Decentralized AI inference sits at the intersection of two hard problems. The first is the ver-
ification problem: a consumer of an inference result whether a human user, an autonomous
software agent, or an on-chain settlement contract cannot re-execute a billion-parameter
language model to check the executor’s answer. The second is the performance problem: ver-
ification mechanisms strong enough to provide cryptographic assurance are, by current prover
technology, far too slow and expensive for interactive use at LLM scale. The maturation of
large foundation models [1,2] and blockchain-native agent economies [3] makes both problems
acutely consequential: when an autonomous DeFi agent bases a six-figure liquidation decision
on the output of an off-chain language model, an incorrect but undetected result translates
immediately into irreversible financial loss.
The canonical blockchain mechanism for verifying computation, namely re-executing every
instruction on every validating node, is infeasible for AI inference: even a modestly sized
transformer would exhaust the Ethereum block gas limit by many orders of magnitude [4].
Three paradigms have developed to address this gap.