Two numbers, two different comparisons

Read the two headline figures separately, because they measure different things. The 7x number is rack against rack, new Microsoft chip against old Microsoft chip. That is a real generational gain if it holds, but Maia 200 is a low base: it arrived in January after delays, and as of this summer it was running in just two US data centers, according to reporting by The Information cited by GeekWire. Beating your own scarce chip by 7x at rack scale is partly a density story: 96 accelerators in one rack will out-produce a smaller, earlier system almost by construction.

The 40% number is the one aimed at Nvidia, and it is the one with no evidence attached. Tokens per megawatt against "leading GPU-based systems," with no named GPU, no model size, no precision format and no workload described. Microsoft's own Maia 200 claim uses the same construction, 1.7x more tokens per watt than "leading GPU infrastructure." A comparison you cannot reproduce is marketing until the company publishes the setup. The technical paper may answer some of this; the blog post does not.

Microsoft is still catching up

The strategic direction is not in doubt, and it is worth saying plainly why. Nvidia's gross margins, estimated at roughly 73 to 75%, are the tax every hyperscaler pays on inference, and inference is now the recurring bill. Google has shipped TPUs since 2015 and is on its seventh generation. Amazon's Trainium is on its third. Microsoft's first Maia arrived in 2023, and analysts at Evercore ISI estimate Maia will still be only about 10% of Microsoft's AI chip spending by early 2027. Nadella has also said Microsoft will be among the first to deploy Nvidia's next-generation Vera Rubin systems. Maia is a hedge being scaled, not a replacement being switched on.

What this changes for anyone buying Microsoft AI

Nothing you can order changes today. There is no Maia instance on a price list, and there will not be one under this plan: the chips serve Copilot, Teams, GitHub and Foundry from inside. If the efficiency claims are even half right, the effect shows up indirectly, as more Copilot capacity per megawatt in a business where power, not chips, is now the binding constraint. That is the number to watch at Ignite next month: not tokens per megawatt in a blog post, but how much serving capacity Microsoft says Maia actually carries by the end of 2027.

Sources

  1. [1] Microsoft (Scott Guthrie), "Microsoft to expand Azure Maia system to multi-gigawatt deployments"Read source
  2. [2] GeekWire, "Microsoft makes 'biggest bet yet' on its own silicon, unveils new AI chip to reduce reliance on Nvidia"Read source
  3. [3] Bloomberg via Advisor Perspectives, "Microsoft Prepares to Put Gigawatts of Its Chips in Data Centers"Read source
  4. [4] Mobile World Live (Reuters), "Microsoft unveils Maia 300 custom AI chip"Read source