The Ethereum Foundation put zkAPI on mainnet on October 1, a payment rail that lets a user buy AI inference without handing the model lab a billing name — or handing Ethereum a record of what was asked.
The dAI team built the protocol with the Open Anonymity Project, implementing a February 2026 Ethereum Research proposal by Davide Crapis and Vitalik Buterin. It is labeled experimental. A user deposits ether or USDC into an on-chain vault in one transaction. That balance becomes a private "note." Client software then produces a Groth16 zero-knowledge proof — Poseidon hashes, a 32-level Merkle tree, nullifiers against double-spends — showing that some funded, unspent note covers the session, without revealing which one.
A zkAPI server checks the proof and issues a short-lived, dollar-capped API key, often compatible with OpenRouter or OpenAI-style endpoints. Prompts travel from the device to the model provider on that key. After the session, a signed usage receipt bills actual consumption against the private balance. Remaining funds can be withdrawn on-chain. The privacy split is deliberate: the zkAPI server sees that payment exists and session totals, not identity or prompts; the AI lab sees the query, not the payer; Ethereum sees deposits and withdrawals, not what any note bought. The Foundation is also pitching the same pattern for RPC access, VPNs, and agent-to-agent payments.
The second-order product is not a mixer. It is a checkout for machines that should not carry a KYC file. zkAPI does not claim full anonymity — IP addresses and timing can still correlate sessions — and the team recommends Tor for a harder trail. What it does claim is a clean cut between money and content. If agents are going to buy inference by the token, the Foundation would rather they not do it with a public wallet glued to every prompt. The vault is live. The experiment is whether labs will accept a key they cannot map to a customer.
Image source: i.ibb.co