4 hours ago
23
Microsoft’s custom AI chip ambitions are back on track. The company’s Maia accelerator program, which hit turbulence with design changes and staffing problems, is now producing real hardware and eyeing an expanded lineup that could reshape how Azure competes for AI workloads.
The Maia 200 inference chip, built on TSMC’s 3nm process, has already begun deploying in Microsoft data centers in Iowa and Arizona. And the company is already designing its successor, the Maia 300, positioning the series as a genuine alternative to Nvidia’s dominance in AI compute.
What Maia 200 actually brings to the table
The Maia 200 ships with 216GB of HBM3e memory and native support for FP8 and FP4 precision formats.
Microsoft claims the chip delivers three times the FP4 performance of Amazon’s Trainium3. It also says FP8 performance exceeds what Google’s TPUv7 can do.
Microsoft says the Maia 200 offers roughly 30% better performance-per-dollar compared to the previous generation of Azure fleet silicon.
The chip is already powering internal applications. Microsoft 365 Copilot and OpenAI’s models are running on Maia 200 hardware in those Iowa and Arizona deployments.
A rocky road to revival
The Maia program first surfaced in November 2023, when Microsoft announced it alongside the Cobalt CPU. Design changes forced timeline shifts, and staffing challenges compounded the delays. Mass production of the Maia 200 was pushed to 2026, well behind the pace that would have matched the original ambitions.
The Maia 200 was formally announced on January 26, 2026, and deployment followed almost immediately.
Maia 300 and the bigger picture
The successor chip, Maia 300, is already in design. The development timeline stretches into 2027.
Discussions in mid-2026 indicated potential collaborations around external use of Maia 200 hardware, with Anthropic mentioned as a possible partner.
Microsoft executives have been vocal about the economics. Both Scott Guthrie and Satya Nadella have emphasized how custom chips improve Azure AI’s cost structure.
Why the chip wars keep escalating
Microsoft’s push into custom silicon follows a pattern that every major cloud provider is now pursuing. Amazon has Trainium and Inferentia. Google has TPUs through version seven. Even Meta has invested heavily in custom AI training infrastructure.
The common motivation is reducing dependence on Nvidia, which currently commands the vast majority of the AI accelerator market.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·