The number Microsoft chose to lead with was not a benchmark score. On the second day of its Ignite conference, Nov 18, the company said its new Maia 300 accelerator delivers 1.7 times the token throughput per watt of the part it replaces, a measure of what it costs to serve a model rather than how quickly one can be trained.
The part it replaces is already shipping. Maia 200 is in volume production, Microsoft said. Maia 300 is not, and the company did not say when it will be, which process it is built on, or which foundry makes it.
That silence carries as much weight as the figure. Microsoft has argued for three years that custom silicon is a hedge against the price of Nvidia’s accelerators, and a hedge pays only if the parts arrive in volume and on time. A per-watt claim with no production date speaks to a future bill, not the present one.
The first Maia was unveiled at Ignite in November 2023, built on a five-nanometer process at Taiwan Semiconductor Manufacturing Co. and offered to Azure customers the following year. Microsoft has not said how much of its own inference now runs on Maia parts, or how many Maia 300 units it intends to deploy.
The efficiency figure lands in a narrow but expensive part of the business. Training a model is a capital cost absorbed once, spread across years of depreciation. Serving it is a bill that arrives every day, in megawatt hours and in rented capacity, for as long as customers keep asking questions. Tokens per watt measures that recurring cost, which is why it has become the number hyperscalers quote when they talk about their own chips.
Power is the constraint behind it. Data center campuses are being built against interconnection queues and transformer lead times, and a chip that produces more tokens for the same watt lets an operator sell more output from a site it has already paid to connect. On that arithmetic, a 1.7 times gain is worth more than an equivalent gain in raw throughput, because it does not require another substation.
The rest of the Ignite agenda pointed the same direction. A day before the chip announcement, Microsoft released seven models of its own, the sort of catalogue that only makes economic sense if the company expects to serve them on hardware it controls. Models and silicon are two halves of one margin argument: whoever pays for the tokens pays for the power, and Microsoft would rather bill for both than rent the middle layer.
What the company did not provide was a benchmark that would let an outsider check the claim. Tokens per watt depends on the model, the batch size, the sequence length and the precision of the arithmetic, and vendors routinely pick the configuration that flatters them. Nvidia has made the same kind of claim for each of its generations, with footnotes.
The comparison that matters is not against Maia 200 anyway. It is against whatever Nvidia is shipping when Maia 300 reaches customers, and that part is not standing still. Microsoft remains one of Nvidia’s largest buyers, and its own accelerators have so far supplemented that spending rather than replaced it.
There is also the software question, which has decided more chip programs than any process node. Nvidia’s advantage sits in a decade of libraries, kernels and debugging tools that engineers already know. A new accelerator has to be good enough that porting work pays for itself, and it has to stay good through several model generations, because the porting cost is paid once and the benefit accrues only if the hardware roadmap holds.
Microsoft’s answer has been to make Maia invisible to the people using it. Azure customers rent capacity, not chips, and the company decides which silicon sits behind a given endpoint. That arrangement hides the porting problem from customers, and it also hides the adoption numbers from everyone else.
For the buyers of Microsoft’s cloud, the practical question is cost per token on the workloads they actually run, and nothing announced on Nov 18 answered it. For the company’s suppliers and rivals, the question is capacity: how much wafer starts Microsoft has reserved, and for which quarter. Neither number was offered.
The strategic logic is easier to state than to execute. Every token served on in-house silicon is a token not served on rented hardware, and the gross margin difference compounds across a business measured in hundreds of billions of dollars of capital spending. That is why the program survives disappointing generations and why the announcements keep coming.
What Nov 18 established is narrower. Maia 200 is real and in production. Maia 300 exists on a slide, with one efficiency ratio attached and no date. The ratio is good if it holds under independent measurement, and the history of accelerator claims suggests it will hold for some workloads and not others.
Microsoft will presumably say more when the part is closer to shipping. Until then, the useful way to read the announcement is as a signal to Nvidia about bargaining power, and to Azure’s largest customers about where the cost curve is headed. Both audiences heard it. Neither got the number that would settle the argument, which is the price per million tokens on a date certain.


