The Ethereum Foundation, together with The Open Anonymity Project, has introduced zkAPI, a zero-knowledge based system for private AI inference, now live on the Ethereum mainnet under the ZkApiVault contract. The deployment supports deposits in ETH and USDC and allows users to access commercial large language models without linking credit cards or static API credentials to every request.
The architecture uses a local cryptographic balance called a private note. Users make a one-time deposit into a smart contract, then a local client generates zero-knowledge proofs certifying available funds without exposing funding origin or remaining balance. Once verified, the system issues a temporary key with a dollar-denominated spending cap that exists only in local volatile memory and routes queries directly to the inference provider. When a session ends, the provider issues a signed receipt reflecting exact usage.
Privacy separation is a core feature: the financial server never sees user prompts, while AI inference providers remain unaware of the payer's identity or account details. The protocol uses the Groth16 proof system over the BN254 curve, Poseidon hashes, and 32-level Merkle trees to prevent double-spending of compute credits without compromising on-chain anonymity.
The local client emulates API specifications of OpenAI and Ollama, making it easier for developers to plug in chat interfaces, code editors, and autonomous agents. However, the Ethereum Foundation notes that zkAPI does not resolve network-level anonymization by itself; users may still need Tor or a VPN, and stylometric identifiers in prompt text could still present privacy risks.
The open-source release includes client code and verification contracts, with a test deployment on Sepolia and an active contract on Ethereum Mainnet. The underlying framework builds on ZK credit research published by Vitalik Buterin and Davide Crapis.
