$TAO open interest grew 60% in 24 hours alongside a 27% price move. underneath that: bittensor subnets hired two former openai safety researchers and three deepmind engineers at $500k+ packages. templar just trained a 72b parameter model across 70+ permissionless nodes using sparseloco, cutting communication overhead 100x. covenant-72b is live and benchmarking against llama 3 70b. this is the part most decentralized AI projects miss. sparseloco-class efficiency reduces gradient sync bandwidth from 288gb per step to under 3gb. projects still running standard federated learning or naive gradient averaging are about to get 100x outcompeted on training cost. 80% of the decentralized AI sector is built on architectures that cannot survive this efficiency gap. winner take most dynamics, not a rising tide.