AI news · October 14, 2026
Microsoft plans Maia 300 for 2027 inference at multi-gigawatt scale
Maia 300 is planned for 2027 with 40% more tokens per megawatt and up to 7x Maia 200 rack throughput; builders should watch total serving cost, not only chip speed.
- expected tokens per megawatt gain
- 40%
- reported rack throughput
- 7x
- planned production start
- 2027
Microsoft says Azure Maia 200 is already serving production workloads, including Copilot, while Maia 300 is planned for production deployments in 2027. The new system targets long-context reasoning, low-latency conversation and sustained multi-step agent work. Microsoft expects 40% more tokens per megawatt than leading GPU-based systems, and says a full rack with 96 accelerators can deliver up to seven times the rack-scale token throughput of Maia 200. The figures are forward-looking company estimates.
The release matters to product teams because inference cost is becoming part of model choice. Maia 300 is not a chip that a small maker can install, and the public plan does not give a general customer price. The useful response is to track tokens per request, requests per task, latency, GPU hours and the cost of retries in your own service. A model route that looks cheap per token can still lose if it needs more context or more attempts to finish the same job.
What you can do with it
Add serving cost and retry count to your model dashboard. For one high-volume workflow, record tokens, latency, hardware hours and accepted output, then compare a fast small model with a stronger model. Use the results to set a cost ceiling before a new accelerator or hosted route becomes available.
Our take
Maia 300 shows that the infrastructure fight is moving toward useful work per watt, not only raw model size. Microsoft owns the comparison, and deployment is still ahead. Makers should treat the announcement as a reason to measure total inference economics now, so future hardware improvements can be priced against a real baseline.
Source: Microsoft — Made With Models writes the brief; the reporting is theirs.