The short answer
● You cannot buy Maia. It is Azure-only silicon, not a product on a price list.
● The claim to watch: 40% more tokens per megawatt than leading GPU systems, from 2027.
● What it changes for you: Azure inference pricing, not your hardware decision.
Microsoft announced Azure Maia 300 on 12 October. It is worth being precise about what this comparison is, because it is not a buying decision for almost anyone reading it.
What was actually claimed
| Metric | Maia 300 | Maia 200 (shipping) |
|---|
| Status | Production 2027 | In production, US/EU/Australia |
| Memory | 288GB | Not restated |
| Bandwidth | 18TB/s (2.5x predecessor) | Baseline |
| Efficiency claim | 40% more tokens per megawatt vs leading GPU systems | Up to 1.7x tokens per watt |
| Rack throughput | 96-accelerator rack up to 7x Maia 200 | Baseline |
| Running today | — | Teams, Copilot, Cowork |
Cobalt 200, an agent-native CPU announced alongside, claims more than 50% higher performance than Cobalt 100, 33% lower agent call latency and 23% higher agentic-workload throughput.
Why the comparison is not what it looks like
Nvidia and AMD sell accelerators to anyone. Microsoft does not sell Maia; it runs Maia. So the practical comparison is not Maia 300 against a GPU you might buy — it is Azure inference priced on Microsoft silicon against Azure inference priced on someone else's, and against AWS and Google doing the same thing with Trainium and TPU.
The efficiency claim is the part that reaches customers. Tokens per megawatt is a cost-of-goods metric for the cloud provider, and in a power-constrained market it is the number that decides whether inference prices fall. A 40% improvement, if real and if passed on, is a price story rather than a performance story.
What to do with this
If you buy GPUs → nothing changes. Maia is not available to you, and Microsoft has said Nvidia remains central to Azure.
If you buy Azure inference → watch 2027 pricing, not the benchmark. That is where a tokens-per-megawatt gain shows up for customers, if it shows up at all.
If you are modelling long-term inference cost → note that all three major clouds are now pushing their own inference silicon. The direction is clear even where individual claims are unverified.
Sources:
Microsoft ·
Yahoo Finance