Microsoft is putting its own chips at the center of its AI infrastructure. On October 12, 2026, CEO Satya Nadella announced that Microsoft will scale its Maia family to multi-gigawatt deployments starting next calendar year, making Maia the primary inference system for the company's own products and services. Alongside that, Microsoft introduced Maia 300, which it says delivers 40% more tokens per watt than leading GPU-based systems.
Nadella called it Microsoft's biggest bet yet on its own silicon.
What Microsoft announced
Three things stand out:
Scale: Maia moves to multi-gigawatt deployments starting next calendar year.
Role: Maia becomes the primary inference system for Microsoft's own products and services.
A new chip: Maia 300 delivers 40% more tokens per watt compared to leading GPU-based systems.
Why tokens per watt is the number to watch
Inference is the work of running a trained AI model to produce answers, and every answer is made of tokens. As more people rely on AI in tools like Microsoft Teams and Copilot, the number of tokens served keeps growing, and the power needed to serve them becomes a central constraint.
Tokens per watt measures how much output a system produces from each unit of power. A higher number means more AI capacity from the same energy, which is why Microsoft led its announcement with this metric.
Maia 200 is already in service
Maia 200 is already deploying across the US, Europe, and Australia. Nadella said it is delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot.
That means the strategy is already running in products people use, not only on a roadmap, and it gives Maia 300 something concrete to build on.
Maia 300: 40% more tokens per watt
Maia 300 is the next step. Microsoft says it delivers 40% more tokens per watt compared to leading GPU-based systems.
The comparison point matters. It is against the leading systems on the market rather than an older baseline, which is the bar any custom chip has to clear to justify running at scale.
Built on full-stack efficiency
Nadella said Microsoft's goal is to build on the full-stack efficiency gains it has already seen. The phrase signals that the gains don't come from the chip alone, but from how the hardware and the software and services running on it work together.
Microsoft also said it will keep bringing the latest innovations from across the ecosystem into its fleet. Read alongside the word "primary" rather than "only," that suggests a fleet where Maia leads inference but isn't the only technology in use.
Why multi-gigawatt scale changes the math
At multi-gigawatt scale, a 40% efficiency gap is not a rounding error. It translates into a large amount of extra serving capacity from the same power. That is why the announcement is as much about energy and capacity as it is about chips.
What to watch next
How quickly Maia deployments scale once the multi-gigawatt rollout begins next year.
Whether Maia 300 extends the Maia 200 gains already seen in Teams and Copilot.
How Microsoft balances Maia with other technologies in its fleet.
The bottom line
The announcement is a statement of intent: Microsoft wants the economics of AI inference to be a strength it builds itself. With Maia 200 already deployed, Maia 300 raising the efficiency bar, and multi-gigawatt scale-up starting next year, Maia is set to become the primary system behind Microsoft's own AI products.
Comments (0)
Join the discussion by logging into your account.
No comments yet. Be the first to comment!