Microsoft plans to make Maia its primary AI architecture — CNBC
Microsoft plans to increasingly use its own Maia artificial intelligence chips and make them the primary architecture for processing incoming requests to AI models. This is expected to begin with the third-generation Maia accelerator, CNBC reports.
Currently, the company’s data centers are largely equipped with Nvidia graphics processing units, which are used for AI workloads and training models from OpenAI and Anthropic. At the same time, Microsoft already uses its own central processing units, as well as Intel and AMD chips, for general-purpose tasks.
Maia 300 production will begin in 2027
Microsoft introduced its first Maia 100 AI chip in 2023 and Maia 200 in January this year. The company’s CEO Satya Nadella said in July that Maia 200 delivers 30% higher performance per dollar compared with the latest generation of equipment in Microsoft’s fleet. According to him, this chip was already processing OpenAI models.
Production of Maia 300 is set to begin in 2027. Microsoft Executive Vice President Scott Guthrie said in a blog post that the company expects to deploy Maia-family chips with a combined capacity of several gigawatts. At the same time, Microsoft General Manager Alistair Speirs did not name a timeframe for such deployment, citing rapid changes in the industry and the need to match chips to specific models.
More current news is available on the UA.News Telegram channel Telegram.
More chips in a server rack
Microsoft expects to place 96 Maia 300 chips in each data-center server rack, twice as many as Maia 200. According to the company, the new version will be able to process up to seven times more AI tokens per rack and 40% more tokens per megawatt than graphics processing units.
Some Microsoft 365 Copilot features are expected to run faster thanks to Maia 300. Evercore ISI analysts estimate that by the first quarter of 2027, Maia will account for about 10% of Microsoft’s spending on AI chips. Nvidia will nevertheless remain an important Microsoft partner, particularly for model training; the company also works with AMD in this area.
Unlike Google’s tensor processors and Amazon’s Inferentia and Trainium chips, Microsoft’s accelerators are not available to other companies for hourly rental.
