Vitalik Buterin printed a analysis proposal that avoids the query everybody retains asking: Are you able to run AI fashions on blockchain?
As a substitute, the examine argues that Ethereum is a privacy-preserving funds layer for pay-as-you-go AI and API utilization. This submit, co-authored with Davide Crapis on Ethereum analysis, argues that the actual alternative is to not put LLM on-chain.
The true alternative lies in constructing an infrastructure that enables brokers and customers to pay for hundreds of API calls with out compromising their identities or making a surveillance path with billing knowledge.
Timing is vital as agent AI is shifting from demonstration to enterprise roadmap. Gartner predicts that 40% of enterprise purposes will embrace task-specific AI brokers by the top of 2026, up from lower than 5% in 2025.
This shift hints at a world the place software program autonomously generates massive numbers of API calls and billing rails develop into strategic infrastructure fairly than back-office plumbing.
Present measurement programs pressure you to decide on between Web2 id billing, which depends on API keys and bank cards and exposes profiling knowledge, and on-chain pay-per-call fashions that hyperlink exercise via clear transaction graphs, that are too gradual and costly.
This proposal introduces ZK API Utilization Credit, a cost and abuse prevention primitive constructed on the speed restrict override performance.
RLN is a zero-knowledge gadget designed to stop spam on nameless programs, and analysis repurposes it for metered entry to providers.
The circulation proceeds as follows. Customers deposit funds as soon as into a wise contract and their dedication is added to the on-chain Merkle tree.
Every API request features a zero-knowledge proof that proves that the consumer is a sound depositor with ample belief for the requested index.
If a consumer makes an attempt to reuse the ticket index and spends twice the allowance, the RLN permits the system to get well the key and scale back the stake as a monetary penalty.
The submit accommodates concrete examples. Consumer deposits 100 USDC and runs 500 hosted LLM queries. One other consumer deposits 10 USDC for 10,000 Ethereum RPC calls.
The structure is explicitly designed for “many calls per deposit,” and on-chain exercise scales with the variety of accounts and settlement frequency, fairly than uncooked inference quantity.
Improve flexibility with variable price assist. Customers pay upfront the utmost payment per name, the server returns a signed refund ticket for the unused quantity, and customers accumulate refunds privately to unlock extra calls with out further deposits.
the infrastructure already exists
This proposal comes at a time when a big utilization credit score cost infrastructure already exists.
In response to DefiLlama, the stablecoin has a circulating market capitalization of roughly $307.6 billion, indicating that the on-chain greenback layer has ample liquidity to assist deposit-based billing for high-frequency providers.
Ethereum’s scaling stack has matured to the purpose the place Rollup handles far more exercise than Layer 1, with L2Beat exhibiting a scaling issue of round 100x, with Rollup processing hundreds of operations per second in comparison with Ethereum Mainnet’s hundreds of operations per second.
Lately, Ethereum’s common transaction payment was measured at round $0.21 on February seventh, suggesting that occasional on-chain measurement and cost flows are doable with out prohibitive prices.
This design explicitly avoids placing LLM on-chain. Ethereum competes not on TPU cycles or inference pace, however on impartial funds, programmable escrow, and verifiable execution.
This structure treats inference as an off-chain service and blockchain as a dependable layer for funds, measurement, and dispute decision. There is no such thing as a want for customers to belief or reveal their id to particular person suppliers.
Ethereum turns into the enforcement layer for AI commerce when AI service suppliers depend on Ethereum or layer 2 sensible contracts to just accept deposits, slash, refund, and adjudicate disputes.
This mannequin is much like how Ethereum turned a stablecoin and DeFi funds layer by offering a impartial basis on which financial agreements are enforced programmatically, fairly than internet hosting the whole utility stack on-chain.
A situation with out the hype
On-chain footprint is proscribed by settlement cadence, not uncooked name quantity.
Assume that 250,000 energy customers or brokers undertake utilization credit in a crypto-native wedge situation concentrating on RPC and infrastructure APIs.
If every performs two on-chain actions per 30 days (deposit or replenishment and withdrawal), roughly 500,000 transactions are generated month-to-month attributable to the railway.
In an AI supplier deployment situation, think about 1 million customers leveraging privacy-preserving credit throughout hosted LLM providers, however solely performing 1 to three on-chain actions per 30 days.
Which means that the 1 million to three million transactions per 30 days related to AI commerce rails could also be concentrated in Layer 2, the place it’s cheaper to execute.
