Microsoft Unveils Maia 300: 40% More Tokens per Megawatt Than a GPU It Won't Name
On October 12, Microsoft unveiled Maia 300, its third-generation in-house AI chip, and said that starting in 2027 its own silicon will become its primary inference system, deployed at multi-gigawatt scale.
Three key facts
40% more tokens per megawatt. In an official blog post, Scott Guthrie, Microsoft's EVP of Cloud and AI, said Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems. The compute chiplets use TSMC's N2P process, with eight HBM4 stacks, 18.0 TB/s of peak bandwidth and a 950 W TDP. A full rack holds 96 chips in a 167 kW envelope and delivers up to 7x the token throughput of the previous Maia 200.
Production deployments start in 2027, feeding Microsoft's own products first. Maia 300 will serve GitHub coding agents, Foundry model serving, MAI model inference and Dragon Copilot. The current Maia 200 already runs Microsoft Copilot in the US, Europe and Australia, serving OpenAI models, MAI models and major open-source models.
The goal is to lean less on other people's GPUs. The Information reported in August that Microsoft was negotiating with TSMC for more than 300,000 units for 2027 delivery. Microsoft still works with Nvidia and AMD on training; its own chips are taking the inference side first.
Wang Dou's Take
The whole post benchmarks against "leading GPU-based systems," and the word Nvidia never appears once. It's like telling a blind date "I'm 40% better than your ex" — without saying who the ex is or what's being measured. No third-party test, no named model: for now, treat the 40% as a wish.
Now do the math: counting rack power alone, 96 chips and 167 kW per rack means roughly 570,000 chips per gigawatt. If the final order is the 300,000-odd chips from the reports, that's just over half a gigawatt. Right now "multi-gigawatt" reads less like a build plan and more like a note to Nvidia's sales team: raise prices again and we'll make our own.
Source: The Official Microsoft Blog / Microsoft to expand Azure Maia system to multi-gigawatt deployments · Techzine