After years of leaning on Nvidia GPUs, Microsoft is moving more of its AI stack onto in-house chips. With the third generation of its Maia accelerator unveiled on Monday, the company indicated it would primarily count on its own silicon to process incoming requests. The Maia line began with Maia 100 in 2023 and grew with Maia 200 in January. In July, CEO Satya Nadella told analysts that Maia 200 "delivers 30% better performance per dollar than the latest generation hardware in our fleet" and added that it was running OpenAI models.
Microsoft's real AI edge is Azure, which runs a broad mix of models for customers. Microsoft 365 Copilot and GitHub Copilot already put AI in front of office workers and developers, and the company also builds its own models, even if the big-name race has featured OpenAI, Anthropic and Google. Expect some everyday speed-ups: parts of Microsoft 365 Copilot should get faster with Maia 300, according to technical fellow Jim Kleewein. And there is a cost angle too. Last month at a Goldman Sachs conference, finance chief Amy Hood said that relying on the company's own chips has produced "margin improvement over the course of a decade."
Using custom silicon for inference rather than training changes the cost structure entirely. Market Briefs covers that economics free every weekday.
Nvidia is still essential. Its GPUs power Microsoft's existing data centers and train models from OpenAI and Anthropic. As demand has surged, top model builders have added more chip options, including Amazon, Advanced Micro Devices, Cerebras, Google and, more recently, Microsoft.
For general-purpose compute, Microsoft uses its own CPUs and silicon from Intel and AMD. For training models, it partners with Nvidia and AMD, said Alistair Speirs, a general manager at the company. Nvidia's presence is not fading either: Last week, CEO Jensen Huang appeared at an event where Microsoft revealed a Surface laptop that includes an Nvidia processor.
One thing you cannot do with Microsoft's accelerators yet is rent them by the hour, unlike Google's TPUs or Amazon's Inferentia and Trainium. As Patrick Moorhead, CEO of Moor Insights and Strategy, put it, "Nobody has tested them outside of Microsoft yet."
Microsoft plans to put 96 Maia 300 chips in each rack, twice the count of Maia 200. In that setup, each rack can handle as many as seven times the AI token volume and provides 40% more tokens per megawatt than GPUs. Production of Maia 300 begins in 2027.
Executive Vice President Scott Guthrie said Microsoft is targeting multi-gigawatt deployments across the Maia family, while Speirs declined to pin down timing. "We reserve a little bit of optionality there, because this is an industry that's changing very fast, and we want to deploy the right silicon for the right models that we're using at the right time as well," he said. The bigger picture is simple: demand keeps climbing, the market wants more choice, and efficiency matters.
As Speirs put it, "We need new architectures that will give us essentially better performance, better use of the power that we have," noting that data center power use maps straight to operating costs.
If Microsoft can run more work on its own chips while keeping Nvidia in the mix for training, it could squeeze more performance per dollar and control capacity as AI use grows. Evercore ISI's call that Maia could be about 10% of AI chip spend by early 2027 signals real budget behind the shift. For everyday investors, the takeaway is not a chip debate - it is that the economics of AI at scale are moving from scarcity to smarter allocation, and that can shape how efficiently the services you use get built and priced.
How much a hyperscaler can shift off Nvidia decides its margins. Join Market Briefs free and follow the shift.
