Microsoft makes Maia its primary inference system, unveils Maia 300 with 40% better tokens per watt

Nadella called it Microsoft's 'biggest bet yet on our own silicon'. Maia 200 already delivers up to 1.7x more tokens per watt for Teams and Copilot, and GitHub Copilot and Microsoft Foundry move onto it by year-end. Microsoft plans to deploy more than one million Maia 300 units in 2027. Evercore ISI estimates Maia will make up about 10% of Microsoft's compute. Microsoft also reshuffled Azure leadership.

The strategy mirrors Google's TPUs and AWS's Trainium and Inferentia. Inference is the growing share of AI spend, and custom silicon tuned for it can undercut NVIDIA GPUs on cost per token. Shifting OpenAI model serving to Maia is especially significant given the scale of that workload.

Also this week, Azure Enclave became generally available for government and regulated workloads, and Microsoft confirmed usage-based billing for Microsoft 365 Copilot Business starting December 1.

Caveats: the 40% claim is vendor-measured, and earlier Maia generations shipped late. Watch for delivery timelines and whether external Azure customers get access.