The biggest obstacles to deploying AI models on the blockchain are privacy and cost: data must be protected while keeping inference costs affordable. The ZK-FHE hybrid model proposed by Boundless Network combines zero-knowledge proofs with verifiable homomorphic encryption into a dual-stack process of 'compress first, prove later': users encrypt their inputs and send them to the computing nodes, which complete inference in the FHE domain and output ciphertext results; then, a succinct proof is generated to prove 'the result is correct and conforms to the model commitment', allowing verification of the contract to confirm inference validity at constant cost. In this way, users' private data and model weights remain encrypted at all times, with only a short proof and hash saved on-chain, without disclosing any confidential content. Compared to pure FHE, Boundless reduces computational overhead by over 60%; compared to pure ZK, it avoids the risk of model weight leakage, achieving 'double-blind trust'.
To prevent the model from 'malicious behavior' or performance degradation, Boundless has established a 'continuous evaluation market': any address can submit challenge samples, and if the model outputs incorrectly, it triggers a pledge reduction and mandatory update; submitters who improve accuracy can receive part of the reduction reward, creating a dynamic incentive. Model developers pledge $ZKC to guarantee service SLA and regularly release 'weight commitments' and performance proofs; once the model is updated, the old and new weights are seamlessly connected through recursive proofs, ensuring continuous credit between versions. This mechanism allows AI services to maintain trust and iteration in a decentralized environment, avoiding the rigid dilemma of 'launching and freezing'.
The AI computing link of Boundless also integrates a 'data treasury' function: data contributors declare data features through selective zero-knowledge summaries, and upon model invocation, they receive traceable rewards based on marginal improvements. All profit-sharing is settled in real-time through micropayment channels. If income falls below the Gas minimum, a unified aggregated batch payment is used to reduce tail costs. This creates an auditable three-sided market among data providers, computing power providers, and model providers, with fees and rights backed by on-chain evidence. The seemingly complex industrial chain is abstracted by a dual-track system of 'proof + payment', greatly simplifying collaboration thresholds.
Economically, the inference calling fee is divided into computation fees and proof fees. The computation fee is paid to computing power nodes, while the proof fee enters the verification pool and buyback pool. As the model invocation volume increases, the surplus in the verification pool is used weekly to buy back ZKC and destroy it. The buyback rate and deflation intensity dynamically follow throughput, truly embodying the principle of 'the more you use, the less supply there is'. If the decentralized AI track continues to expand, Boundless will become the preferred channel for high-value models on-chain, while ZKC plays a dual role of fuel and equity.
