Microsoft Announces Maia 300 Chip: 40% More Tokens Per Watt Than GPU Systems, Says Nadella

Microsoft is going bigger on its own silicon. CEO Satya Nadella announced what he called the company’s biggest bet yet on its own chips, including a new AI accelerator called Maia 300 and a plan to scale the Maia family to multi-gigawatt deployments starting next calendar year.

The announcement was made in a post on X on October 12, 2026. Here is what was said, what the numbers mean in plain terms, and which claims are still unverified. For more on the AI hardware race, see our Microsoft Maia 300 tag page.

Microsoft Maia 300 Chip: What Was Announced

What Does Tokens Per Watt Mean?

AI models read and write in small units called tokens. Tokens per watt measures how much work a chip does for each unit of electricity. A higher number means cheaper and cooler AI, which matters because power and data-centre space are now among the biggest limits on AI growth. A 40% gain over leading GPU systems, if it holds in practice, would be a meaningful saving at Microsoft’s scale.

Maia 200 Is Already Running

Nadella said Maia 200 is already deploying across the United States, Europe and Australia, and that it delivers up to 1.7x more tokens per watt across experiences in Teams and Copilot. The words “up to” matter here, since the best case rarely matches the average workload.

Why This Move Matters

Cloud companies have relied heavily on GPUs from outside suppliers. By making Maia the primary inference system for its own products, Microsoft is signalling that it wants more control over cost and supply. Inference is the stage where a trained model answers questions, so it is where most daily AI usage happens. Running that on its own chips could lower costs for services like Copilot over time. Nadella also said Microsoft will keep bringing the latest innovations from across the ecosystem into its fleet, which suggests GPUs will not disappear from its data centres.

What Is Not Confirmed

The 40% and 1.7x figures are Microsoft’s own claims, shared in a social post without a published benchmark method, the models tested or the exact GPU systems used for comparison. Pricing, availability to outside customers and the precise chip specifications were not given. Total power and chip counts behind the multi-gigawatt plan were also not detailed. Wait for technical papers or independent tests before treating the numbers as settled. You can read more about the company on Microsoft’s website and Wikipedia.

Quick Answers

Details in early reports and social posts can change as companies publish official information, so please treat the figures above as reported at the time of writing. We will update this article when new or corrected details arrive, and we suggest checking the official pages linked here before you buy, invest or plan around anything mentioned in this story.

Source: Satya Nadella on X.

The bigger picture: Custom chips are becoming a core part of the AI race. Large cloud companies spend heavily on AI hardware, and even small efficiency gains can save a great deal of money and power when multiplied across many data centres. Microsoft’s message is that it wants to run most of its own inference on chips it designs, while still using outside hardware where it makes sense. Developers and businesses using Copilot or Azure should watch for any pricing or performance changes that follow, though none were announced in this post.