In enterprise agent situations, the dimensions of deposits will increase, the danger of dependable execution will increase, and the significance of the slash mechanism will increase.
Metadata points
The proposal seeks to make funds unlinkable, however the analysis thread itself highlights potential weaknesses.
Commenters argue that even when nullifiers can’t be cryptographically linked, servers can affiliate customers via inference-based metadata akin to timing patterns, token counts, and cache hits.
This critique proposes mounted, bucketed pricing for enter and output courses to scale back leakage. The stress between cryptographic privateness and behavioral metadata is central as to if a design really achieves the objective of anonymity.
One other hurdle exists in precise implementation. Though this proposal makes use of RLN as a primitive, the Privateness and Scaling Explorations venture web page notes that RLN is inactive or deprecated.
Productizing ZK API Utilization Credit could require you to take care of a fork or implement a brand new answer fairly than counting on present instruments.
The RLNJS benchmark stories round 800ms for proof era and round 130ms for verification on M2 Macs, offering an early efficiency sanity verify, however leaving open questions on cell constraints and enormous production-grade circuits.
The proposal additionally envisions suppliers integrating deposit and proof flows, accepting stablecoin funds, and adopting Ethereum or Layer 2 contracts for dispute decision.
It isn’t only a technical drawback, it is a coordination drawback. Web2 API suppliers have present billing infrastructure and clear laws concerning identity-linked transactions.
To persuade them to undertake a ZK-based various, they might want to exhibit both a compelling price benefit or a differentiated market section the place privateness safety expenses can seize income they might not in any other case have.
| mannequin | Billing technique | Issues that leak or break | Somebody who fits you |
|---|---|---|---|
| Web2 ID billing (API key + card) | Account-based billing tied to your ID (API key + cost technique). Supplier centrally measures requests and invoices | leak: Id linkage and profiling path throughout requests. break: Pseudonym/self-custody norms. danger: Centralized administration (suspension/censorship, single supplier belief) | Mainstream SaaS/API suppliers. Firms that prioritize compliance, simplicity, and present billing rails |
| On-chain pay-per-call | Every request (or batch) pays on-chain per name by way of a transaction/sensible contract | break: Excessive name price/delay. leak: On-chain linkability (transaction graph ties utilization collectively). friction: UX overhead for repeated TX | Crypto-native providers which can be known as occasionally. When transparency/auditability is extra essential than privateness/throughput |
| ZK API utilization credit (one time deposit, limitless calls) | Customers deposit as soon as. Every request contains ZK proof of membership and remaining credit. Slash for double use. Non-obligatory refund ticket for variable bills | danger: Metadata correlation (timing/token patterns may be relinked). Burden: Supplier integration + collaboration. Maturity: ZK instruments/operational complexity, circuit upkeep | Excessive-frequency APIs (LLM, RPC, knowledge) the place privateness is a promoting level. Agent toolchain. Customers who require metering with out identity-based monitoring |
What this implies for Ethereum
If this design positive aspects traction, Ethereum’s worth proposition will additional shift towards serving as a impartial enforcement layer for digital commerce fairly than a general-purpose computing platform.
The proposal treats blockchain not as a spot the place purposes run, however as a cost infrastructure that ensures financial guidelines are enforced.
The rate of stablecoins will increase as deposits circulation into utilized credit score contracts, doubtlessly creating a brand new class of on-chain financial exercise distinct from DeFi hypothesis and NFT buying and selling.
Layer 2 utilization is more likely to enhance as suppliers and customers resolve disputes, course of refunds, and deal with thrashing occasions in a throughput-optimized chain.
The query is whether or not a parallel ecosystem will emerge the place privateness safety expenses develop into a prerequisite for sure consumer segments.
Companies involved about knowledge leakage via billing logs, builders constructing agent toolchains that require unsupervised and auditable instrumentation, and energy customers who worth nameless entry to high-frequency providers are all potential early adopters.
Ethereum’s alternative is to function a layer on which the AI providers market can settle with out individuals having to belief particular person platforms or sacrifice the privateness of their billing infrastructure.
The proposal claims that Ethereum can implement cost agreements, adjudicate disputes, and allow pay-as-you-go entry with out id federation in methods which can be structurally not doable with conventional programs.
The success of this argument relies on fixing the metadata correlation drawback, sustaining sturdy ZK implementations, and convincing suppliers that the market-unlocked integration prices are justified.
