Microsoft’s latest AI-chip push is less a clean break with Nvidia than a calculated attempt to take control of the costliest part of the AI boom: serving models at scale.

2023 to early 2026 — laying the groundwork. Microsoft introduced the first Maia accelerator in 2023, then unveiled Maia 200 in January. The company says the second-generation chip is already powering production workloads including Copilot, GitHub Copilot, Microsoft Foundry and Microsoft’s own models. Satya Nadella told analysts in July that Maia 200 delivered “30% better performance per dollar” than the latest hardware in Microsoft’s fleet.¹

October 2026 — the bigger bet. The newly announced Maia 300 is designed to become the “primary AI inference system” across Microsoft products and services, with production deployment due in 2027.² Scott Guthrie, Microsoft’s cloud chief, said a full rack of 96 accelerators could produce up to seven times Maia 200’s token throughput, while delivering “40 percent more tokens per megawatt compared to leading GPU-based systems.”²

That efficiency claim goes to the heart of Microsoft’s argument: data-center power translates directly into operating costs. The company says it will scale Maia to multi-gigawatt deployments from 2027, a commitment Nadella called Microsoft’s “biggest bet yet on our own silicon.”Satya Nadella (satyanadella)³

2027 — independence, with an asterisk. Maia is meant chiefly for inference — answering user requests — not for replacing every Nvidia system in Microsoft’s estate. Nvidia GPUs still fill Microsoft data centers and remain central to training models; Microsoft also works with AMD for that task. Alistair Speirs, a Microsoft general manager, said the company wants “the right silicon for the right models” as the market changes quickly.¹

The limitations are equally clear. Unlike Google and Amazon’s custom chips, Maia cannot be rented by outside cloud customers. Patrick Moorhead of Moor Insights and Strategy put it bluntly: “Nobody has tested them outside of Microsoft yet.”¹