Microsoft has officially unveiled its third-generation Maia AI accelerator chip. The company intends to shift the majority of its future AI processing tasks to its proprietary hardware, reducing its long-standing reliance on Nvidia GPUs.
The Maia 300 is scheduled for mass production in 2027. Microsoft plans to deploy 96 Maia 300 chips per data center rack, doubling the density of the previous Maia 200 generation. This architecture aims to increase AI token processing throughput by up to 7 times per rack and improve energy efficiency by 40% compared to traditional GPUs.
While Microsoft is expanding its internal chip usage, it continues to utilize CPUs from Intel and AMD, and maintains partnerships with Nvidia and AMD for model training. Nvidia CEO Jensen Huang recently attended a Microsoft event where new Surface laptops featuring Nvidia processors were introduced.
Unlike competitors such as Google or Amazon, Microsoft does not currently offer its Maia chips for external rental. Analysts at Evercore ISI estimate that Maia chips will account for 10% of Microsoft's total AI chip procurement spending by the first quarter of 2027.
