This framing makes sense. What we call “AI thinking” is really large-scale, non-sentient computation—matrix ops and inference, not cognition. Once you see that, progress is less about “intelligence” and more about access to compute. I’ve run training and batch inference on decentralized GPU aggregators (io.net, Akash) precisely because they treat models as workloads, not minds. You trade polished orchestration and SLAs for cheaper, permissionless access to H100s/A100s, which works well for fault-tolerant jobs. Full disclosure: I’m part of io.net’s astronaut program.