The public-market story on AI infrastructure has come unstuck from the usage story it used to be a proxy for.

Token volumes at the major model labs are up — quarter on quarter, not just year on year. Enterprise commitments are higher than at any point in the cycle. None of that is showing up cleanly in the share prices of the companies most of us would have called "AI infrastructure" eighteen months ago.

The capex that doesn't show up

Part of the story is that the marginal dollar of AI capex isn't going where it used to. The model providers are increasingly building, leasing or renting compute through paths that bypass the public-equity names that benefitted from the first wave. The disclosed deals make this obvious; the undisclosed ones, by reputation, even more so.

Part of the story is also that "AI infrastructure" was always a label investors applied loosely. Half the names that ran in 2024 had revenue that was AI-adjacent, not AI-driven, and that distinction has caught up with several of them.

What to actually watch

If you want a leading indicator that's still correlated with usage, the most honest one we've found is agent-mediated session count on enterprise SaaS — the very metric most companies in our coverage cannot give us cleanly. The companies that can measure it are the ones that haven't moved on multiples; the companies that can't are the ones that have.

That's not financial advice. It's a research note about what's measurable.