Microsoft is making its biggest push yet to run its own AI services on its own silicon. The company said Monday that its newest in-house accelerator, the Maia 300, will start rolling out next year, and that it expects to install its Maia chips at multiple-gigawatt scale across its data centers. The plan would make Maia the primary inference system behind Microsoft products — the hardware that answers every Copilot prompt, summarizes every Teams meeting, and generates every research response.
Inference is where the economics of AI are now decided. Training a model is a one-time expense; running it happens billions of times a day, and the bill is measured in electricity. Microsoft says the Maia 300 will deliver 40 percent more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, the company expects up to seven times the rack-scale token throughput of its current Maia 200. For an operator building data centers by the gigawatt, that efficiency gap translates directly into operating cost.
The chip itself reflects how specialized this hardware has become. Maia 300 uses a chiplet design built on TSMC's N2P process, with a software-defined dataflow architecture intended to cut data movement, higher memory bandwidth, and networking integrated into the design rather than bolted on. Microsoft tuned it for the workloads its products increasingly depend on: long-context reasoning, low-latency responses, and multi-step agent tasks that chain many model calls together.
This is the third generation of a program that is already in production. Microsoft began installing the Maia 200 early this year, and it now runs in data centers in the United States, Europe, and Australia, powering built-in AI features in Teams and Copilot. The chip also serves models from OpenAI and Microsoft's own AI teams, along with open-source model families. Rani Borkar, the executive who oversees Microsoft's data center hardware and chip design teams, framed the scale of the next step as a statement of confidence in the design, saying the company is backing that confidence by deploying Maia at gigawatt scale starting next year.
The move puts Microsoft alongside Amazon and Google, which both started their in-house accelerator programs earlier and are generally seen as ahead. Notably, Microsoft does not currently plan to rent Maia computing power directly to cloud customers, as its rivals have done with their chips — Maia is being built for Microsoft's own services first. Nvidia's GPUs still handle most inference work at Microsoft today, so this is a gradual shift in the industry's center of gravity rather than a sudden break. But when the world's largest software company designs its own path around the merchant GPU market, every data center operator watching power bills should be paying attention.