Microsoft pushes homegrown AI chips with Maia 300 launch and cost‑saving deployment plan

MSFT on Monday, October 12, unveiled Maia 300, the next generation of its in-house AI accelerator, and said it plans a large deployment of its own chips with an eye on saving costs. In a blog post, Microsoft cloud chief Scott Guthrie said the company will scale its Azure Maia family to multi-gigawatt deployments starting in 2027, which is also when Maia 300 begins production deployments. Maia 300 has key components built by TSM; Microsoft says its compute chiplets use TSMC's N2P process node.

The specifications target inference cost. Each chip uses eight HBM4 memory stacks for 18.0 TB/s of peak bandwidth within a 950-watt power budget, and a rack holds 96 accelerators within 167 kilowatts. Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and up to 7x the rack-scale token throughput of the current Maia 200, which is already in production in the US, Europe and Australia serving Teams and Copilot. Rani Borkar, who runs Azure hardware systems, told Reuters the chip delivers "more intelligence for every watt, every dollar, and every byte of memory". Reuters compared the rack with NVDA's GB300, which holds 72 GPUs and draws about 142 kilowatts.

The efficiency figures are Microsoft's own projections for a chip that does not enter production until next year, and Microsoft gave no dollar savings estimate and said nothing about reducing its purchases of Nvidia or AMD GPUs. Volume is also unconfirmed: The Information reported in August that Microsoft was seeking more than 300,000 Maia 300 chips from TSMC for 2027, a figure a Microsoft executive said does not reflect the scale of the program. Microsoft shares traded about 1.3% higher late Monday morning, per SentiSense data. What to watch: independent benchmarks once Maia 300 ships, any capital-spending commentary on Microsoft's October earnings call, and whether other hyperscalers' custom chips pressure GPU pricing.