Conflicting tools give different Ethereum client shares

On Sept. 16 three commonly used tools produced divergent estimates of which Ethereum consensus client held the largest share. Blockprint’s panel on clientdiversity.org showed Teku at 99.83%. Miga Labs’ Ant crawler returned Lighthouse at 51.32%. Rated’s on-chain analysis reported Teku at 53.86%.

Client diversity affects how the Ethereum validator set reacts to software bugs and network events. Protocol guidance identifies two thresholds: a client used by more than one-third of validators can stop finality, and a client with two-thirds of voting weight could finalize an incorrect chain. Researchers measure stake-weighted distributions as the relevant metric for those risks.

The three tools use different signals and methods. Blockprint applies a machine-learning classifier that infers client behavior from block production patterns. Sigma Prime’s archived repository later indicated the classifier became inaccurate after the Electra upgrade and described the project as defunct, while the clientdiversity.org panel remained labeled as updated daily.

Miga Labs obtains client metadata by discovering peer nodes with its Ant crawler and requesting their reported client versions. That approach can miss nodes behind firewalls, nodes that refuse connections, peers that are not discovered, and peers that rotate identifiers.

Rated estimates client share by grouping validator keys by deposit address and mapping those groups to operators using on-chain transactions, block graffiti and voluntary disclosures. Rated’s documentation warns there is no standard method for higher-level operator attribution and that operator mapping depends on persistent public links.

A count of visible nodes does not equal staked voting power. One hosting node can run many validators, and distinct validator keys can share a single operator or hosting provider. Methods that sample nodes or keys therefore do not directly reveal how much stake supports each client observation.

Protocol changes and privacy proposals would reduce some public signals. A July research post proposed rebuilding the active validator registry daily, requiring fresh validator keys, and replacing some on-chain accounting with ZK-STARK proofs. The proposal also suggested hiding links between deposit and withdrawal addresses. The author described the change as creating “strong validator anonymity” and noted that large operations might still leak aggregate signals.

Some measurement techniques could still work if identifiers rotate. Peer-discovery tools and behavior-based classifiers can extract signals that do not rely on stable keys, although protocol or client changes can invalidate fingerprints, as occurred with Blockprint after Electra. A 2025 study found that four observer nodes located more than 15% of validators in the peer-to-peer network over a three-day window, demonstrating that network traces can reveal hosting concentration.

Researchers have explored cryptographic ways to publish authenticated aggregate measurements without exposing individual validators. One project examined encrypted voting where validators commit encrypted client choices, prove ballot validity, and enable a set of authorities to recover only the aggregate using homomorphic encryption, distributed key generation and zero-knowledge proofs. An IETF research draft on verifiable distributed aggregation describes related primitives for private sums and histograms. Open questions include authenticating stake, preventing fake reports, sampling, software attestation, selection of decryption authorities and performance. Multiplexed validator setups that run multiple clients complicate single-client labels.

Several teams have proposed that measurement become an explicit part of any privacy design. Suggested components include stake-authenticated reporting, verifiable aggregation of results, published uncertainty bounds around estimates, and separate methods to measure client share, operator concentration and stake distribution.