The shift centers on the new Maia 300 chip, which uses TSMC's N2P process and 8 HBM4 memory stacks to deliver 40% more tokens per watt than top GPU systems. Microsoft will scale its Maia family to multi-gigawatt deployments by 2027 to serve as the primary inference system for all products, reducing reliance on Nvidia. The Maia 300 also offers up to 7x the token output per rack compared to the Maia 200, which is already active across the US, Europe, and Australia.
Tuesday, October 20, 2026
Story
Microsoft Cuts Claude AI Spend by Over a Third for Maia Chip Pivot
AI8 days ago
The software giant is shifting internal artificial intelligence workloads to its own custom silicon to lower costs and increase profit margins. Microsoft reduced its projected annual expenditure on Anthropic's Claude AI models by more than a third from an original budget of over $1 billion, while slashing per-employee AI spending from $100,000 a month to $10,000. This migration allows the company to keep token costs internal as it builds both its own models and hardware.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.