Microsoft to rely more on homegrown AI chips, reducing dependence on Nvidia
By increasingly counting on its own chips, Microsoft stands to expand margins and bolster its AI stack.
TL;DR
- Microsoft will increase its use of homegrown AI chips, starting with the third-generation Maia AI accelerator.
- The company aims to reduce its reliance on Nvidia's graphics processing units (GPUs) for AI workloads.
- Microsoft's own chips are intended to bolster its AI stack, including services like Azure, Microsoft 365 Copilot, and GitHub Copilot.
- The Maia 300 chip, planned for production in 2027, will feature 96 chips per data center rack, double that of the Maia 200.
- The Maia 300 is expected to handle seven times more AI tokens per rack and 40% more tokens per megawatt compared to GPUs.
- Using its own chips has already contributed to margin improvement for Microsoft.
- Nvidia will continue to be a significant partner for Microsoft, particularly in model training and for certain hardware deployments.
- Unlike Google and Amazon, Microsoft's AI accelerators are not available for rent by other companies.