July 2026 AI Insights

Three structural patterns dominated the July conversation with the boards we work with, and none of them were about a specific model release.

First: the cost curve bends in both directions. Inference keeps getting cheaper per token, but total AI spend keeps rising, because consumption is growing faster than unit cost is falling. The CFO question is no longer "what does it cost per call" — it is "what is the ceiling on consumption, and who sets it".

Second: the sovereignty question moved from slide decks into procurement contracts. Swiss and European buyers are asking where weights are hosted, where inference runs, and whose jurisdiction logs sit under. The answer is often inconvenient, which is why it was ignored for two years. It cannot be ignored any longer.

Third: the gap between "we have an AI strategy" and "we have deployed AI that changes a number on the P&L" widened, not narrowed. Strategies are easier to write than ever. Deployment is as hard as it ever was.

None of this is dramatic. All of it is compounding.