No cryptocurrency can be called the best for AI integration, and the label itself is misleading, because large AI models cannot run inside a blockchain's own computation. On Ethereum, every validating node repeats each transaction, a contract's code is capped at 24,576 bytes, and a block holds at most 60 million units of gas. Projects that advertise AI therefore run the models off the chain and connect the results back, which brings in the trust problem that blockchains were designed to avoid.
Key takeaways
- Ethereum's documentation defines gas as the unit measuring the computational effort of an operation, and a plain transfer costs 21,000 units.
- A contract's deployed code is limited to 24,576 bytes under EIP-170, while GPT-3, a 2020 language model, has 175 billion parameters.
- Smart contracts cannot read outside data by default, so AI output has to arrive through an oracle, and oracle operators must be trusted.
- "AI integration" in a crypto project can mean very different things, and each meaning can be checked separately.
- This page ranks nothing and gives no investment advice.
Why a blockchain is a poor place to run a model
A public blockchain works because many computers reach the same answer from the same inputs. The Ethereum documentation on gas describes gas as the unit that measures the computational effort of an operation, with a fee equal to the gas used multiplied by the price per unit. A simple ETH transfer needs 21,000 units. The page on blocks lists a target of 30 million gas per block and a maximum of 60 million, which the network's validators can adjust only in small steps from one block to the next.
Because the limit is shared by all users, computation is rationed on purpose. Dividing 60 million by 21,000 shows that a full block could hold at most about 2,857 plain transfers if it held nothing else. A model that needs billions of arithmetic steps for one answer has no room in that budget, and every node would have to repeat the work.
Size is the second barrier
EIP-170, a finalized Ethereum standard created on November 4, 2016, caps contract code at 24,576 bytes. Trained models are far larger. The paper introducing GPT-3 reports 175 billion parameters. An illustrative calculation, assuming two bytes per parameter, gives about 350 billion bytes (350 gigabytes) for the weights alone. That is roughly 14 million times the contract size limit. Real systems compress, split, or shrink models, so the figure shows the order of magnitude, not a design.
Smaller models and simple decision rules are a different case. A short rule written in a contract, such as a fixed formula, runs fine on-chain, but it is ordinary code and not a trained model.
The oracle route and its trust cost
The usual workaround is to run the model on ordinary servers and send the result to a contract. Ethereum's documentation on oracles explains why this needs a special mechanism: contracts cannot by default access information stored outside the network, since outside data could differ between nodes and make agreement impossible. The same page states that having to trust oracle operators undermines the trustless property, that paying an oracle does not guarantee honesty, and that decentralized designs reduce these risks without removing them.
For AI the problem is sharper. The question is no longer only whether a price feed is accurate, but whether the model that produced an answer was the one advertised, ran on the stated input, and was not altered. A token or network can coordinate payments for such work, yet it cannot by itself prove that the work was done well.
What "AI integration" can mean in a project description
| Claim in a description | What it usually means | Question to ask |
|---|---|---|
| "Runs AI on-chain" | A small rule or a hash of an off-chain result is stored on the chain | What exactly executes inside the contract, and how big is it? |
| "Decentralized AI network" | Computers offer processing time and are paid in a token | Who checks that results are correct? |
| "AI agent with a wallet" | A program signs transactions on its own | Who is liable for a mistaken or fraudulent payment? |
| "AI-powered trading" | Marketing language | Is the seller registered, and where is the track record audited? |
The last row connects to regulator warnings covered in the article on how AI and cryptocurrency differ. A technical-sounding description is not evidence that anything works.
Documents worth reading before trusting a description
- The project's technical paper or documentation, looking for what runs on the chain and what runs elsewhere.
- Any independent security audit, including its date and the version of the code it covered.
- The terms that say who operates the oracle or compute network and how errors are handled.
- Regulator registers and warnings in the reader's own province, state, or country.
This article is general information about how blockchains work. It does not recommend, compare, or rate any token, exchange, wallet, or product, and it makes no forecast about prices or returns. Digital assets can lose their full value, and transfers are usually irreversible, so questions about personal finances belong with a registered adviser and the regulator's own pages.
FAQ
Can an AI model be stored on a blockchain?
Not within the limits of a contract on Ethereum. The code cap is 24,576 bytes, so large models are kept elsewhere, and only a reference or a result is recorded on the chain.
Does a blockchain make AI output more trustworthy?
Recording an output makes it hard to alter later, but it does not show that the output was correct. Accuracy depends on the model, the input, and the oracle that delivered the answer.
Why do gas limits matter for AI?
Gas caps how much computation fits in each block. Work that needs billions of operations cannot fit in a budget shared by all users.
Is there an official list of cryptocurrencies that support AI?
No regulator or standards body maintains one. Any such list is a private opinion and may carry commercial interests.
Related reading
- AI and cryptocurrency compared
- Pitfalls of cryptocurrency
- Concerns around Bitcoin
- How vectors work in AI
More in AI in finance.
Last verified: October 2026
Sources
- Gas and fees, Ethereum developer documentation, Ethereum Foundation, 2026
- Blocks, Ethereum developer documentation, Ethereum Foundation, 2026
- Oracles, Ethereum developer documentation, Ethereum Foundation, 2026
- EIP-170: Contract code size limit, Ethereum Improvement Proposals, 2016
- Brown et al., Language Models are Few-Shot Learners, arXiv, 2020