Microsoft makes Maia its main AI inference chip

Microsoft just put its own chip first in line for its own AI products.

Scott Guthrie, Microsoft's executive vice president of Cloud + AI, wrote on Oct. 12 that Microsoft will scale its Azure Maia chips to multi-gigawatt deployments starting in 2027, making Maia "the primary inference system for Microsoft products and services."

Maia 200 already runs Copilot

Maia 200 is in production now, Microsoft says, deploying across the United States, Europe and Australia. It runs OpenAI models, Microsoft's own MAI models and leading open-source model families. Microsoft Copilot already runs on it.

The next moves have dates. GitHub Copilot and Microsoft Foundry follow by the end of the year, according to the post.

Maia 300, the third generation, begins production deployments in 2027. Its compute chiplets are built on TSMC's N2P process. Eight HBM4 memory stacks give it 18.0 TB/s of peak memory bandwidth inside a 950-watt power budget. A rack of 96 of them runs within 167 kilowatts.

The number Microsoft wants you to remember

Microsoft expects Maia 300 to deliver

40% more tokens per megawatt than leading GPU-based systems.

That is Microsoft's projection, against GPUs it does not name, for a chip not yet deployed. It also says a full Maia 300 rack delivers up to 7x the token throughput of Maia 200. And it says Maia 200 already delivers up to 1.7x more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork.

Every cloud company says its own chip beats the GPU on its own workload. Grant that. What counts here is the commitment: gigawatts of capacity and Microsoft's biggest AI products, on Microsoft's chip.

Inference is where your bill lives

Training gets the headlines. Inference, running a trained model every time you send a prompt, is the cost paid every day. A chip that serves more tokens per megawatt can, in principle, serve them for less.

If you use GitHub Copilot, nothing changes on your screen. Your requests land on different hardware by year end.

If you build on Microsoft Foundry, you are on the same path. Microsoft says the two Maia generations together are designed for cloud-scale model serving in Foundry, and for multi-step agentic coding work at GitHub. Record latency, error rates and cost on your own workloads now. A baseline from before the move is the only way to see what changed after it.

For Nvidia, the post is narrower than some headlines. It covers inference for Microsoft's own products and services.

Watch Ignite next month

Microsoft says it will share more at Ignite next month in San Francisco and online. Two things to look for there: what Foundry charges for models served on Maia, and any outside benchmark of Maia 300.

Until then, 40% is a promise. Price it like one.