The capex is still "early" for sure (i.e. data centers are still being planned and built out, we're hardware/energy constrained). If model scaling holds out, w

The capex is still "early" for sure (i.e. data centers are still being planned and built out, we're hardware/energy constrained). If model scaling holds out, we're "early-ish" in terms of the reliability and performance of these systems, just based on utilization of the compute from the planned capex. If we hit hard diminishing returns and we don't find architectural/data workarounds, that would put a wrinkle in things, but I suspect that the AI we have now is capable of helping us find those workarounds and keep things moving.