Vitalik Buterin simply revealed a analysis proposal that sidesteps the query everybody retains asking: can blockchains run AI fashions?
As a substitute, the analysis claims Ethereum because the privacy-preserving settlement layer for metered AI and API utilization. The put up, co-authored with Davide Crapis on Ethereum Analysis, argues that the actual alternative is not placing LLMs on-chain.
The actual alternative lies in constructing the infrastructure that allows brokers and customers to pay for 1000’s of API calls with out compromising id or creating surveillance trails by way of billing knowledge.
The timing is important as a result of agentic AI is shifting from demonstrations to enterprise roadmaps. Gartner forecasts that 40% of enterprise functions will embrace task-specific AI brokers by the top of 2026, up from below 5% in 2025.
That shift implies a world during which software program autonomously generates large volumes of API calls, making billing rails strategic infrastructure moderately than back-office plumbing.
Present metering programs power a selection between Web2 id billing, which depends on API keys and bank cards and leaks profiling knowledge, and on-chain pay-per-call fashions which are too sluggish, too costly, and hyperlink exercise by way of clear transaction graphs.
The proposal introduces ZK API utilization credit, a fee and anti-abuse primitive constructed on Price-Limiting Nullifiers.
RLN is a zero-knowledge gadget designed to forestall spam in nameless programs, and the analysis repurposes it for metered entry to companies.
The stream proceeds as follows: customers deposit funds as soon as into a wise contract, and their dedication is added to an on-chain Merkle tree.
Every API request features a zero-knowledge proof demonstrating that the person is a sound depositor with ample credit score for the requested index.
If a person makes an attempt to reuse a ticket index, double-spending their allowance, RLN permits the system to get well their secret and slash their stake as an financial penalty.
The put up consists of concrete examples. A person deposits 100 USDC and makes 500 hosted LLM queries. One other deposits 10 USDC for 10,000 Ethereum RPC calls.
The structure is explicitly designed for “many calls per deposit,” that means that on-chain exercise scales with the variety of accounts and settlement frequency moderately than uncooked inference quantity.
Variable-cost assist provides flexibility: customers prepay a most value per name, servers return signed refund tickets for unused quantities, and customers privately accumulate refunds to unlock extra calls with out extra deposits.
Infrastructure is already there
The proposal arrives when the fee substrate for utilization credit already exists at scale.
Stablecoins have a circulating market cap of roughly $307.6 billion, in keeping with DefiLlama, indicating that the on-chain greenback layer is sufficiently liquid to assist deposit-based billing for high-frequency companies.
Ethereum’s scaling stack has matured to the purpose the place rollups course of way more exercise than layer-1, with L2Beat displaying a roughly 100x scaling issue, with rollups dealing with 1000’s of operations per second in comparison with tens on the Ethereum mainnet.
Common Ethereum transaction charges not too long ago measured round $0.21 on Feb. 7, suggesting that occasional on-chain metering and settlement flows are possible with out prohibitive value.
The design explicitly avoids placing LLMs on-chain. Ethereum competes on impartial settlement, programmable escrow, and verifiable enforcement, not TPU cycles or inference velocity.
The structure treats inference as an off-chain service and the blockchain because the layer that makes fee, metering, and dispute decision credible, with out requiring customers to belief particular person suppliers or to disclose their identities.
If AI service suppliers settle for deposits and depend on Ethereum or layer 2 sensible contracts to adjudicate slashing, refunds, and disputes, Ethereum turns into the enforcement layer for AI commerce.
The mannequin parallels how Ethereum turned the settlement layer for stablecoins and DeFi, not by internet hosting the complete software stack on-chain, however by offering a impartial substrate the place financial agreements are enforced programmatically.
Eventualities with out hype
The on-chain footprint is bounded by settlement cadence, not uncooked name quantity.
In a crypto-native wedge state of affairs concentrating on RPC and infrastructure APIs, suppose 250,000 energy customers or brokers undertake utilization credit.
If every performs two on-chain actions monthly, a deposit or top-up plus a withdrawal, that generates roughly 500,000 transactions month-to-month attributable to the rail.
In an AI supplier adoption state of affairs, think about a million customers make use of privacy-preserving credit throughout hosted LLM companies however nonetheless carry out just one to 3 on-chain actions month-to-month.
That means a million to 3 million transactions monthly tied to AI commerce rails, doubtless focused on layer 2s the place execution is cheaper.
