The case for putting AI on a blockchain is not decentralisation for its own sake. It is four specific gaps in centralised AI that happen to map onto four things a blockchain already does well — and the honest assessment is that only two of those four are close to working at scale.
What a chain actually adds
Verifiability. You can prove that a given output came from a specific model version. This matters most where the inference decision has legal or financial consequences, and it is the hardest of the four to deliver cheaply.
Incentives. Compute providers, data contributors, and model trainers can be paid in a shared unit without a central payroll or a bank account. This works today, and it is the least controversial part of the thesis.
Composability. A model output becomes an on-chain asset other contracts can consume directly, so an agent's decision can settle a trade without a human in the loop.
Censorship resistance. No single provider can revoke access to a model or a training run.
Compute markets are the mature part
Selling GPU time by auction works, and the mechanism is straightforward: providers list capacity, buyers bid or pay per GPU-hour, and settlement happens on-chain. Akash sells general compute with a reverse auction for bids. Render targets rendering and increasingly general GPU work. Filecoin's storage market is the same pattern applied to disk.
The limits are not conceptual. Decentralised markets are worse at the thing that matters most to a customer — predictable latency and a support contract when something breaks. They win where the workload is batchable, interruption-tolerant, and cost-sensitive, which is a real but narrow slice of the market.
Verifiable inference is where the cost lands
Two approaches, both expensive.
Zero-knowledge proofs wrap the computation so a verifier can check it without re-running it. The overhead is the problem: proving a model's forward pass costs somewhere between ten and a hundred times the raw inference, which is fine for an audit trail and prohibitive for a production endpoint.
Optimistic challenge games assume the output is correct and allow anyone to dispute it within a window, with the dispute resolved by re-execution. Cheap in the common case, and it introduces a fraud window during which the wrong answer is the accepted one. Whether that is acceptable depends entirely on what the output controls.
Agents are the interesting frontier, and the least proven
Autonomous programs holding a wallet and transacting — Eliza from ai16z, GAME from Virtuals, and a long tail of frameworks — are the most active area in the space. The technical pattern is unremarkable: an agent reads state, decides, and calls a contract.
What is unproven is whether agent token economics persist. An agent that generates revenue has a real claim on value; an agent whose token exists to fund the development of the agent has a claim on a fundraise. The two look identical in a dashboard.
Decentralised training is the hardest and most speculative
Bittensor coordinates specialised subnets — text, image, audio — where miners produce model outputs and validators score them, with consensus determining the ranking and TAO emissions, roughly 7,200 per day, distributed to participants. Gensyn and similar efforts attempt to coordinate many machines on one training job.
Two problems are structural. Verification of a training step is much harder than verification of inference, because the computation is enormous and the output is a set of weights rather than a prediction. And sybil resistance in training is weak: fake nodes contributing nothing useful while collecting rewards is a live attack, and it scales with the openness of the network.
What goes wrong
Model drift means an on-chain model degrades as the world changes and nobody has an incentive to retrain it. Proof overhead caps what can be proven economically today. Latency makes decentralised inference a poor fit for interactive products. Token economics attach a volatile asset to a fixed service, so an agent whose costs are in dollars and revenue is in a token is short volatility whether it wants to be or not. Compute cartelisation is a plausible outcome of any open market where capital buys advantage.
And the regulatory load is arriving: the EU AI Act and successive US executive orders create compliance obligations that a permissionless network is structurally awkward at satisfying, while tokenised models invite securities scrutiny.
The honest read
Decentralised compute and incentive mechanisms work. Verifiable inference works at a cost that limits it to high-stakes, low-volume use cases. Decentralised training is an open research problem. If you are building in this space, the question worth asking about any project is which of the four properties it actually needs — because several of them can be bought more cheaply from a centralised vendor whose failure modes you at least understand.
References
Next, read Decentralized Oracles 101 to understand how AI models get trustworthy data in the first place. More in Web3 Applications.