A Microsoft Azure Maia 300 prototype chip.
Microsoft
Microsoft will accelerate the use of its local artificial intelligence chips after years of dependence on from Nvidia graphics processing units, as the software giant attempts to strengthen its AI stack.
Starting with its third-generation Maia AI accelerator, unveiled Monday, Microsoft said it will rely primarily on its own silicon to handle incoming requests.
Microsoft’s biggest strength in AI is its Azure cloud, which runs a series of models for customers. Its Microsoft 365 Copilot and GitHub Copilot offer AI-based services to software engineers and other office workers. And the company develops models in-house, even if they have not yet made a splash in a market dominated by OpenAI, Anthropic And Google.
The emerging chip opportunity stems from a relentless demand for advanced processors and a desire for choice in the market. Nvidia’s GPUs are the standard for AI workloads and models, and they fill Microsoft’s existing data centers, while also being used to train models from OpenAI and Anthropic.
But as demand continues to grow, model leaders have adopted a variety of silicon from companies including Amazon, Advanced microdevices, CerebralGoogle and, in recent months, Microsoft.
For general use, Microsoft already leverages its own central processing units as well as silicon from Intel And AMD. Chief Financial Officer Amy Hood told a Goldman Sachs conference last month that using its own chips had contributed to “margin improvement over the course of a decade.”
Microsoft introduced its first AI chip, the Maia 100, in 2023. It touted the Maia 200 last January, and CEO Satya Nadella told analysts in July that Maia 200 “delivers 30% better performance per dollar than the last generation hardware in our fleet” and said it runs OpenAI models.
Production of the Maia 300 begins in 2027. Microsoft is targeting multi-gigawatt chip deployments in the Maia family, according to a blog post from Executive Vice President Scott Guthrie. But Alistair Speirs, chief executive, declined to provide a timetable.
“We reserve a bit of optionality because this is a very fast-moving industry and we also want to deploy the right silicon for the right models that we use at the right time,” Speirs said.
Microsoft plans to fit 96 Maia 300 chips into each data center rack, double the number in the Maia 200. The new version can handle up to seven times more AI tokens per rack, with 40% more tokens per megawatt compared to GPUs.
“We need new architectures that will essentially give us better performance and better use of the power that we have,” Speirs said. “You can also imagine that power translates pretty directly into operational costs for something like a data center. »
Parts of Microsoft 365 Copilot should speed up with Maia 300, said Jim Kleewein, the company’s chief technology officer.
In a note distributed to clients Wednesday, Evercore ISI analysts estimate that Maia will account for about 10% of Microsoft’s spending on AI chips by the first quarter of 2027. The analysts recommend buying Microsoft stock.
Nvidia will still play an important role in the equation. CEO Jensen Huang spoke at an event last week where Microsoft announced a Surface laptop packing an Nvidia processor. And when it comes to training models, Microsoft is partnering with Nvidia and AMD, Speirs 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 has tested them yet outside of Microsoft,” said Patrick Moorhead, CEO of consulting firm Moor Insights and Strategy.
WATCH: Melissa Otto: Nvidia is the AI business barometer
Gn bussni
