Microsoft plans gigawatt-scale Maia rollout to curb Nvidia reliance

Microsoft plans to deploy its Maia 300 chips at gigawatt scale starting next year, aiming to power its AI software mostly with homegrown hardware. The processors target inference, the work of running AI models to summarize conference calls, answer research questions and perform other tasks. Microsoft expects the chips to deliver 40% more tokens per megawatt than leading GPU-based systems.

The company is betting that chips tailored to its data centers can improve efficiency and reduce the expense of buying Nvidia graphics processing units, which currently handle most of its inference workloads. Amazon and Google are also developing in-house accelerators and are perceived to lead Microsoft after beginning their chip efforts earlier.

Microsoft began installing Maia 200 early this year. Those chips now support Copilot, Microsoft Teams and other services in data centers across the US, Europe and Australia, running models from OpenAI, Microsoft’s AI team and open-source families. Unlike Amazon and Google, Microsoft does not currently plan to rent Maia computing capacity directly to cloud customers. Data center hardware executive Rani Borkar said its own AI services business is large enough to take priority.