Microsoft’s Maia chip strategy is moving toward a bigger role in its AI services. Unveiled on October 12, 2026, Maia 300 is the company’s third-generation in-house AI accelerator, with production workloads scheduled to begin in 2027 and a plan to make its own silicon the main system for AI inference across Microsoft products.
Key takeaways
- Microsoft expects Maia 300 to produce 40% more tokens per megawatt than leading GPU systems.
- A 96-accelerator rack targets up to seven times Maia 200’s token throughput.
- Microsoft will continue working with Nvidia and AMD on model training.
According to The Verge, the shift is intended to reduce Microsoft’s reliance on Nvidia and AMD GPUs while improving the cost and efficiency of cloud AI workloads. The chip also targets more demanding reasoning models and agentic tasks.
Microsoft’s Maia chip targets reasoning and agentic workloads
Maia 300 is built for long-context reasoning, delay-sensitive conversational experiences, and sustained tasks involving multiple agent steps. Scott Guthrie, executive vice president of Microsoft’s cloud group, describes it as a complete AI system for those workloads.
The accelerator follows Maia 200, introduced in January 2026. That generation already handles production workloads for Microsoft Copilot, GitHub Copilot and Microsoft Foundry, alongside Microsoft’s own AI models. CNBC reported that Maia 200 also runs OpenAI models; in July, CEO Satya Nadella said it delivered 30% better performance per dollar than the latest-generation hardware in Microsoft’s fleet.
More tokens from each megawatt and rack
Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, Guthrie said. In a full configuration of 96 accelerators, he said, rack-level token throughput reaches up to seven times that of Maia 200.
CNBC reported that the planned rack contains twice as many accelerators as its predecessor. Microsoft general manager Alistair Speirs linked better use of available power to data-center operating costs, while technical fellow Jim Kleewein said parts of Microsoft 365 Copilot should become faster with Maia 300.
The design uses a chiplet architecture built on TSMC’s N2P process, separating compute, memory-controller and I/O functions. Microsoft’s Azure hardware and infrastructure team said that separation dedicates more silicon area to computation and lets each function use technology suited to its role.
Production in 2027, with Nvidia still involved
Maia 300 will begin serving production AI workloads in 2027. Guthrie’s stated plan is to expand the Azure Maia family into multi-gigawatt deployments starting that year and make it Microsoft’s primary inference system.
That shift does not extend to all AI computing. Speirs told CNBC that Microsoft continues partnering with Nvidia and AMD for model training. He also emphasized flexibility in choosing silicon for particular models.
Evercore ISI analysts estimated in a client note that Maia would account for about 10% of Microsoft’s AI-chip spending by the first quarter of 2027, CNBC reported. Microsoft’s accelerators are not available for other companies to rent by the hour.
Co-designing Microsoft’s models and silicon
Microsoft AI CEO Mustafa Suleyman identifies model-and-chip co-design as a priority for the Microsoft Maia chip strategy. The report connects that work to the company’s effort to catch up with Google DeepMind, OpenAI and Anthropic.
“We have to work across the full stack, all the way down to the silicon,” Suleyman said. “At that scale we need a platform that we own, and that we co-design with our models. That’s why running our Microsoft AI models on top of the Maia chip is a critical focus for our team.”
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.