Maia 300: a promise of more tokens per megawatt doesn't set the cost of AI
The announcement includes the vendor's own projections for datacenter efficiency, but doesn't say what a customer will pay for an adequate answer. The buying decision depends on comparing cost per accepted task, quality and availability, under conditions similar to real use.
The announcement measures efficiency at datacenter scale
Maia 200 already serves production workloads, according to Microsoft, which also plans to expand the family in multi-gigawatt datacenters. Maia 300 is the next generation of accelerators aimed at inference, the stage where the model receives a request and produces an answer. Deployment is planned for 2027.
Microsoft projects that Maia 300 will process 40% more tokens per megawatt than leading GPU systems. In a full configuration with 96 accelerators, the company estimates up to seven times the rack token throughput, compared with Maia 200.
Efficiency doesn't equal savings for each customer
Tokens are units of text that models process. “Tokens per megawatt” compares processing volume to the power consumed in a given infrastructure. Rack throughput expresses the processing rate in the configuration described. Neither measure says how many valid tasks a company completes, what the service costs or how much it will pay for an accepted answer.
The figures are Microsoft's own forecasts for a system that hasn't gone into production yet. The announcement brings no independent test with a workload representative of each customer, comparison of answer quality or guarantee of lower Azure prices. More tokens per unit of energy can increase capacity or help the vendor manage costs, but passing that on commercially depends on contract, demand and usage profile.
The metric that matters for the decision
Managers can compare vendors by cost per accepted task, response time, quality, availability and volume of human review. The test needs to use requests similar to the real work, record acceptance criteria and keep the same security configuration. Measuring only tokens or power can favor a fast but inadequate answer.
This kind of reading also shows why training AI leaders should include basic technology and business indicators. The team doesn't need to choose a chip to ask the vendor how the announced efficiency reaches its bill, task or service level.
Sources
Microsoft Official Blog, 12 Oct 2026. https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/
Previously published on dg5i.com.br.