I spent some time digging into OpenGradient’s HACA, and the part that stayed with me was surprisingly simple: the model work doesn’t happen inside blockchain consensus.
Inference runs on specialized nodes, while the chain verifies the proof and settles the result. That avoids forcing every validator to repeat the same heavy computation.
I also liked that verification can change depending on the use case, from faster hardware attestations to stronger ZK proofs.
It still feels early, but the architecture makes a lot more sense after following the full request flow.
Do you think separating execution from validation is the right path for scalable on-chain intelligence?