The real hardware story is agent efficiency, not scale. Papers on complexity-aware reasoning and on-device agents (PalmClaw) matter far more than another batch of llama.cpp increments—they're signaling where the actual compute bottleneck sits: not raw inference speed, but agents that know when to stop thinking and when to punt. TerraZero's procedural sim work is the unglamorous foundation that'll make self-play viable at scale, which every robotics company pretends to care about but almost none are actually cracking. Ignore the AI-can't-find-life panic and the patent hand-wringing; watch who ships sub-100mW agent inference on phones and who gets procedural sim stacks production-ready—that's where hardware gets designed next.