Video alone can't train physical AI. Been watching how the AI category prices infrastructure vs data. Interesting gap. $TAO is priced on compute. ~$5.79B FDV against ~$15M annualised revenue. The decentralised training thesis has clearly worked, and the market has assigned it a category-leader premium. $ICP is priced on on-chain infrastructure, canister compute, cross-chain AI workflow, agents. Multi-billion FDV. Revenue attribution thin. Priced on the thesis that AI infra can live on-chain. Both trades placed on the compute and infrastructure side. Fine. Physical AI training is where nobody's built the equivalent trade yet. Because physical AI isn't a video problem. It's a video + audio + gesture + spatial + haptic + timing problem. Six data streams, all aligned in time, captured from a human doing a task in a specific environment. No LLM shortcut solves this. No decentralised subnet generates it. It has to be captured. $KGEN is the only network I've seen operating on this specific piece. KAI ships 20,000+ hours of first-person multimodal video into AI laboratory training pipelines. All layers aligned. Contracts routed through Humyn Labs. Revenue disclosed. ~$85M annualised. FDV ~$200M. Compute is priced. Infrastructure is priced. The captured multimodal stream underneath the whole category isn't. #Altcoin Season#