The Chip Built to Answer

The announcement described the chip in the language of power rather than transistors. Maia 300, unveiled October 12, is Microsoft’s third-generation custom AI accelerator and its first designed primarily for inference — the work of answering, not teaching. A rack of 96 accelerators is expected to deliver up to seven times the token throughput of a Maia 200 rack, at a projected 40% more tokens per megawatt than leading GPU-based systems. Production begins in 2027 across deployments Microsoft measures in gigawatts, with the chip set to become the company’s primary inference platform for Copilot, agents and everything else it ships.

The numbers Microsoft led with are the operating numbers. Maia 300 is the third generation of the company’s custom AI accelerator and the first built primarily for inference — generating the answer rather than training the mind. Microsoft projects 40 percent more tokens per megawatt than leading GPU-based systems, and a rack of 96 accelerators delivering up to seven times the token throughput of a Maia 200 rack. A rack, a watt and a token: the chip is announced in the units its buyers actually pay in.

The workload list tells the same story. Long-context reasoning, low-latency conversation and sustained multi-step agentic tasks are the jobs of a model already trained — the traffic that spikes when a person asks, and spikes again when an agent asks on their behalf. Training runs once and makes the news. Inference runs forever and makes the invoice.

Training is a capital expense with a press cycle. Inference is rent, and Microsoft just bought the building.

Training Was the Headline

For three years the industry measured its chips by training runs — the cluster big enough to teach a frontier model, announced with a photograph of the rack and a founder standing beside it. Maia 300 points the other way: Microsoft says the chip is intended to become its primary inference platform for its own products and services, reducing its reliance on external GPU suppliers for the workloads it runs every day. The customers of that decision are Copilot, the agents, and everything else Microsoft ships.

Production is planned for 2027, with Microsoft targeting deployments it measures in gigawatts across the Maia family — capacity described in power rather than chip counts. It is a telling unit of account. The limiting factor has stopped being wafers and started being substations, and a company that once bought its accelerators by the pallet now forecasts its roadmap in megawatts.

One Platform, One Silicon

The strategic line is stated plainly: Maia becomes Microsoft’s primary inference platform for its own products. That is vertical integration in the oldest sense — the platform company deciding that the most expensive input in its business should not be priced by someone else. Every large platform is drawing the same conclusion at the same time, which is why the announcement matters beyond Redmond: the accelerator market is being split into chips made for sale and chips made for use.

What Microsoft did not claim matters too. There is no challenge to Nvidia’s training dominance here, no benchmark war on that ground — just a chip built to answer, sized for the second half of the sentence the industry has been finishing for three years. The race to train produced the models. The race to answer pays for them.