Microsoft Expands Production of Custom Maia Chips to Reduce Reliance on Nvidia
Microsoft has announced its third generation of custom AI accelerators, "Maia," in a strategic move aimed at strengthening its cloud infrastructure and reducing its historical reliance on graphics processing units (GPUs) from Nvidia. This step is designed to support the Azure platform and Copilot services, while improving power efficiency and lowering data center operational costs.
Evolution of Maia Chip Generations and Production Roadmap
The company unveiled its first dedicated AI chip, "Maia 100," in 2023, followed by "Maia 200" last January, which delivered 30% better performance per dollar. Microsoft aims to begin production of the new "Maia 300" generation by 2027, with plans to deploy the Maia family at scale—reaching multi-gigawatt capacities—giving the company greater flexibility to respond to changing market demands.
Breakthrough in Performance and Energy Efficiency
"Maia 300" is engineered to fit 96 chips per rack in data centers, doubling the density of the previous generation. The new accelerator can process up to 7 times more AI tokens per rack compared to its predecessor, while delivering 40% more tokens per megawatt than traditional GPUs, directly contributing to reduced operational costs.
Competition and Ongoing Partnership with Suppliers
Although analysts at Evercore ISI estimate that Maia chips will account for around 10% of Microsoft's total AI chip expenditure by early 2027, Microsoft confirmed that its long-standing partnerships with Nvidia and AMD for training complex models will continue. Unlike competitors such as Google and Amazon, Microsoft does not yet offer its AI accelerators for direct hourly rental to third-party companies, keeping them focused on powering its internal systems.