Microsoft AI Strategy: Maia 300, Agent 365 and Claude Code Cuts

Microsoft is trying to own more of the artificial intelligence stack, from the silicon that runs its models to the tools engineers use to write code. Three recent developments show this: a new in-house inference chip, a consolidated agent-building toolkit with a governance layer on top, and a pullback from a rival's coding assistant. Taken together, they suggest a company putting cost and control ahead of openness, even while it markets itself as model-agnostic.

Maia 300: betting on its own silicon

Microsoft unveiled Maia 300, its third-generation AI accelerator, on October 12, with production deployments planned for 2027. 478 The chip is built mainly for inference, and Microsoft says it delivers 40% more tokens per unit of power than leading GPU-based systems. 45 Microsoft wants it to become the main system answering requests for products like Copilot and Teams. 48

Microsoft's published rack configuration puts 96 accelerators in a roughly 167-kilowatt envelope. 57 Analytics Insight compares that with Nvidia's GB300 racks, which hold 72 GPUs at about 142 kilowatts. 7 Android Headlines adds a claim of seven times the token output of the prior Maia 200. 5 IBTimes cautions that the gain is more capacity within a fixed electricity budget, not necessarily faster individual responses. 8

The details vary by outlet. Coverage based on Microsoft's own documents cites 288 GB of HBM4 memory. 4 Earlier leaked reports described 216 GB of HBM3e on a 3-nanometer TSMC process. 6 Analytics Insight instead describes a chiplet design, with compute on TSMC's 2-nanometer node and data-movement parts on 3-nanometer. 7 Volume estimates also differ. August reports had Microsoft seeking TSMC capacity for more than 300,000 chips. 6 Post-announcement coverage cites plans for over a million units in 2027. 7 Where they conflict, the official October figures should be treated as authoritative, and the pre-launch leaks as superseded.

The broader point is less disputed. With capital spending reportedly expected to top $175 billion in 2026, cutting reliance on expensive third-party GPUs is a financial necessity as much as a technical one. 57

Agents: a unified framework, a still-crowded stack

On the software side, Microsoft shipped version 1.0 of its Agent Framework on April 3. 2 It merges two previously incompatible open-source SDKs, AutoGen and Semantic Kernel, into one toolkit. 23 Atlan notes that Semantic Kernel v1.x will get critical fixes for at least a year. 3 The framework also supports interoperability standards including MCP, A2A, AG-UI, and OpenAPI. 3

Governance comes through Agent 365. It gives each agent an Entra Agent ID, a registry entry, and tenant-wide observability, and it extends Defender and Purview controls to non-human identities. 1 It costs $15 per user per month on its own, or comes bundled in the $99 Microsoft 365 E7 tier. 1 The license does not cover agent runtime costs, and unregistered external agents can still escape visibility. 1

Critics argue the consolidation is incomplete. Forbes contends that Google Cloud and AWS offer clearer default paths from framework to managed runtime. In its view, Microsoft's agent story still spans too many surfaces across build, deployment, governance, and distribution. 2 Atlan makes a related point: orchestration is solved inside the SDK, but supplying agents with governed business context is not, and that gap determines accuracy in production. 3

Claude Code cuts: cost control and self-preference

The most revealing move may be internal. In May 2026, Microsoft began cancelling most Claude Code licenses in its Experiences + Devices division, which covers Windows, Microsoft 365, Outlook, Teams, and Surface. 11 Engineers were told to move to GitHub Copilot CLI by June 30, the end of Microsoft's fiscal year, under direction from executive vice president Rajesh Jha. 911 Byteiota, citing reporting from early October, adds that per-employee AI budgets were cut by 90%, and that Meta halved its own Claude Code users to 30,000. 9

The framing differs by outlet. AI Weekly presents the cuts as evidence that enterprise AI returns are failing. It cites one client that reportedly burned $500 million in a month without usage controls. 10 Byteiota emphasizes competitive logic: Microsoft owns Copilot and partners closely with OpenAI. 9 Space Daily points to Uber reportedly exhausting its 2026 AI-tools budget in four months. 11

There is a tension here. Superblocks notes that GitHub Copilot's model picker runs Claude, GPT-5.x, and Gemini side by side, and that Microsoft 365 Copilot enabled Claude by default in January. Yet it describes the license cuts as Microsoft pushing customers back onto Copilot. 12 Microsoft welcomes rival models inside its own interfaces but resists rival interfaces.

The reading

The pattern is consistent across all three areas. Maia 300 lowers dependence on Nvidia. Agent 365 makes Microsoft the identity and compliance layer for AI agents. Steering engineers to Copilot keeps spending in-house. The risk is fragmentation, as Forbes warns. 2 Still, the direction is clear: Microsoft wants to supply the AI, the hardware it runs on, and the rules governing it.