Microsoft projects Maia 300 at 40% over GPUs, below Maia 200’s 1.7x

Microsoft expects its next AI accelerator, Maia 300, to deliver 40% more tokens per megawatt than leading GPU-based systems. In the same announcement on 12 October, the company said the already deployed Maia 200 is delivering up to 1.7x more tokens per watt than leading GPU infrastructure. Those are the same unit. On Microsoft’s own numbers, the chip it hasn’t shipped yet carries a smaller claimed efficiency lead over GPUs than the one already running in its data centers.

The two figures aren’t measured the same way, and Microsoft doesn’t present them as a pair. The Maia 200 number is scoped to built-in AI capabilities in Microsoft Teams, Copilot and Cowork experiences. The Maia 300 number is a company expectation for a part that enters production in 2027, as CNBC reported. Neither claim names the GPU generation it was measured against.

That gap matters because the GPU comparison is the whole reason this launch got written up. Microsoft’s own technical deep dive, published the same day by Rani Borkar and Saurabh Dighe, benchmarks the part against Maia 200 rather than against anyone’s GPU. Its headline performance claims are all set against the previous generation.

The 7x is a rack number, and the rack changed

Take the biggest of those claims. The deep dive says Maia 300 delivers up to 7 times greater rack-scale token throughput in its full 96-accelerator configuration, and more than 2.5 times the performance per watt. The catch is that the rack it describes holds 96 accelerators across 24 four-accelerator trays, “doubling the accelerator density of Maia 200”. On the silicon itself, the same post credits its chiplet strategy with a 36% increase in compute FLOPs.

Which means the headline throughput figure rests on twice the silicon in the rack, plus memory and networking work, rather than a 7x jump in any single chip. That’s a reasonable way to build an inference system, and it’s also why the per-accelerator numbers are worth reading next to it.

Maia 300 figure	Value	Compared with
Tokens per megawatt	40% more (expected)	Leading GPU-based systems
Rack-scale token throughput	Up to 7x	Maia 200
Performance per watt	More than 2.5x	Maia 200
Compute FLOPs	36% more	Maia 200
Accelerators per rack	96, in 24 trays	Double Maia 200 density
Rack power envelope	167 kW	n/a
Memory per accelerator	288 GB, 18 TB/s	Eight HBM4 stacks

Nobody outside Microsoft can check

Even so, independent verification is where this gets thin. Microsoft also posted a quote sheet of analyst commentary to its Maia 300 press site. The document states its own purpose in the first paragraph: the quotes “can be used as third-party perspective to help contextualize the announcement and reinforce its broader industry relevance”. Austin Lyons of Chipstrat is quoted there calling inference “increasingly a data movement problem”, which is the case the architecture is built around.

Nobody has tested them outside of Microsoft yet.

Patrick Moorhead, CEO, Moor Insights and Strategy, via CNBC

That’s the structural limit here. Unlike Google’s tensor processing units or Amazon’s Inferentia and Trainium, CNBC reports, Microsoft’s accelerators can’t be rented by the hour, so no outside party can run its own benchmark. Every number above comes from the vendor. We flagged the same pattern when Microsoft benchmarked its RTX Spark laptops at 64GB rather than the 128GB it sells.

None of this is a change of direction. Microsoft has been pushing its own silicon and its own models to cut inference costs for a while, and finance chief Amy Hood told a Goldman Sachs conference last month that in-house chips have helped with “margin improvement over the course of a decade”. Tokens per watt is becoming the standard unit of that argument, the same frame OpenAI used for its Jalapeño chip.

What to watch is the ramp, not the slide. Evercore ISI analysts estimated in a client note that Maia will account for about 10% of Microsoft’s AI chip spending by the first quarter of 2027, per CNBC. General manager Alistair Speirs declined to give a timeline for the multi-gigawatt deployments, saying Microsoft wants to “reserve a little bit of optionality there”. Guthrie said more is coming at Microsoft Ignite 2026 next month in San Francisco.