Microsoft to Deploy Custom AI Chips at Massive Scale, Reducing Nvidia Reliance

Microsoft is preparing to deploy its proprietary Maia AI chips across multiple gigawatts of data center capacity, marking a major push to reduce dependence on Nvidia's dominant graphics processing units. The company announced Monday that its newest Maia 300 chip will begin rolling out next year.

The custom semiconductors are specifically designed to handle AI inference tasks, such as generating responses to queries or summarizing content. According to Microsoft, the Maia 300 is projected to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems currently on the market.

The current generation Maia 200 chips are already operational in data centers across the United States, Europe, and Australia, powering AI features in Microsoft Teams, Copilot, and other products. These chips also run models from OpenAI and Microsoft's own AI teams, alongside open-source alternatives.

Rani Borkar, the Microsoft executive overseeing data center hardware and chip design, expressed confidence in the company's approach. By customizing chips specifically for its infrastructure, Microsoft aims to run inference operations more efficiently while controlling costs associated with procuring third-party hardware.

Unlike competitors Google and Amazon, which offer their custom AI chips to cloud customers, Microsoft has no immediate plans to rent out Maia computing power. The company states its own AI services business is substantial enough to prioritize internal use first.