Microsoft unveils 2nm Maia 300 and targets multi-gigawatt custom-silicon deployments by 2027

Microsoft is making its biggest bet yet on in-house AI silicon. Maia 300 is a full system on TSMC's N2P process with 288GB of HBM4 at 950W. Microsoft claims 40% more tokens per megawatt than GPU-based rivals and up to 7x the rack-scale token throughput of Maia 200. Multi-gigawatt deployments start in 2027, with Maia positioned as the inference backbone for Copilot and partner models.

Maia 200 is already in production. It serves Microsoft 365 Copilot, OpenAI models and Microsoft's MAI foundation models, and it is deploying in the US, Europe and Australia. GitHub Copilot and Microsoft Foundry are slated to move onto it by end of 2026. The chip news coincides with a reported reshuffle of Azure leadership.

The competitive logic is about cost per token and Nvidia leverage. Google has TPUs, AWS has Trainium and Inferentia, and Microsoft has been the hyperscaler most dependent on Nvidia. A tokens-per-megawatt claim matters because power, not chips, is the binding constraint on AI buildouts.

The caution: Evercore ISI estimates Maia will cover only about 10% of Microsoft's AI compute near term, so Nvidia remains dominant. The 40% efficiency claim comes from Microsoft itself, not independent testing.