Microsoft will run its Maia AI chips at gigawatt scale from 2027

Microsoft will make its in-house Azure Maia chips the main AI inference system across its products starting in 2027, deploying them at multi-gigawatt scale.

The plan marks the clearest signal yet that Microsoft intends to reduce its dependence on Nvidia for the compute that powers AI features across its product lineup. The company has designed its own accelerators, validated them against leading GPU systems, and set a date — 2027 — for them to carry the primary inference workload.

CEO Satya Nadella frames the commitment in unusually strong terms, calling it "our biggest bet yet on our own silicon."

What performance claim does the Maia 300 make?

The upcoming Maia 300 should deliver 40 percent more tokens per megawatt than leading GPU systems, according to the company.

That metric matters more than raw speed. Tokens per megawatt measures how efficiently a chip converts electricity into AI output. At multi-gigawatt scale — think power consumption on the order of entire cities — efficiency per watt determines both operating cost and how much AI capacity Microsoft can actually build within real-world power constraints.

A 40 percent gain in tokens per megawatt means Microsoft could either serve more users on the same energy budget or cut the energy bill for a fixed workload by roughly a third.

Why does 2027 matter?

The date sets a hard deadline on Microsoft's silicon roadmap. Starting in 2027, Maia chips become the main inference system across Microsoft products — not a supplement to GPUs, but the primary platform.

The stakes for the AI infrastructure market are significant:

• Microsoft currently spends heavily on Nvidia GPUs to run AI workloads for products such as Copilot features across its portfolio.
• Shifting primary inference to in-house silicon at gigawatt scale would redirect a substantial share of that demand away from external suppliers.
• Custom chips tuned to a company's own workloads often deliver better cost efficiency than general-purpose GPUs — which is exactly the claim Microsoft is making with its tokens-per-megawatt figure.

What does this mean for the AI chip market?

Microsoft is not alone in pursuing custom silicon, but a multi-gigawatt commitment on a fixed timeline is among the most aggressive such plans disclosed to date. The decision touches every layer of the AI supply chain: chip design, data center construction, and the power contracts needed to feed these facilities.

For Nvidia, the world's dominant AI accelerator supplier, the announcement signals that its largest cloud customers are actively building alternatives rather than simply buying more GPUs. For hyperscale rivals, it raises pressure to accelerate their own accelerator programs.

For Microsoft itself, the bet cuts both ways. Owning the silicon means controlling costs and optimizing hardware directly for its own inference workloads. It also means carrying the full risk of chip development — design delays, manufacturing constraints, and the possibility that GPU vendors move faster on efficiency.

The so-what

If the Maia 300 delivers on its 40 percent tokens-per-megawatt advantage and Microsoft hits its 2027 deployment target, in-house silicon moves from an experiment to the backbone of one of the world's largest AI infrastructures. Nadella's words — "our biggest bet yet on our own silicon" — now come with a date, a performance number, and a gigawatt-scale commitment attached.