Inside zkAPI: Ethereum's AI Billing Layer, Its 30-Day Decay, and the Pause Switch
On September 30, the treasury contract went live. On October 1, the Ethereum Foundation announced it. One day of silence between deployment and disclosure โ and in that gap, the mechanism that matters most was already coded: expired deposits flow to the treasury. Not refunded. Not burned. Routed. If you are an options trader, you recognize the shape of this immediately. It is a decay curve with a counterparty on the other side, and the counterparty wrote the contract.
The system is zkAPI โ an on-chain billing layer for AI API calls, designed by Vitalik Buterin and Davide Crapis and built with a partner calling itself Open Anonymity. It runs on Ethereum mainnet. The premise is simple and, frankly, overdue: AI APIs bill by the request, and that billing currently lives inside a Stripe-shaped black box. zkAPI moves the meter on-chain, wraps it in ZK privacy, and settles on L1.
The architecture is a prepaid state channel. You deposit funds; the deposit becomes a note. Your wallet keeps the private spending state locally. Each API call prompts the server to sign a successor state โ your remaining balance. No per-request transaction touches the chain. Settlement is deferred.
This is not new. It is Lightning with a ZK skin. Terra's code was poetry; Luna's exit was prose. The difference here is who holds the pen.
Zoom out and the strategic picture clarifies. Solana's AI payment system claims a million transactions per second; zkAPI claims something harder to market โ a way to get your money back when the seller stops cooperating. Two chains, two answers to the same question: who holds the risk in machine-to-machine payments? Solana optimizes for throughput. Ethereum is betting that the binding constraint in the AI agent economy is not speed but settlement finality and dispute resolution. That is a defensible thesis. It is also a slow one.
Two exit routes exist. Mutual close, which requires server cooperation and is fast. And escape withdrawal, which is permissionless โ but subject to a 24-hour challenge window. Read that number again. Twenty-four hours. Typical optimistic rollups give you seven days. This gives you one. A nullifier prevents state reuse, and the treasury keeps an active root of request history so challenges can be verified against it.
Now the part nobody puts in the headline. Escape withdrawal does not let you assert any balance you like. You must present the correct spending state โ a cryptographic proof that the funds are yours and unspent. Lose that local state, and the chain has no way to know you ever existed. This is the inherited wound of every state channel: the state is the money.
When the billing server goes dark, mutual close becomes impossible and you are forced onto the escape route โ which is exactly the scenario the design anticipates. That is genuinely better than a pure IOU model. But it also means your exit latency is measured in challenge windows, not seconds. For a high-frequency consumer of AI compute, that latency is a working-capital cost.
The lifecycle is 30 days. The treasury code rounds deposit time and lifecycle up to a day boundary โ a subtle asymmetry that can expire a note up to a day early. Small, but deliberate. Engineering choices leave fingerprints.
Here is the discrepancy that should stop you cold. The announcement says the treasury holds USDC credit. The mainnet manifest reads native ETH, accounted in integer gwei. The block explorer prices it in ETH. Either the disclosure is loose, or there are multiple deployments, or the asset migrated mid-build. Any of those is a due-diligence signal.
I have audited this shape before. In 2017 I forked TokenSale contracts to prove reentrancy to founders who preferred the whitepaper to the proof. The lesson has not changed: the code tells you who gets paid when the music stops. Here, the music stops at day 30.
The pause switch is the other one. The vault owner can halt deposits, mutual closes, and the initiation of new escape withdrawals. Pending finalizations, active challenges, and expiry claims sit outside that gate. So the owner cannot reach into a withdrawal already in flight โ but can stop you from starting one. Risk isn't the volatility you see; it's the liquidity you cannot.
The challenge mechanism carries no economic penalty. A challenge costs the challenger nothing and pays them nothing. Free to file, free to be wrong. That is a game-theory hole: you have built a watchtower nobody is paid to staff.
The narrative sells "no server permission required." The code says permission is required from a different party. That is not a scandal โ it is a design boundary, and the boundary is drawn carefully. But retail will read the marketing, deposit, and treat the escape hatch as guaranteed. Smart money will read the contract, note the 24-hour window, note the local-state dependency, and price the operational risk accordingly. The gap between those two readings is where capital gets transferred. The spread is the gap between belief and reality.
And understand what is actually being transferred. This system does not eliminate trust in the billing server. It relocates it. The provider can still mis-meter your usage, and there is no on-chain enforcement against overbilling. You have decentralized the settlement rail while leaving the meter in the same hands. That is progress, not absolution.
Note what is absent: there is no token. No presale, no points program, no governance coin. From a regulatory standpoint that is elegant โ the Howey test has almost nothing to grab. From an adoption standpoint it is a handicap. There is no speculative carrot to pull users onto a rail that requires them to manage cryptographic state. Good for compliance, bad for growth.
Watch the treasury balance. If expired deposits start accumulating, you are watching a user-experience failure in real time โ not a hack, just a thousand people who forgot to manage a state file. The real question is not whether zkAPI is decentralized. It is whether the AI agent economy will accept a settlement layer where forgetting your keys means funding someone else's treasury. My bet: the technology is sound and the incentives are not. Options don't lie about decay. Neither does a vault with a calendar.