Decentralized AI on Blockchain: DFINITY, SingularityNET, Fetch.ai & Ocean Protocol

By · Updated

Blockchain is useful around AI when it provides auditable payments, model or dataset commitments, and governance among parties that do not share an operator. It is usually a poor place to train a large model. Claims of "decentralized AI" must state where inference runs, who can reproduce it, how private data is protected, and whether a token incentive measures quality or merely participation.

Separate compute, verification, and settlement

“Decentralised AI” can refer to unrelated designs: a marketplace pays independent model providers, a ledger records dataset licences, a protocol aggregates training contributions, or replicated nodes execute the same inference. DFINITY Foundation develops ICP, while the network is operated and governed through its own nodes and protocol. Assess the specific execution and governance path rather than treating the organisation and network as interchangeable.

DFINITY was founded by Dominic Williams, and the Internet Computer mainnet launched in May 2021. ICP canisters execute WebAssembly and serve certified content under subnet replication. Chain-key cryptography is a suite of threshold protocols used for certification, signing, cross-subnet functions, and related system operations; it is not itself “the consensus mechanism.” Consensus is a separate layer that orders messages for a subnet.

A chain can commit to model hashes, record licenses and payments, and coordinate providers. Those records do not prove that a hidden accelerator ran the promised model. Reproducible inference, zero-knowledge proofs, trusted execution environments, or replicated deterministic execution provide different assurance and cost. AI can assist monitoring or audits, but probabilistic output must not participate directly in consensus.

Marketplaces and the ASI consolidation

SingularityNET has published marketplace and platform components for discovering and paying for model services. Its organisation, contracts, supported networks, and token migration must be checked independently: a marketplace listing does not attest to model quality, training rights, privacy, availability, or a provider's ability to reproduce an output.

Fetch.ai focuses on software agents and a network for agent services. In 2024 it joined SingularityNET and Ocean Protocol in the Artificial Superintelligence Alliance and began a staged token consolidation around FET/ASI branding. The participating organisations remain separate entities, and migration state has changed over time; integrations should use current official contracts, chain identifiers, and governance material.

Ocean Protocol has explored data-market and “compute-to-data” patterns in which an algorithm runs near a controlled dataset. This can reduce direct distribution of raw data, but the data host, execution environment, output review, access policy, and code still form a trust boundary. It does not by itself prove consent, prevent extraction, satisfy deletion duties, or make training lawful.

Federated learning keeps raw examples at participating sites and exchanges model updates, but those updates can leak information or be poisoned. Adding a ledger can order contributions and payments; it cannot score model quality objectively without a benchmark, evaluator, or governance process. Systems such as Bittensor should therefore be assessed by validator incentives, collusion resistance, reproducibility, and the cost of gaming their reward signal.

On-chain inference on ICP

ICP can run deterministic inference that fits canister memory, instruction, replication, and cycles limits, and its ecosystem documents model-serving experiments. Larger models may require specialised system services or external compute. An “on-chain” claim should disclose model bytes, subnet replication, determinism, accelerator or API dependencies, update authority, latency, and cost per verified result.

Privacy and a deployment decision

Zero-knowledge proofs, secure multiparty computation, trusted execution, homomorphic encryption, and federated learning provide different confidentiality and verification properties. Combining any of them with a blockchain does not make arbitrary model training private. Specify the protected input, information leaked by outputs and metadata, key holder, proof statement, deletion path, and failure response.

Benchmark the proposal against a conventional signed API. Measure end-to-end latency and cost, reproduce outputs, test provider disappearance, and define data-deletion, model-update, appeal, and liability rules. Decentralised coordination can broaden provider choice or make payments auditable, but it does not guarantee model quality, privacy, fairness, or resistance to collusion.

Primary references

AI, ML, Dfinity, Blockchain, Governance

Published · Updated