Modular Blockchains Explained 2026: Celestia vs EigenDA vs Avail

Why Blockchains Got Split Apart

Every blockchain, however it is marketed, does four jobs at once.

• Execution. Running transactions and computing the new state. This is the part users feel as speed.
• Settlement. Being the court of final appeal - where disputes are resolved, proofs verified and withdrawals honoured.
• Consensus. Agreeing on the order of blocks so everyone sees the same history.
• Data availability. Publishing the raw bytes of each block so anyone can download them, rebuild the state and check the work.

A monolithic chain does all four itself. Bitcoin does. Solana does. Ethereum L1 did, for most of its life. The cost is that every full node has to be good at everything, and the weakest link caps the system.

A modular architecture unbundles them. A rollup executes on its own, settles on Ethereum where its proofs are verified, orders blocks with its own sequencer, and publishes its data wherever data is cheapest. The four jobs become a supply chain - and like any supply chain, the interesting question is which link is the bottleneck.

Data Availability Is the Bottleneck, Not Execution

Execution is not the expensive part of running a rollup. Processing a few thousand transactions on a beefy sequencer is cheap. The recurring bill is publishing the compressed transaction data somewhere everybody can see it - because every safety property a rollup claims depends on it:

• An optimistic rollup relies on someone noticing a bad state root and submitting a fraud proof. You cannot prove a computation was wrong if you cannot see the inputs.
• A ZK rollup proves the state transition was valid - but validity is not usability. If the operator publishes a valid proof and hides the data, the proof says the state changed correctly while nobody can tell what the new state is. Users cannot construct a withdrawal. The chain becomes provably correct and practically frozen. (We unpack both designs in our guide to ZK rollups vs optimistic rollups.)
• Forced exits - the escape hatch that lets you leave even if the sequencer censors you - require reconstructing your balance from published data.

DA is what makes a rollup a rollup rather than a database with a nice website. And it is priced by the byte, forever, which is why it dominates the cost structure. For how those bytes become a number on your wallet screen, see our explainer on gas fees.

What Data Availability Sampling Actually Does

The naive way to guarantee data was published is to make every node download every byte. That works, and it is exactly why blockspace is scarce: the block size ceiling is set by what a hobbyist node can handle. DAS breaks that link, in two steps.

Erasure coding. Block data is expanded with redundancy so the original can be reconstructed from any sufficiently large subset of the pieces. Per the Celestia documentation, block data is split into k by k shares, arranged in a matrix, and extended with parity data into a 2k by 2k extended matrix, producing separate Merkle roots for rows and columns. Because of that redundancy, a producer who wants to hide even one share must withhold a large fraction of the extended data - which is loud and easy to catch.

Random sampling. Light nodes each request a few random coordinates from that matrix, with Merkle proofs. A single sample proves almost nothing, but the probability that a withheld block survives many independent nodes each asking for random pieces falls off a cliff. Per L2BEAT, a Celestia light node performs a minimum of 16 samples per block - roughly 99 percent confidence that unavailable shares would be detected.

The pay-off: the more light nodes sample, the bigger blocks can safely get. Security scales with participation instead of being consumed by it. Celestia notes that for block data of n squared bytes, each light node downloads only on the order of n bytes - so improvements in light-node bandwidth have a quadratic effect on DA throughput. It is the same intellectual family as sharding, which is why DAS appears in every serious scaling roadmap, Ethereum's included.

Ethereum's Own DA: Blobs, EIP-4844 and Fusaka

Before March 2024, rollups posted data as ordinary calldata, competing with every swap and mint for the same gas. EIP-4844, shipped in Dencun, introduced a separate lane. Straight from the spec:

• A blob holds 4,096 field elements of 32 bytes each - 131,072 bytes, or 128 KB.
• Blobs are committed with KZG polynomial commitments, exposed to the EVM as a 32-byte versioned hash (version 0x01) rather than readable bytes.
• Blob gas has its own exponential fee market, independent of execution gas, with a minimum base fee per blob gas of 1.
• Blobs are pruned. MIN_EPOCHS_FOR_BLOB_SIDECARS_REQUESTS is 4,096 epochs, roughly 18 days. Ethereum guarantees the data was available; it does not host it forever.
• The original parameters were a target of 3 and a maximum of 6 blobs per block.

