Satya Nadella bets multi-gigawatt AI future on Microsoft's own Maia chips

"Today we are announcing our biggest bet yet on our own silicon," Microsoft chairman and CEO Satya Nadella said in a post on X, setting out plans to put the company's custom Maia chips at the centre of its AI infrastructure.

"Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own products and services," he added.

The announcement also introduces the next generation of the chip. "We're introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems," Nadella stated.

"We are already deploying Maia 200 across the US, Europe and Australia, delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot," Nadella said.

The Maia 200 is built on TSMC's 3nm process with more than 140 billion transistors, and delivers over 10 petaFLOPS in 4-bit precision within a 750-watt power budget, with 216GB of HBM3e memory, according to Techzine. Microsoft designed the accelerator for inference rather than training. The international footprint marks a step up from June, when Nadella said two data centres were using Maia 200 and that the company planned to scale it further. Andrew Wall, General Manager of Azure Maia, has said he expects Maia 200 to remain in use for another four to five years.

In practical terms, the announcement means that from 2027 the AI features inside Microsoft's products, including Copilot and Teams, will be served predominantly by chips the company designs itself. Third-party GPUs will remain in the fleet, though they will no longer be the default engine for Microsoft's own workloads.

The metric Nadella chose matters. Tokens per watt measures how much AI output a system produces for each unit of electricity it consumes, so a 40% gain means more Copilot responses from the same power supply. Across deployments, Nadella is sizing in gigawatts; that saving compounds into lower running costs for every AI feature Microsoft sells.

"Our goal is to build on the full-stack efficiency gains we've already seen, while continuing to bring the latest innovations from across the ecosystem into our fleet," Nadella noted, signalling that Microsoft will continue to buy external hardware alongside its own. Nadella has previously said building the company's own silicon to work alongside Nvidia and AMD chips is a priority, Cloud Wars reported.

Runtime Wire noted that Maia 300 could lower Microsoft's inference costs and free up Nvidia systems for paying Azure customers, though the strategy depends on manufacturing scale and winning an outside customer. The Information reported that Microsoft is confident it can persuade more customers to rent Maia 300 chips, potentially including Anthropic, partly on the strength of the cost savings. The push has deep roots: in a 2022 email later published in court documents, Nadella wrote: "right now we are a very thin layer on top of Nvidia".

Microsoft has been in talks with TSMC to secure manufacturing capacity for more than 300,000 Maia 300 chips for delivery in 2027, and ultimately aims for more than 1 million, though component shortages and capacity negotiations could constrain those plans, Reuters reported. "The figures reported don't reflect the scale of our programme," Wall told Reuters.

Even at that volume, Microsoft's ambitions sit below those of its hyperscale rivals. The company introduced Maia in November 2023 but has lagged Alphabet and Amazon in scaling its in-house chip efforts, and the Maia line is widely seen as less mature. The Information, citing Morgan Stanley estimates, reported that Google plans to make more than 3 million TPUs in 2026 and 5 million in 2027. The multi-gigawatt commitment now gives Microsoft a public benchmark against which that gap will be measured.