Enterprise agent eventualities enhance deposit sizes, elevating the stakes for credible enforcement and making slashing mechanisms extra consequential.
The metadata downside
The proposal tries to make funds unlinkable, however the analysis thread itself highlights a possible weak point.
A commenter argues that even when nullifiers are cryptographically unlinkable, servers can correlate customers by way of inference-based metadata corresponding to timing patterns, token counts, and cache hits.
The critique proposes bucketed pricing, with fastened enter and output courses, to scale back leakage. That rigidity between cryptographic privateness and behavioral metadata is central as to if the design truly delivers on its anonymity targets.
Implementation actuality presents one other hurdle. The proposal makes use of RLN as a primitive, however the Privateness and Scaling Explorations challenge web page notes that RLN is inactive or has been sundown.
Productionizing ZK API utilization credit doubtless requires sustaining forks or implementing new options moderately than counting on present tooling.
RLNJS benchmarks report roughly 800 milliseconds for proof technology and 130 milliseconds for verification on an M2 Mac, offering an early sanity examine on efficiency however leaving open questions on cellular constraints and production-grade circuits at scale.
The proposal additionally assumes that suppliers will combine the deposit-and-proof stream, settle for stablecoin settlements, and undertake Ethereum or layer 2 contracts for dispute decision.
That is a coordination downside, not only a technical one. Web2 API suppliers have present billing infrastructure and regulatory readability round identity-linked transactions.
Convincing them to undertake a ZK-based various requires demonstrating both a compelling value benefit or a differentiated market phase during which privacy-preserving billing unlocks income they might not in any other case seize.
ModelHow it billsWhat it leaks/breaksWho it suitsWeb2 id billing (API keys + playing cards)Account-based billing tied to id (API key + fee methodology); supplier meters requests and invoices centrallyLeaks: id linkage + profiling trails throughout requests. Breaks: pseudonymity/self-custody norms. Threat: centralized management (suspension/censorship, single-provider belief)Mainstream SaaS/API suppliers; enterprises prioritizing compliance, simplicity, and present billing railsOnchain pay-per-callEach request (or batch) pays onchain per name by way of transactions/sensible contractsBreaks: value/latency for high-frequency calls. Leaks: onchain linkability (transaction graph ties utilization collectively). Friction: UX overhead for repeated txsCrypto-native companies with low name frequency; circumstances the place transparency/auditability is extra necessary than privateness/throughputZK API utilization credit (deposit as soon as, many calls)Consumer deposits as soon as; every request carries a ZK proof of membership + remaining credit score; slashing for double-use; optionally available refund tickets for variable costRisk: metadata correlation (timing/token patterns can re-link). Burden: supplier integration + coordination. Maturity: ZK tooling/ops complexity, circuit maintenanceHigh-frequency APIs (LLMs, RPC, knowledge) the place privateness is a promoting level; agent toolchains; customers needing metering with out identity-based surveillance
What this implies for Ethereum
If the design beneficial properties traction, Ethereum’s worth proposition shifts additional towards serving as a impartial enforcement layer for digital commerce moderately than a general-purpose computing platform.
The proposal treats blockchain because the settlement substrate the place financial guidelines get enforced credibly, not the place the place functions run.
Stablecoin velocity may rise as deposits stream into utilization credit score contracts, creating a brand new class of on-chain financial exercise distinct from DeFi hypothesis or NFT buying and selling.
Layer 2 utilization may enhance as suppliers and customers resolve disputes, course of refunds, and deal with slashing occasions on throughput-optimized chains.
The query is whether or not a parallel ecosystem emerges during which privacy-preserving billing turns into a prerequisite for sure person segments.
Enterprises involved about knowledge leakage by way of billing logs, builders constructing agent toolchains that require auditable metering with out surveillance, and energy customers who worth pseudonymous entry to high-frequency companies are all potential early adopters.
Ethereum’s alternative is to function the layer on which AI service markets settle, with out requiring individuals to belief particular person platforms or to sacrifice privateness to billing infrastructure.
The proposal claims Ethereum can implement fee agreements, adjudicate disputes, and allow metered entry with out id linkage in ways in which conventional programs structurally can’t.
Whether or not that declare holds relies on fixing the metadata correlation downside, sustaining sturdy ZK implementations, and convincing suppliers that the market justifies the combination value it unlocks.
