and the disagreement is the point. The trackers say a tool that’s widely bought is unevenly used. Microsoft’s own numbers say engagement is deepening fast where it lands. Both are true at once.

The tell is that Microsoft itself wired Anthropic’s Claude into its own Copilot, “to alleviate its dependency on OpenAI” — hedging its flagship product. The saving grace, per above: Copilot is the minority of the AI money. Microsoft’s franchise doesn’t live or die on it.

Build 2026: the namespace claim on agents

The real play surfaced at Build 2026 (June 2–3; conference reporting — verify any single figure against the official record before banking it). Microsoft stopped pitching AI as a feature inside Windows and started re-architecting Windows as the place agents run. Four moves, read together, are a land claim on the next platform layer.

By defining how agents register, communicate, authenticate, and persist on Windows, Microsoft is doing to the agent layer what Windows APIs did to application developers in the 1990s: setting the standard everyone else has to conform to. That, not Copilot’s seat count, is the durable bet.

Drop the dependency, own the runtime

The OpenAI relationship is unwinding by design, and from Microsoft’s chair it reads as liberation rather than loss. The April 2026 restructuring ended Microsoft’s exclusivity and the revenue share it paid OpenAI, while OpenAI keeps paying Microsoft 20% of revenue (capped at $38B) and Microsoft holds the IP license to 2032 and a ~27% equity stake worth ~$230B. Microsoft now pays OpenAI less, receives more, builds its own model (Polaris) on its own silicon (Maia), and routes Copilot across whichever model performs best. Where OpenAI experiences this as the ally walking away, Microsoft experiences it as finally being free to own the runtime instead of renting the model.

The quiet substitution

In July 2026 the hedge stopped being a plan and started being production traffic. On July 23, Satya Nadella said Microsoft is now routing tasks inside GitHub Copilot, Excel and Outlook to its own MAI models whenever those models match or beat OpenAI and Anthropic on cost and quality — with Copilot Chat and PowerPoint next in line (Times of India). This is not a lab announcement. It is a procurement decision being executed inside products hundreds of millions of people already open every morning.

What makes it credible is that Microsoft did not ship one in-house model. At Build 2026 it shipped a fleet of seven — reasoning, coding, image, speech, transcription and more (Microsoft AI). Two matter commercially: MAI-Thinking-1 for reasoning and MAI-Code-1-Flash, a small model (roughly 5B active parameters) built for GitHub Copilot and VS Code. Small, specialised, cheap to run — and good enough on the one surface it was designed for.

That two-track pattern sharpens the thesis rather than contradicting it. Microsoft never claimed the best model; it claimed the layer the models plug into. The July routing decision makes the claim falsifiable and, so far, true: the orchestration layer owns the customer, and the models are interchangeable parts bought on price and quality per surface. The moat is optionality, not loyalty.

The structural cause sits one section above. Once the April 2026 rewrite made Microsoft’s OpenAI IP license non-exclusive and freed OpenAI to serve any cloud, an in-house model fleet stopped being a nice hedge and became necessary. Every task MAI absorbs is a task Microsoft no longer buys at someone else’s list price — which matters more each quarter that inference volume grows faster than the price of inference falls.

What makes the franchise defensible

Three structural assets keep Microsoft above the model wars — and one of them, model independence, is the move every other incumbent wishes it had made earlier.

The one question Microsoft can’t dodge

Here is the crack in the strongest franchise in this set, and it’s the same force bearing down on the whole industry: the per-seat model is dying. Microsoft’s AI app is priced per logged-in human — 20 million Copilot seats at ~$30 a month. But the entire promise of agents is that the work gets done without a human in the seat, and the moment an agent resolves the ticket or drafts the deck, the seat it replaces stops being a thing you can bill for. Microsoft has the best answer in the field to “where do agents run” and the weakest answer to “what is a seat worth once the agent does the work.” Its agent-OS bet is, in part, a wager that it can move its meter from seats to agent-runtime tolls — the Agent Store, Azure consumption, Maia inference — before the seat it sells today melts. That migration is unproven, and a second risk rides alongside it: a meaningful slice of the “AI revenue” is OpenAI resale that compresses as OpenAI diversifies to AWS and GCP.

My read — the toll, not the app

This is the strongest revenue quality in AI and, I’d argue, the most underrated position in the set. The $37B disclosed run-rate sits on audited profit rather than floating free like the labs’ estimates; the $627B backlog, the in-house silicon (Maia), the model independence (Polaris), and a credible claim on the agent-OS layer make it the most structurally defensible AI franchise I track. History rhymes here — whoever builds the operating system for an era tends to win it. The risks are real — OpenAI-resale compression and the unanswered per-seat question — but the structural read is clear, and it starts with refusing the headline: read “$37B AI” for what it is — mostly the toll, not the app.

Whoever builds the operating system for an era tends to win it — and Microsoft is quietly building this one’s.

If I’m building: the binding constraint that sends me to Microsoft is enterprise reality — identity, compliance, procurement — not raw model quality, and its model-agnostic routing means I’m not wedding the integration to any one lab. I’d price Copilot internally on proven usage, never on seat count, and I’d watch the Agent Store and Azure Agent Mesh closely, because that’s where Microsoft is trying to move the meter before the seat dies under it.

Sources: FY26 Q4 results (~$90B quarter, ~$332B year), Microsoft IR; the $37B AI run-rate as last disclosed, Microsoft IR (FY26 Q3); MAI routing inside GitHub Copilot, Excel and Outlook (July 23, 2026), Times of India and product analysis; the seven-model MAI fleet, Microsoft Build 2026 and microsoft.ai/models; the partnership rewrite, Microsoft; Copilot model availability in July 2026, M365 Copilot update log; Q3 FY26 financials, Microsoft and CNBC; what the $37B actually is, Motley Fool and GeekWire; the Copilot ~$7B estimate, vaasblock analysis; Copilot usage trackers, Stackmatix; Copilot deepening engagement (Accenture 740k, 50k+ seats quadrupled, “Outlook-level”), TechCrunch; Claude-in-Copilot, Microsoft 365 blog; Project Polaris, Let’s Data Science; Windows Agent Framework / Azure Agent Mesh, byteiota and aitoolsrecap; SynthID adoption, CNBC; the OpenAI restructuring, CNBC and Bloomberg. The honest frame: Microsoft’s profit is the firmest ground in this series; the “$37B AI” is a disclosed run-rate, mostly Azure rental, not Copilot; Build 2026 product claims are conference reporting until the filings catch up.

This deep dive is part of “Deep Research on Frontier Companies.” Crafted with intent by Ravi Teja Palanki · June 2026.