Microsoft to increase use of homegrown AI chips, reduce Nvidia dependence
Microsoft announced it will significantly increase the use of its own AI chips, such as the Maia accelerators, to reduce reliance on Nvidia. The new Maia 300 chip enters production in 2027.
Microsoft plans to substantially increase its utilization of homegrown artificial intelligence chips, like its Maia accelerators, as it moves to reduce its reliance on Nvidia's graphics processing units (GPUs). The software giant stated on Monday that its third-generation Maia chips will become the primary architecture for running AI models. This strategic shift aims to bolster Microsoft's AI infrastructure and services.
The new Maia 300 chip, scheduled for production starting in 2027, promises significant performance enhancements. Microsoft claims the chip can handle seven times more AI tokens per rack compared to the previous Maia 200 model and delivers 40% more tokens per megawatt-hour than Nvidia's GPUs. This increased efficiency is expected to improve data center performance and potentially lower operational costs.
By developing its own silicon, Microsoft seeks to address the escalating demand for AI processors and offer greater choice within the market. While Nvidia chips remain crucial, particularly for model training, the use of in-house developed chips could improve Microsoft's long-term margins. The company first introduced its AI chip, the Maia 100, in 2023.
Analysts estimate that Microsoft's own chips will account for approximately 10% of the company's AI chip spending by the first quarter of 2027. However, Nvidia is expected to remain a key supplier, and Microsoft will continue its partnership for model training. Unlike comparable solutions from Google or Amazon, Microsoft's accelerators are not currently available for customer rental.