Those ceilings have moved twice. Per the Ethereum Foundation, pre-Fusaka limits were a target of 6 and a maximum of 9. Fusaka activated on mainnet on 3 December 2025, shipping PeerDAS (EIP-7594) - Ethereum finally doing data availability sampling at the protocol level. Two blob-parameter-only forks then raised capacity without a full hard fork: BPO1 on 9 December 2025 to a target of 10 and maximum of 15, and BPO2 on 7 January 2026 to a target of 14 and maximum of 21.

Fusaka also shipped EIP-7918, putting a floor under the blob base fee by bounding it against execution cost with a BLOB_BASE_COST constant of 2^13. Before it, blob fees routinely collapsed to 1 wei. Blobs are no longer free by default.

So why do external DA layers still exist? Two reasons, both arithmetic. Bandwidth: even at 21 blobs per 12-second slot, L2BEAT puts Ethereum's maximum DA throughput at roughly 0.22 MiB/s - the entire shared budget for 49-plus projects. Celestia's live ceiling is listed at about 5.33 MiB/s; EigenDA advertises 100 MB/s. And price under contention: the blob fee market is exponential, so when demand pushes past target the base fee climbs fast, and posting a megabyte to Ethereum becomes far more expensive than posting it to a chain built to sell nothing but bandwidth. Every major L2 that has left for alt-DA left on that trade.

How We Compared

This is an editorial synthesis, not a lab test. We read each project's official documentation, blogs and improvement proposals, the relevant Ethereum EIPs, and the L2BEAT data availability dashboard - the one neutral public place tracking value secured, validator counts, throughput ceilings and adoption side by side. Every L2BEAT figure below is a mid-July 2026 snapshot and will drift: value secured moves with token prices, throughput ceilings move with governance votes. We deliberately publish no dollar cost-per-megabyte figures, because DA prices are set by live fee markets and any number we printed would be stale by the time you read it.

1. Celestia — Best for cheap bandwidth with real light-client verification

Best for: rollups that want the largest cheap-bandwidth ecosystem plus genuine data availability sampling, and can accept a young independent validator set.

Celestia made modular a category. It launched mainnet beta on 31 October 2023 as a purpose-built DA layer: it does not execute smart contracts and does not settle anything - it orders and publishes blobs. Its security model is a standard proof-of-stake set staking TIA, but what distinguishes it is that light nodes actually sample. You do not have to trust validators to be honest about publication; you can check, from a laptop.

• Security model: own proof-of-stake set. Per L2BEAT, 98 validators and roughly 137.2M USD in economic security, with stake slashable in a data-withholding attack.
• Throughput: L2BEAT lists a live maximum of about 5.33 MiB/s. The Matcha (v6) upgrade, live on mainnet in late November 2025, lifts the ceiling much further: per the Celestia blog, CIP-38 raises the maximum block size from 8MB to 128MB, the data square from 128 to 512, and maximum transaction size from 2MB to 8MB. The live cap is a governance parameter, so the deployed number lags the protocol capability.
• Trust assumptions: DAS plus bad-encoding fraud proofs. L2BEAT flags the residual risks - a dishonest supermajority finalising an unavailable block with no light nodes sampling, or too few light nodes to collectively reconstruct.
• Adoption: 11-plus projects, about 170.9M USD total value secured - Manta Pacific, Eclipse, Derive, Orderly Network and others.
• Retention and economics: roughly 7 days of storage after CIP-34 cut the pruning window; Matcha also halved issuance from 5 to 2.5 percent and cut unbonding to 14 days.

