Microsoft is putting 1M+ of its own Maia 300 AI chips into Azure next year — should small teams care?
Microsoft announced the Maia 300, its next-gen custom AI chip, and says it will mainly run its own silicon for incoming AI requests — multi-gigawatt deployments starting 2027, potentially over a million chips next year. The stated goal is cutting dependence on Nvidia. I run a small side project on Azure's AI APIs and my bill is whatever the price-per-token says it is. Real question: does any of this ever reach small buyers as cheaper inference, or is it purely Microsoft improving its own margins? People who watched the Graviton/Trainium playbook at AWS — how long did it take before custom silicon moved prices for the little guys?
1 answer
- watch-the-price-per-token-lag
Custom silicon shows up in your bill with a lag, and the first move is always the provider's margin. AWS Graviton took roughly 12–18 months to turn into list-price cuts small buyers could feel, and Trainium followed the same pattern. The checkable thing: write down the price-per-million-tokens on the exact models you use today, then check again in six months. If Maia 300 really ships a million-plus units in 2027, expect the first visible inference price moves in the back half of the year. Don't re-architect around it yet — track it like a futures contract.