Ethereum cannot run a modern AI model on its own chain. A smart contract is limited to a maximum size of 24 KB, every operation costs gas, and the network is built so that every node reaches the same result from the same input. AI and Ethereum therefore meet at the edges: models run off-chain, while the chain can hold identities, records, and rules. One example is ERC-8004, a draft standard for registering and rating AI agents.
Key takeaways
- According to Ethereum Foundation, a smart contract can be at most 24 KB, and contracts cannot fetch offchain data without an oracle.
- The Ethereum Virtual Machine produces a deterministic output from a given input, and each operation is metered in gas.
- ERC-8004, "Trustless Agents," lists its status as Draft, so it is a proposal and not a finished standard.
- A registry entry or a reputation score on a blockchain does not prove an AI agent works as advertised, and the draft itself says so.
- This article is general information about technology. It gives no investment advice and no view on any token.
Why a blockchain is a poor place to run a model
Ethereum is described by its own documentation as a distributed state machine. The Ethereum Virtual Machine (EVM) executes code consistently across all nodes, and the state transition function takes an old state and a set of transactions and returns a new state. Given an input, it produces a deterministic output. Every node must be able to repeat the computation and agree on the answer.
That design clashes with how most AI systems are used. Large models run on specialized hardware, often produce different text from the same prompt when sampling is used, and rely on very large files of numbers. Each of these traits is a problem for a network where many independent nodes must replay the same step and arrive at the same result.
Gas, size limits, and a back-of-the-envelope comparison
The Ethereum Foundation gas page explains that gas measures the computational effort of an operation, that a fee is charged for every computation to deter spam and infinite loops, and that the fee is paid whether or not a transaction succeeds. A plain ETH transfer uses 21,000 units of gas. The page works through an example in which the base fee is 10 gwei and the tip is 2 gwei, so the fee is 21,000 times 12, or 252,000 gwei (0.000252 ETH). That is a sum for the simplest possible transaction.
The size limit makes the gap clearer. The following is an illustrative example, not a measurement of any product. A model with seven billion parameters stored at 16 bits per parameter needs 14 billion bytes, or 14 GB. A contract capped at 24 KB (24,576 bytes) is about 570,000 times smaller than those weights. Even before gas is counted, the weights could not be stored in one contract.
What the chain can hold instead
| Task | Runs on-chain (EVM) | Runs off-chain |
|---|---|---|
| Model inference (text, images, predictions) | Not practical, because of size, cost, and determinism | The model runs on ordinary servers or the user's device |
| Rules such as "pay when condition X is confirmed" | Yes, as a smart contract | Evidence about condition X comes through an oracle |
| Identity of a service or agent | A registry entry that points to a description file | The description file and the service itself |
| Proof that a result was checked | Records of validation requests and responses | The checking work, such as re-running a job |
Ethereum's documentation says smart contracts cannot retrieve data from offchain sources by design, because relying on outside information could jeopardize consensus. Oracles are the tools that bring such data in. An AI output is offchain data, so any contract that acts on a model's answer depends on whoever supplies it.
What ERC-8004 proposes
ERC-8004 is listed on the Ethereum Improvement Proposals site as a Draft Standards Track ERC, created August 13, 2025. Its authors include contributors from MetaMask, the Ethereum Foundation, Google, and Coinbase. Its abstract describes using blockchains to discover, choose, and interact with agents across organizational boundaries without pre-existing trust. It defines three registries:
- An Identity Registry, based on ERC-721, that gives each agent a portable identifier resolving to a registration file.
- A Reputation Registry, an interface for posting and fetching feedback signals.
- A Validation Registry, with hooks for independent checks such as stakers re-running a job, zero-knowledge machine learning verifiers, or trusted execution environment oracles.
Payments are described as outside the scope of the protocol. The draft's own security section warns that Sybil attacks can inflate the reputation of fake agents, and that the standard cannot cryptographically guarantee that advertised capabilities are functional and non-malicious. Because the status is Draft, details may change and nothing here confirms real-world adoption.
Checks before believing an "Ethereum AI" claim
- Where does the model run, and who operates that server?
- What exactly is recorded on the chain: the output, a hash of it, or only a payment?
- Is the standard cited final, or a draft or proposal?
- Who can change the model, and does the chain record that change?
- Does the claim describe a working product, or a plan?
Readers weighing financial products should treat any page that mixes AI with a token as a marketing claim until checked against regulator guidance, such as that published by the securities and consumer protection authorities in their province or state. Nothing on this page is financial advice.
FAQ
Can an AI model run inside an Ethereum smart contract?
Not in a practical sense. The contract size limit is 24 KB according to Ethereum Foundation, and every step costs gas, so models are run off-chain and the chain holds records or rules.
What does a smart contract need to use real-world data?
An oracle. The Ethereum documentation says contracts cannot get information about real-world events by default, and oracles ingest offchain data and make it available to them.
Is ERC-8004 an approved standard?
The proposal's page lists its status as Draft as of October 24, 2026. A draft can change or be abandoned, so it should be read as a proposal.
Does recording an AI agent on a blockchain make it trustworthy?
No. A registry can make an identity and its feedback public, but the draft says it cannot guarantee that an agent's advertised capabilities work or are harmless.
Related reading
- A decentralized AI network on TON
- Another look at where AI technology is heading
- How an asset manager describes its AI use
- AI and robotics in society
More in AI Basics.
Last verified: October 2026
Sources
- Introduction to smart contracts, Ethereum Foundation developer documentation, 2026
- Ethereum Virtual Machine (EVM), Ethereum Foundation developer documentation, 2026
- Gas and fees: technical overview, Ethereum Foundation developer documentation, 2026
- Oracles, Ethereum Foundation developer documentation, 2026
- De Rossi M, Crapis D, Ellis J, Reppel E, ERC-8004: Trustless Agents [Draft], Ethereum Improvement Proposals, August 2025