Microsoft's Maia 300 aims to become its default inference chip

Microsoft unveiled its third-generation in-house AI accelerator on Monday, the Maia 300 — and paired it with a commitment it has never made before. Starting next calendar year, CEO Satya Nadella said, the Maia family scales to "multi-gigawatt deployments" and becomes the primary inference system for Microsoft's own products and services. For a company that has spent three years buying more Nvidia chips than anyone else on earth, that is a public promise to run its own AI on its own silicon.

What was actually announced

Maia 300 enters production in 2027 and starts rolling out to production AI workloads the same year, according to Microsoft's announcement and subsequent coverage. The headline claims: 40% more inference tokens per megawatt than "leading GPU-based systems," and up to seven times the rack-scale token throughput of the Maia 200. Microsoft expects to pack 96 Maia 300 chips into each rack — double the Maia 200's density — and parts of Microsoft 365 Copilot should speed up as a result, technical fellow Jim Kleewein said.

The hardware details are still thin. A configuration breakdown circulated on X ahead of a fuller dossier describes a main compute die built on TSMC's N2P process node flanked by two I/O dies and four memory dies, with eight HBM4 stacks totaling 288 GB and a thermal design power around 950 W; trade press reported the rack as 96 accelerators, HBM4-equipped and mostly liquid-cooled. None of those figures has been independently verified, and Microsoft has not published a full spec sheet.

We flagged this moment back in August, when The Information reported that Microsoft planned to unveil Maia 300 in the fall and was negotiating TSMC capacity for more than 300,000 wafers in 2027 — see our brief, "Microsoft to unveil Maia 300 AI chip in September". The unveiling landed in October rather than September, and today's version dropped the two things the rumors promised: a production date and a scale commitment.

Why tokens per megawatt is the only number that matters

Every AI infrastructure announcement now speaks the same currency, because power — not chips — is the binding constraint on the buildout. A Bloomberg analysis out Monday put hard numbers on the stakes: constructing and equipping a 1-gigawatt AI data center in Abilene, Texas could cost roughly $30 billion, with about $17 billion of that going to chips, $3 billion to power and $2.4 billion to construction. Chips are more than half the bill of materials, and the electricity that feeds them is the scarce input that decides how much of that bill you can amortize.

That is the logic behind Microsoft's claim. If Maia 300 really delivers 40% more tokens per megawatt than the GPU systems filling Microsoft's data centers today, the same power envelope generates more billable work — and finance chief Amy Hood's line at a Goldman Sachs conference last month, that in-house chips have helped with "margin improvement over the course of a decade," stops being a hedge and becomes a strategy. The building-side numbers are worth understanding on their own terms; we laid them out in AI 101 — What is an AI data center?.

Who wins and who loses

Nvidia is the obvious target, but the near-term damage looks smaller than the rhetoric. Alistair Speirs, a Microsoft general manager, said the company still partners with Nvidia and AMD for training, and Evercore ISI estimated in a client note last Wednesday that Maia will represent about 10% of Microsoft's AI chip spending by the first quarter of 2027 — which means roughly 90% of the spend stays where it is. Jensen Huang shared a Microsoft stage a week ago for a Surface announcement. The gradualist reading: Maia buys Microsoft a pricing lever, not an exit.

The structural winner, if the numbers hold, is the custom-silicon thesis itself. Google has TPUs, Amazon has Inferentia and Trainium, Meta has MTIA, and OpenAI's Broadcom-co-designed Jalapeño already published benchmark claims of 1.5 to 1.9 times more work per watt than Blackwell-generation systems. Custom inference silicon is the one part of the AI hardware market where Nvidia's pricing power is genuinely contestable — which is also why Nvidia spent $3.5 billion on MediaTek to pull custom-accelerator makers into its NVLink Fusion ecosystem, as we analyzed in Nvidia pays $3.5B for MediaTek's custom-silicon handshake.

The loser, for now, is everyone hoping to rent Maia. Unlike Google's TPUs or Amazon's Inferentia and Trainium, Microsoft's accelerators cannot be leased by the hour to other companies. Microsoft keeps the advantage internal and gives up a third-party validation loop and a revenue line in one move.

The skeptic's case

Patrick Moorhead, CEO of Moor Insights and Strategy, put the problem in one sentence: "Nobody has tested them outside of Microsoft yet." The 40%-per-megawatt figure is a vendor claim measured against an unspecified "leading GPU-based system," with no independent benchmark to check it — a sharp contrast with OpenAI's Jalapeño, which at least showed up on SemiAnalysis's public InferenceX leaderboard. Then there is the cadence: Maia 100 in 2023, Maia 200 in January, Maia 300 now, each generation arriving before the last one could be measured at scale. That reads less like a confident ramp than a company that decided the power constraint will not wait.

There is also no named external customer. When Microsoft talked up Maia 300 in August, Anthropic was reportedly being courted as a buyer; Microsoft did not produce one today. And Evercore's 10% figure cuts both ways: it is a real number for a chip program barely two years into production, and a rounding error next to Nvidia's installed base.

What to watch

The load-bearing questions are simple. Does Microsoft publish independent numbers for Maia 300, or keep benchmarking itself? Does the multi-gigawatt commitment actually show up in 2027 deployments — and does HBM4 supply, tight for everyone, hold? Does Anthropic or any other frontier lab ever buy a Maia? And does Azure ever rent the chip out? The first answer will tell you whether Monday's announcement was an engineering milestone or a negotiating position aimed at Nvidia's price list.

Is running your own AI on your own silicon worth giving up the outside world's ability to measure it? Tell us in the comments.