This strategy seeks to reduce reliance on Nvidia and AMD while bypassing energy constraints in data centers. The Maia 300 features 288GB of HBM4 memory and 18 TB/s peak bandwidth to handle low-latency chat and complex agent tasks. Microsoft is already running Maia 200 accelerators for Copilot and Teams across the US, Europe, and Australia, where it achieves 1.7x more tokens per watt. This hardware pivot follows internal cuts to its projected annual Claude spending by more than a third from over $1 billion to lower costs and improve margins.
Wednesday, October 21, 2026
Story
Microsoft Makes Maia AI Chips Primary System at Multi-GW Scale
AI9 days ago
The company is in negotiations with TSMC to secure more than 300,000 Maia 300 accelerators for delivery in 2027 to spearhead a transition to in-house AI hardware. CEO Satya Nadella has characterized the move as the firm's "biggest bet yet" on proprietary silicon, stating the Maia family will become the primary inference system for all Microsoft products and services at multi-GW scale. The Maia 300, built on TSMC's 2nm process with a multi-chiplet design, delivers 40% more tokens per megawatt than leading GPU systems and up to 7x the rack-scale token throughput of the Maia 200.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.