Microsoft is preparing to unveil its next-generation Maia 300 AI accelerator as soon as September, positioning the third-generation chip as the workhorse of a cloud fleet that Redmond wants weaned off Nvidia. According to a Reuters report citing The Information, Microsoft has been in talks with TSMC to secure manufacturing capacity for more than 300,000 units of the chip for delivery in 2027, with eventual ambitions exceeding one million units.
From Maia 100 To A 300,000-Unit Order
Microsoft introduced its first Maia AI chip in November 2023 and followed with the Maia 200 in January 2026, built on TSMC's 3-nanometer process with a large on-die SRAM store. Volume for Maia 200 remained modest, in the tens of thousands. Maia 300 changes the scale of Microsoft's silicon ambitions and its dependency on TSMC as the sole leading-edge foundry able to hit that volume.
Reducing The Nvidia Bill
Microsoft's push mirrors a broader hyperscaler shift toward custom accelerators for internal AI workloads. Andrew Wall, general manager for Azure Maia, said Microsoft "continues to invest in custom silicon as part of our long-term AI infrastructure strategy," adding that the reported production figures "don't reflect the scale of our program." The company already ships its Cobalt CPU line alongside Maia; the July debut of the Anthropic in-house silicon team and continued M&A across the inference stack underscore how quickly the frontier lab economy is rewriting the compute supply chain.
TSMC Capacity As The Bottleneck
Booking 300,000+ leading-edge accelerators pulls Maia 300 into the same TSMC allocation queue as Nvidia's next Rubin generation and AMD's Instinct Helios roadmap. That is why the "one million" aspiration Microsoft is briefing internally is being taken with caution: capacity, advanced packaging and HBM supply all sit outside Microsoft's control. The Sony-TSMC image-sensor joint venture unveiled the same week, another billion-dollar Japanese fab, shows just how many suitors are lined up at TSMC's door. Microsoft's Maia 300 program will only be as fast as the foundry allows.
Reporting based on coverage from Reuters, Quartz, The Information and Bloomberg.