Microsoft unveils Azure Maia 300, claims 40% more tokens per megawatt

Microsoft's Maia program is now on its third generation, with the Maia 300 joining the Maia 200 and Maia 100. The Maia 200 already runs Microsoft production workloads such as Microsoft 365 Copilot, along with OpenAI models, Microsoft's MAI models, and open-source models, so the new chip builds on silicon that is already in service rather than a fresh experiment.

Microsoft is not alone in this push. Its fellow hyperscalers Amazon, Google, and Meta, as well as AI labs OpenAI and Anthropic, either already offer or are working on their own custom chips. Amazon and Google now rent out their custom chips to third-party customers, and Amazon is looking into selling chips for customers to install in their own data centers. Nvidia regularly generates upwards of 50% of its quarterly revenue from hyperscalers, so the shift toward in-house silicon is a potential problem for it.

That does not mean hyperscalers will stop buying third-party chips. Customers still want access to a number of chips for different workloads, so companies like Amazon and Microsoft will need to keep stocking the most in-demand processors — and for now, those are Nvidia's.