Microsoft to Deploy Maia 300 AI Chip in 2027
Microsoft is moving its own chips to the center of its inference infrastructure, which shifts a large share of AI compute demand away from merchant GPUs and toward in-house silicon.
Reporting from 2 sources: ASCII.jp, GIGAZINE.
Microsoft announced on October 12 that it will scale its Azure Maia AI accelerator to several gigawatts starting in 2027 and deploy its third-generation chip, Maia 300, the same year. The current Maia 200 already runs Microsoft Copilot and is deployed in the United States, Europe, and Australia. Microsoft claims Maia 300 can raise tokens per megawatt by 40% versus preceding GPU-based systems and hit up to 7 times Maia 200's rack-level throughput in a 96-accelerator configuration.
Maia 300 is built for inference workloads the company calls long-context, latency-sensitive, and multi-step agent processing. Microsoft says a rack fitted with 96 accelerators reaches up to 7 times the token throughput of Maia 200, and it puts the per-megawatt gain at 40% over preceding GPU-based systems.
The Maia 200 is already serving Copilot and built-in features in Teams across the United States, Europe, and Australia, with Microsoft claiming up to 1.7 times more tokens per unit of power than GPU infrastructure. The company says its own chips will become its primary inference hardware, tying models, software, and silicon together.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.