Microsoft launches third-generation Maia 300 AI chip, plans to reduce reliance on NVIDIA

According to CNBC, Microsoft officially released its third-generation in-house AI accelerator chip, the Maia 300, on October 12, announcing plans to use the Maia series as the primary architecture for running AI models to gradually reduce reliance on NVIDIA GPUs. The Maia 300 is slated for production in 2027, with each data center rack capable of hosting 96 chips (doubling over the previous generation), increasing AI token processing volume per rack by seven times and improving token processing efficiency per megawatt by 40% compared to GPUs. Microsoft aims to deploy multi-gigawatt-scale Maia chip clusters, with Microsoft 365 Copilot expected to be the first to benefit from the performance improvements. Analysts at Evercore ISI estimate that by Q1 2027, Maia chips will account for approximately 10% of Microsoft’s AI chip spending. Notably, Maia accelerators are currently reserved exclusively for internal use and are not available for external leasing; NVIDIA and AMD will continue to participate in Microsoft's model training operations.