Microsoft's Sept. 2 8-K creates two new fiscal-year 2027 segments: Agents and Infrastructure, and Devices and Consumer. The filing restates fourth-quarter Azure revenue at US$29.42 billion and guides first-quarter Agents and Infrastructure revenue to US$75.15 billion to US$75.75 billion. Microsoft also expects Azure constant-currency growth of 44% to 45%; the midpoint of the new segment range is US$75.45 billion, with a US$0.60 billion width, or roughly plus or minus 0.40% around that midpoint.
The numbers change the reporting lens. Agents and Infrastructure combines Azure infrastructure with Microsoft 365, GitHub and agent products. Devices and Consumer contains Windows, Xbox, search and advertising. The filing is a reporting architecture change, not a claim that token billing has been separated from every other workload.
A new map for an old business
Azure was previously reported inside Intelligent Cloud. Restating the fourth quarter gives analysts a bridge into the new presentation, not a historical series for every agent, model or inference workload.
Customers may purchase Azure capacity, Microsoft 365 seats, GitHub services and an agent through overlapping agreements. Revenue can arrive in one segment while compute, networking and support costs sit across teams. Treating the label as a pure AI bucket overstates what the filing proves.
Reporting architecture is not billing architecture
The source pack contains no verified 65% token-mix figure. Microsoft customers use multiple models, and capacity remains constrained. An agent that calls a model repeatedly can raise consumption, yet the economics depend on model routing, reserved capacity, storage, network egress and support. Seat revenue can mask variable compute; a usage surge can do the opposite.
For investors, the test is whether the segment shows consumption that is measurable, collectible and profitable. For customers, invoices must expose prompts, retrieval, tools and data movement well enough to govern usage. The filing does not disclose token or agent-run unit economics.
The cost-allocation test
The guidance range is narrow enough to be operationally relevant. A US$0.60 billion range around US$75.45 billion is about 0.8% of the full span and 0.4% from midpoint. A miss of that scale could come from demand timing, foreign exchange, capacity availability or an accounting allocation; the filing alone cannot identify the cause.
The next layer is margin. AI infrastructure requires accelerators, networking, data-center power and depreciation before an agent produces cash flow. Fixed-price compute can pressure gross margin; usage pass-through can scale revenue with cost. Analysts still need Microsoft's cost base and allocation between cloud infrastructure and products.
What technology teams should measure
Buyers should reconcile Azure invoices with model calls, retrieval volume, tool execution, storage and egress. They should record which workloads can move between models and which depend on Microsoft's identity, data or development services. That evidence shows portability and concentration without assuming that a branded agent is a separate economic product.
Procurement teams can use the new segment disclosure as a prompt to request usage ceilings, rate cards, capacity commitments and renewal economics. Engineering leaders should watch latency and throttling alongside spend: constrained capacity can make a cheap token unusable for a production workflow.
The investor test is in the next filing
Microsoft shares were shown at US$496.82, down 0.90%, in the frozen market snapshot; the quote has no event-study attribution. The next 10-Qs should show whether the new segment produces clearer Azure usage, AI infrastructure cost, agent revenue and margin commentary.
The Sept. 2 filing changes what Microsoft can show investors. It does not prove a token mix, a separate AI profit pool or a billing revolution. The durable test is whether the new meter links customer consumption to capacity cost and cash conversion.