Data availability: where a rollup publishes its evidence

An optimistic rollup posts a new state and lets anyone dispute it with a fraud proof. Building that proof requires the transactions that produced the state, so the challenger can re-execute them and show the result differs. If the operator publishes the state but withholds the transactions, nobody can construct the proof, and an incorrect state goes unchallenged.

A validity rollup using zero-knowledge proofs is in a slightly different position, because the proof establishes correctness without re-execution. But a user who wants to exit, or anyone who needs to reconstruct the current balances, still needs the data, so withholding it means the state is provably correct and nobody can act on it. In both designs the data has to be available to someone other than the operator, and that is the whole subject.

Before EIP-4844, a rollup published its data as ordinary Ethereum transaction calldata, which competed for the same block space as everything else and was stored forever by every node, as the EIP itself sets out. That made publishing expensive and dominated a rollup's costs.

EIP-4844 introduced blob transactions: a separate data space with its own fee market, committed to with KZG commitments, and pruned by nodes after a few weeks because it is needed only long enough for verification and dispute. The effect on cost was large. After blobs went live, median fees on rollups such as Arbitrum, Optimism and Base dropped from tens of cents to fractions of a cent, a reduction of roughly a hundredfold. The layer 2 blockchains page covers the rollup designs that sit on top, and Arbitrum, Base and Polygon the individual networks.

Data availability sampling lets a light node establish that a block's data was published without downloading it. The data is first expanded with erasure coding, so that any sufficient fraction of the pieces reconstructs the whole, and the node then requests a few pieces at random. If the pieces come back, the data is almost certainly there; if a publisher withheld part of it, random sampling finds the gap with high probability. KZG commitments are used to prove the encoding itself was done correctly, which matters because a publisher could otherwise encode wrongly and pass sampling.

The committee is the cheaper and weaker alternative. A data availability committee is a named set of parties that signs an attestation that it holds the data, and the rollup's contract accepts those signatures. It costs almost nothing and its security is the honesty and availability of the committee, so a rollup using one has an external trusted party that a sampling-based design does not. Ethereum's own blobs, notably, do not yet use sampling: verifying availability there still means running a full node that downloads the data.

Celestia is a chain built for data availability alone, with sampling from launch, Tendermint finality in around 15 seconds, and fraud proofs for incorrect encoding. In January 2026 it reached one terabit per second of data throughput across 498 nodes, and its Matcha upgrade raised the maximum block size from 8 to 128 megabytes. Avail also supports sampling and uses KZG validity proofs, with finality of around 20 seconds under GRANDPA.

Ethereum with blobs has the longest finality, 12 to 15 minutes under Casper, no sampling, and KZG validity proofs. EigenDA inherits Ethereum's finality and uses KZG validity proofs. The trade is consistent: using Ethereum for availability means the rollup's data inherits Ethereum's security and costs more, while an external layer is cheaper and faster and adds a second system whose failure is now a dependency.

Three things, and each maps to a concrete obligation. The ability to exit without permission: if a user cannot reconstruct their balance from published data, leaving the rollup depends on the operator's cooperation, which makes the holding a claim against an operator. The ability to verify independently: an institution that reports a position derived from a rollup needs the data to reproduce the figure, and a committee's attestation is not a reproduction.

And the ability to recover from an operator's failure: a rollup whose data is available can be restarted by someone else from that data, while a rollup whose data was held by a defunct party cannot. For a regulated firm those three are the questions a supervisor would ask about any outsourced record-keeping, which is also the frame the DORA regulation in Germany page applies to dependencies of this kind.

A rollup transaction's cost has three parts and the proportions have shifted. Execution on the rollup is cheap, because the rollup's own compute is not scarce. Publishing the data is the part blobs made cheap and the part an external layer makes cheaper still. Settlement on the base chain, posting the state root and the proof, is amortized across all the transactions in a batch, so it shrinks per transaction as batches grow.

After EIP-4844 the data part stopped dominating, which is why rollup fees fell by two orders of magnitude and why the discussion moved to the base chain's blob capacity. For a financial application the useful consequence is that per-transaction cost is no longer the constraint it was, and the remaining questions are about finality time and about who holds the data, which are the ones this page is about.

Where the data goes. A rollup publishes its transaction data to the base chain, so availability inherits the base chain's security. A validium uses validity proofs for correctness but keeps the data off the base chain, typically with a data availability committee, which makes it cheaper and introduces a party that can withhold. The correctness of a validium's state is still proven; what is not guaranteed is that anyone can read the data they would need in order to exit. That is the trade, and it should be a conscious one for anyone holding assets there.

Only if it holds or records assets on a layer 2, and then it is one of the few technical properties that changes the legal position. Where the data is published decides whether the bank can reconstruct its own holdings without the operator's help, which is the difference between holding an asset and holding a claim. A bank that is deciding which network to issue or hold on should get the answer in writing: which layer carries the data, whether availability is sampled or attested, and who the parties are if it is attested.