Microsoft just fired its biggest shot yet at Nvidia's dominance. On Monday, the company unveiled Maia 300, the third generation of its custom AI chip — and announced plans to scale it to multi-gigawatt deployments starting in 2027. CEO Satya Nadella called it "our biggest bet yet on our own silicon."
What's new
Maia 300 is built to run the AI workloads Microsoft actually cares about: coding, reasoning, and agentic tasks — the kind where an AI carries out multi-step business work on your behalf. Microsoft says the chip will become the "primary AI inference system for Microsoft products and services," meaning Copilot, Teams, GitHub, and Azure would all eventually run on Microsoft's own chips instead of Nvidia's.
The key specs
The chip uses a new chiplet architecture, with compute chiplets built on TSMC's 2-nanometer process and cheaper 3-nanometer chiplets handling data movement. Microsoft claims a full rack of 96 Maia 300 chips will deliver 40% more useful computing per megawatt than "leading GPU-based systems" — though it didn't name which systems it compared against. It also promises up to seven times the token throughput of the previous Maia 200 chip at rack scale. A full rack uses about 167 kilowatts, compared with 142 kilowatts for Nvidia's latest GB300 server racks.
Why it matters
This is about money. Running AI models is wildly expensive, and Microsoft expects to spend more than $175 billion on capital expenditures this year alone. Its predecessor, Maia 200, is already running Microsoft's in-house models — and OpenAI models on Azure — and Microsoft says it costs 30% to 40% less than comparable Nvidia hardware. Maia 300 is designed to push that gap wider. The company is reportedly in talks with TSMC to secure more than 300,000 Maia 300 chips for 2027, aiming for over a million eventually.
The bigger picture
Microsoft isn't alone in this bet. Google and Amazon have been running custom AI silicon at scale for years, OpenAI is developing its own inference chip with Broadcom, and Meta has started producing its Iris chip. The whole industry is trying to cut the cost of every AI request — a cost that only grows as AI agents become more capable. Microsoft says it will keep buying Nvidia and AMD chips too, so this isn't a divorce. But the direction is unmistakable: the cloud giants want to own the whole stack, down to the silicon.
Verdict
Maia 300 won't show up in any product you can buy — it's data-center hardware, not a gadget. But for Microsoft's bottom line, it might be the most important thing the company announced this year. If its efficiency claims hold in the real world, cheaper tokens mean cheaper AI for everyone downstream. Worth watching: whether Microsoft's own models really run better on Maia, as Mustafa Suleyman's team claims — and whether Nvidia feels the pressure.
Comments
Post a Comment