Microsoft Azure Maia 300 prototype chip.
microsoft
microsoft Plans to step up use of in-house artificial intelligence chips that it has relied on for years Nvidia’s Graphics processing units are being beefed up as the software giant looks to beef up its AI stack.
Microsoft said it will primarily rely on its own silicon to process incoming requests, starting with the third-generation Maia AI accelerator announced Monday.
Microsoft’s biggest strength in AI is the Azure cloud, where it runs a set of models for its customers. The company’s Microsoft 365 Copilot and GitHub Copilot provide AI-powered services to software engineers and other office workers. Although the company has developed its own models, they have not yet gained much traction in a market dominated by: OpenAI, human and google.
New opportunities in chips are emerging from the continued demand for advanced processors and the desire for choice in the marketplace. Nvidia’s GPUs have become the standard for AI workloads and models, filling Microsoft’s existing data centers while also training models from OpenAI and Anthropic.
However, as demand continues to grow, model leaders have adopted a variety of silicones from companies such as: Amazon, advanced micro device, cerebrumGoogle, and in recent months, Microsoft as well.
For general-purpose applications, Microsoft already relies on its own central processing units and silicon. intel and AMD. Finance chief Amy Hood said at a Goldman Sachs conference last month that the use of in-house chips has contributed to “a decade of improved margins.”
Microsoft introduced its first AI chip, the Maia 100, in 2023. The company touted the Maia 200 in January, and CEO Satya Nadella told analysts in July that the Maia 200 offers “30% better performance per dollar than our latest generation hardware” and works with OpenAI models.
Production of the Maia 300 will begin in 2027. According to a blog post by Executive Vice President Scott Guthrie, Microsoft is targeting multi-gigawatt deployments of the Maia family of chips. But general manager Alistair Spears refused to provide a schedule.
“This industry is changing so rapidly that we want to reserve a little bit of options there and bring the right silicon to the right model that we’re using at the right time,” Spears said.
Microsoft plans to install 96 Maia 300 chips in each data center rack, twice as many as Maia 200. The new version can run up to 7x more AI tokens per rack, delivering 40% more tokens per megawatt when compared to GPUs.
“We need a new architecture that inherently provides better performance and allows us to better leverage the capabilities we have,” Spears said. “Even in data centers, you can think of power as a direct reflection of operating costs.”
“Parts of Microsoft 365 Copilot should be accelerated by Maia 300,” said Jim Kleewein, a technical fellow at the company.
Evercore ISI analysts estimate in a note distributed to clients Wednesday that Mya will account for about 10% of Microsoft’s AI chip spending by the first quarter of 2027. Analysts recommend buying Microsoft stock.
Nvidia will continue to be a big part of the equation. CEO Jensen Huang spoke last week at an event where Microsoft announced Surface laptops powered by Nvidia processors. Microsoft also partners with Nvidia and AMD when it comes to training models, Spiers said.
Unlike Google’s tensor processing units or Amazon’s Inferentia and Trainium, Microsoft’s accelerators cannot be rented by the hour to other companies.
“No one outside of Microsoft has tested it yet,” said Patrick Moorehead, CEO of advisory firm Moore Insights and Strategy.
Featured: Melissa Otto: Nvidia is the standard bearer in the AI industry
