Microsoft today announces the successor to its first proprietary AI chip, the Maia 200. Built on TSMC’s 3nm process, Microsoft claims its Maia 200 AI accelerator “delivers 3x faster fourth-training performance compared to third-generation Amazon Trainium and eighth-training performance superior to Google’s seventh-generation TPU.”
Each Maia 200 chip has over 100 billion transistors, designed to handle large-scale AI workloads. “Maia 200 can easily support today’s largest models, leaving plenty of room for even larger models in the future,” says Scott Guthrie, executive vice president of Microsoft’s Cloud and AI division.
Microsoft will exploit Maia 200 to host OpenAI’s GPT-5.2 model and others for Microsoft Foundry and Microsoft 365 Copilot. “Maia 200 is also the most efficient inference system Microsoft has ever deployed, delivering 30 percent better performance per dollar compared to the next-generation hardware currently in our fleet,” says Guthrie.
Microsoft’s performance gap against close Gigantic Tech competitors is different from when Microsoft first released Maia 100 in 2023 and didn’t want to get drawn into head-to-head comparisons with Amazon and Google’s AI cloud capabilities. However, both Google and Amazon are working on next-generation AI chips. Amazon is equal works with Nvidia to integrate the upcoming Trainium4 chip with NVLink 6 and Nvidia’s MGX rack architecture.
Microsoft’s Superintelligence team will be the first to exploit Maia 200 chips, and the company is also inviting researchers, developers, AI labs, and open source modelers to an early preview of the Maia 200 development kit. Microsoft is starting to deploy these modern chips today in the US Azure data center region, with additional regions to follow.