Microsoft’s new AI chip makes it less reliant on Nvidia’s GPUs

Microsoft announced Maia 300, the third generation of its in-house AI chip, and said it will begin rolling into production workloads in 2027. The company describes it as the primary inference system for Microsoft products and services, with plans to scale the Maia family to multi-gigawatt deployments.

The headline claim, from cloud chief Scott Guthrie, is 40 percent more tokens per megawatt than leading GPU-based systems, and up to seven times the rack-scale token throughput of Maia 200 in a 96-accelerator configuration. If those numbers hold up, the cost of running Copilot-style services drops, and one of Nvidia's biggest customers starts buying a lot less from Nvidia.

We think this is the most consequential kind of chip news: not a benchmark war, but a big buyer deciding to become its own supplier. Microsoft is not saying it will stop using Nvidia or AMD. It is saying its own products will run mainly on its own silicon. That shifts leverage across the whole AI market.

The pace is the striking part. Maia 200 was only announced in January and is already running production work across Microsoft Copilot, GitHub Copilot, Microsoft Foundry and Microsoft's own models. Now its successor is on the calendar for next year. That cadence suggests Microsoft believes inference, the everyday work of answering prompts and running agents, is where the money and the electricity go, and where custom hardware pays off fastest.

The design choices back that up. Maia 300 is built on TSMC's N2P process with a chiplet layout that splits compute, memory control and I/O so more silicon goes to compute. Microsoft says it targets long-context reasoning, latency-sensitive conversation and sustained multi-step agent tasks. Those are exactly the workloads that make AI useful and also make it expensive. Cheaper tokens per megawatt is good for users and good for the grid, if efficiency gains are not simply swallowed by more usage.

Now the caution. Every performance number here comes from Microsoft, about a chip that is not yet in production. Tokens per megawatt compared to unnamed leading GPU systems is a vendor framing, and we want independent measurement before treating it as settled. Microsoft also calls Maia 200 the best performing AI infrastructure in its fleet, by its own real-world metrics. That may be true, but it is not something outsiders can check yet.

There is a bigger trajectory to name. When the companies that run the models also design the chips, co-design the models with them, and own the data centers, the stack gets more vertically integrated. Mustafa Suleyman says plainly that Microsoft needs a platform it owns, co-designed with its models. That can produce real efficiency. It can also lock customers deeper into one cloud, and make it harder for anyone without their own silicon to compete on price.

Our read: this is a credible, well-timed bet and a healthy pressure on Nvidia's pricing. I'm watching whether Maia shows up only inside Microsoft's own services, or whether Azure customers get real access to it, because that decides whether the savings stay in-house or reach everyone else.

Microsoft is one of several hyperscalers building custom AI chips to reduce dependence on Nvidia and AMD GPUs. The Verge reports that Maia 300 is also central to Microsoft AI's effort to catch up with Google DeepMind, OpenAI and Anthropic in frontier models, with the chip designed alongside Microsoft's own model work.