Limitations: economic security roughly 350 times smaller than Ethereum's, denominated in a token correlated with the ecosystem it secures. Sampling is only a strong guarantee if enough light nodes are actually running.

2. EigenDA — Best for raw throughput backed by restaked ETH

Best for: high-volume chains - gaming, perps, anything with MegaETH-class transaction counts - that need bandwidth first and will accept a committee trust model to get it.

EigenDA is EigenLayer's flagship service and the clearest expression of the restaking thesis: rather than bootstrap a new validator set and a new token, borrow Ethereum's capital. Operators restake ETH (and EIGEN) and take on the obligation to store and serve blob chunks. It went live on mainnet in April 2024; EigenDA V2 launched at 100 MB/s in mid-2025, the highest advertised figure of any live DA layer.

Architecturally it is deliberately not a blockchain. A disperser erasure codes each blob, sends chunks directly to operators over unicast rather than gossiping them, collects signatures, and posts an attestation on Ethereum. That is what buys the throughput - and where the trade-off lives.

• Security model: EigenLayer restaking, with operators meeting a minimum stake in the ETH quorum (32 ETH) or the EIGEN quorum (1 EIGEN). Per L2BEAT, EigenDA has not activated slashing, so L2BEAT records its economic security as effectively zero. The stake is there; the punishment mechanism is not switched on.
• Throughput: 100 MB/s on V2 per the EigenCloud blog, with average end-to-end latency around 5 seconds.
• Trust assumptions: no DAS for end users - you trust a quorum of operators to have signed honestly. L2BEAT records no fraud detection mechanism and warns that users can be censored if the disperser does not distribute data to operators. The disperser is currently run by Eigen Labs.
• Adoption: 7-plus projects, about 580.5M USD total value secured across 91 operators - the largest alt-DA footprint by value - including Celo, Ronin, MegaETH, Fuel Ignition and Aevo.
• Retention and pricing: blobs reach end of life about 14 days after on-chain confirmation. Two payment modes - on-demand, and discounted pre-paid reserved bandwidth - Because published third-party rate figures for EigenDA go stale quickly, treat the live rate card as the source of truth — this guide deliberately publishes no dollar-per-megabyte number for it.

Limitations: the single disperser is a real centralisation and liveness dependency, there is no user-side sampling, and until slashing is live the economic security is nominal rather than enforced.

3. Avail — Best for DAS purism and a unification-first roadmap

Best for: teams that want Ethereum-style DAS cryptography on an independent validator set, and are betting on Avail's cross-chain coordination layer rather than DA alone.

Avail shipped its DA mainnet on 23 July 2024, and is technically the closest of the three to what Ethereum's own danksharding roadmap describes: KZG polynomial commitments plus validity proofs plus data availability sampling, on a Substrate-based validator set. Light clients verify blocks without downloading them.

Where Avail differs is scope. It sells not DA as a standalone product so much as a base layer for unification: Nexus, its cross-chain coordination protocol, and Fusion, a design for letting external assets like BTC and ETH contribute to Avail's economic security. Fusion is a stated direction, not shipped security.

• Security model: own proof-of-stake set. Per L2BEAT, 80 validators and roughly 8.9M USD in economic security, with slashable stake for withholding.
• Throughput: L2BEAT lists a maximum of about 0.2 MiB/s - the lowest of the alt-DA set, roughly in line with Ethereum blobs. Avail's DA page claims 250ms preconfirmations with 20-second DA finality; block size doubled to 4MB in early 2025, with a roadmap toward far larger blocks.
• Trust assumptions: DAS, with a caveat L2BEAT records plainly - the block reconstruction protocol that lets light nodes collectively rebuild a block is still under development. Sampling detects withholding; reconstruction is what lets the network recover from it.
• Adoption: 2 projects, about 20.1M USD total value secured - Sophon and Lens, with Lens responsible for most recent data posting.

