Ethereum Foundation launches zkAPI for private AI payments

The Ethereum Foundation and the Open Anonymity Project put a tool called zkAPI live on Ethereum's main network on October 1, 2026. It lets a person pay for a metered online service, such as an AI model, without the provider learning who paid.

Vittorio Rivabella of the Foundation's dAI team described the release in a post on the Ethereum Foundation blog. The design traces to a February 2026 research note by Davide Crapis, who leads the Foundation's AI team, and co-founder Vitalik Buterin. They had sketched the idea on the Ethereum Research forum months earlier. The code shipped this week.

What zkAPI actually does

Most paid online services tie each request to an account. That account ties back to a card, an email, or a wallet. zkAPI breaks that chain. A person deposits money once into a contract, then draws against it in small amounts, and never hands over an identity to spend.

Think of a prepaid transit card bought with cash. The gate checks that the card holds credit and opens. It never learns the rider's name. zkAPI aims for that same split between paying and being seen, only for software that bills by use.

In its announcement, the Foundation summed up the split this way: "The provider sees the requests, and the payment layer sees the spend. Neither learns the link between them."

How the privacy works

A user locks ETH or USDC into a vault contract on Ethereum. The deposit is stored as a coded entry in a Merkle tree 32 levels deep, a structure that lets the system confirm an entry exists without exposing it. To spend, the user's software builds a zero-knowledge proof. That proof shows two things at once: that a funded deposit covers the charge, and that the deposit has not already been used.

zkAPI uses Groth16 proofs on the BN254 curve, with Poseidon hashing for the coded entries and the markers that block double spending. A server checks each proof off the chain and returns a short-lived key capped at a set dollar amount. The vault contract checks the same kind of proof when money moves in or out. None of it needs the user's name.

What zkAPI does not hide

Privacy here is partial, and the team says so plainly. zkAPI does not cover network data. A provider can still log an IP address and tie sessions together that way. Prompt wording can act as a fingerprint too. Two requests in the same style, sent minutes apart, can be matched even when the payment trail stays blank.

Timing is another leak. Requests that arrive in a steady rhythm can be grouped. The tool cuts one link in the chain, the one between a payment and a request. It does not claim to cut them all. Full anonymity would still need a private network connection layered on top.

Why private payments for AI matter

AI firms meter usage by the token. Every prompt sent to a model gets logged against a paying account, and that log can be subpoenaed, sold, or spilled in a breach. A lawyer researching a case, a reporter checking a tip, or a firm probing a rival's product may not want that record to exist at all.

zkAPI points at that problem without claiming to end it. The launch also tests whether Ethereum can work as a payment rail for ordinary software, not only for trading. Buterin has argued for years that privacy tools belong in everyday use. Whether AI providers wire zkAPI into their billing is the open question. None has announced support yet.

What to watch

Watch for the first AI provider to take zkAPI payments in production. That would be the signal the tool has moved past a demo. Watch too for an audit. The Foundation published the code and the method, but a launch post is not a security review, and contracts that hold deposits draw attackers. For now the tool is live, open, and modest in what it promises. The next few months decide whether anyone builds on it.