Story
Oct 13, 2026
Microsoft’s Maia 300 Escalates the Race to Loosen Nvidia’s Grip
Microsoft is betting its Maia 300 chip can make AI inference cheaper and more efficient from 2027, while keeping Nvidia central to model training. The move underscores how Nvidia’s biggest buyers are becoming chip rivals without yet breaking free of its ecosystem.
Microsoft’s push toward homegrown AI chips has been years in the making. It introduced Maia 100 in 2023, then brought Maia 200 into production workloads including Copilot, GitHub Copilot and its own models this year. The company says the earlier chip is already its best-performing AI infrastructure on real-world service measures.1
On Oct. 12, Microsoft unveiled Maia 300, its third-generation accelerator, framing the chip as a response to a costly new phase of AI: long-context reasoning, fast conversational systems and multi-step agents. Scott Guthrie, the company’s cloud chief, said a 96-chip configuration could produce up to seven times Maia 200’s rack-scale token throughput, while delivering 40% more tokens per megawatt than leading GPU-based systems.1
Satya Nadella called it Microsoft’s “biggest bet yet on our own silicon,” saying the Maia family would scale to multi-gigawatt deployments from 2027 and become the primary inference system for Microsoft’s own products and services.
2 The distinction matters: Microsoft’s chips are not currently available for outside companies to rent through Azure, and the commitment does not automatically extend to customers such as OpenAI.3
The strategy is less a declaration of independence from Nvidia than a bid for leverage. Microsoft expects Maia 300 production deployments to begin in 2027, but executives have stressed that the company wants the flexibility to match the right silicon to each model. Nvidia and AMD will remain partners for training workloads, where Nvidia remains dominant.4
That split captures the wider AI-chip contest. Google, Amazon, OpenAI, Meta and Anthropic are also pursuing or examining alternatives, largely targeting inference, where operating costs and margins are most exposed. Yet Nvidia’s revenue more than doubled to $96 billion in the second quarter, and custom-chip builders still depend on the same scarce manufacturing, packaging and high-bandwidth-memory supply chain.4
Microsoft is therefore trying to capture more of AI’s economics without betting that Nvidia’s central role disappears anytime soon.