Limitations: the smallest security budget and thinnest adoption of the three, and its throughput advantage over Ethereum blobs is currently marginal. Nexus is genuinely differentiated, but it is a bet on a roadmap.

4. Ethereum blobs — Best for anyone who cannot afford a new trust assumption

Best for: rollups holding serious value, where a DA committee failure would be existential and the premium is worth paying.

The baseline. Data availability is enshrined in consensus - no bridge, no attestation, no committee. If your data is in a blob, the validator set that secures ETH itself vouched for its publication.

• Security model: enshrined. Per L2BEAT, 880.5K validators and about 48.4B USD of economic security.
• Throughput: about 0.22 MiB/s maximum after BPO2 (21 blobs of 128 KB per 12-second slot), with 18 days retention.
• Adoption: 49-plus projects, roughly 35.4B USD total value secured - Arbitrum One, Base, OP Mainnet and the overwhelming majority of L2 value, as our Ethereum layer 2 guide covers in depth.

Limitations: the least bandwidth by a wide margin, a shared fee market that spikes exactly when you most want to post, and a hard price floor since EIP-7918.

DA Layers at a Glance

All L2BEAT figures are a mid-July 2026 snapshot and will move.

Which Should a Rollup Choose?

Ethereum blobs if your chain holds real money. A DeFi rollup with hundreds of millions in TVL should not save four figures a month by adding a committee that can, in principle, attest to data it never received. Post to blobs, stay a rollup in L2BEAT's classification, and let the 48B USD security budget work. The economics only break when your data volume genuinely outgrows blobspace.

EigenDA if throughput is the product. A game, an orderbook exchange or a MegaETH-class chain cannot fit in a 0.22 MiB/s shared lane. A 100 MB/s ceiling is not marginally better, it is a different category. You accept a quorum trust model and a disperser dependency - rational when the average transaction is worth cents, not thousands.

Celestia if you want cheap bandwidth without abandoning verifiability. The middle path: far more room than blobs, real DAS so users can check rather than trust, and the most mature alt-DA tooling. Its security budget is the honest limitation.

Avail if you are buying the unification story. On DA metrics alone it leads on none of throughput, security budget or adoption. It leads on architectural purity and on Nexus - a strategic bet, not a cost optimisation.

And the honest fourth answer: many teams should not choose yet. DA adapters are increasingly swappable. Build so you can switch.

The Trust Assumption Nobody Reads

When a rollup posts data to Ethereum, availability is a consensus fact. When it posts to an external DA layer, its Ethereum contracts cannot see that layer - so a DA bridge contract accepts a signed attestation as evidence that publication happened. If a dishonest supermajority of that layer's validators or operators attests to data that was never released, the Ethereum contract accepts a state root nobody can verify, and users cannot reconstruct their balances to exit.

That is why L2BEAT does not call these chains rollups. They are validiums (with validity proofs) or optimiums (with fraud proofs). It does not make them bad - it makes them a different security product, and the discount on your DA bill is the price of that difference. Comparing DA layers purely on dollars per megabyte is reading the invoice and ignoring the contract.

Conclusion

Modularity is not a philosophy, it is a decomposition: four jobs, four markets, four prices. Data availability turned out to matter most, because it is the layer every rollup pays for continuously and the layer every rollup's safety story quietly depends on.

Ethereum has responded far more aggressively than the alt-DA thesis assumed - blobs, then PeerDAS, then a target of 14 and a maximum of 21 blobs across two BPO forks. Celestia answered by lifting its ceiling toward 128MB blocks, EigenDA competes on a bandwidth number nobody else is close to, and Avail competes on a different axis entirely. Work out what your data actually costs you, work out what a DA committee failure would cost you, and pick the layer where the second number is smaller than the first.

This is an editorial synthesis of vendor documentation and public sources; see our [methodology](/methodology). Verify current details with each provider.