Maia 300: the race for cheaper tokens runs through electricity
On October 12, Microsoft announced plans to expand the Azure Maia family to multi-gigawatt deployments starting in 2027. Maia 200 already serves production workloads; Maia 300 production deployments are planned to begin in 2027.
The company expects 40% more tokens per megawatt than leading GPU-based systems. A 96-accelerator configuration is said to deliver up to seven times Maia 200's rack-scale token throughput. These are Microsoft's performance claims, without independent validation in the sources reviewed.
What are you actually buying when you buy tokens?
A token is a unit of text a model processes or generates, sometimes a word and sometimes part of one. Every response involves chips, memory, and networking, all with a cost. A multistep agent task can trigger a long sequence of model calls, so even a small saving per step can matter at scale.
Throughput per megawatt describes how much work fits within a given power envelope. This matters when new data-center power connections cannot be added as quickly as equipment. Yet tokens are not equal units of value: a wrong answer produced quickly and cheaply still needs correction.
AnalysisCo-designing hardware and services lets Microsoft tailor its systems to its own workloads. Success could improve service margins or support more usage on the same infrastructure. The announcement alone does not establish customer prices or when they might fall.
When comparing two document-summary services, for example, check completeness, generation time, and editing effort. Price per token is one part of the bill.