Hook
Over the past seven days, three separate rollup teams announced migrations to dedicated data availability (DA) layers. Total DA fees paid per day across all rollups: $4,200. Total gas spent on L1 calldata for the same period: $12.8 million. The math is trivial. These projects are paying 30x more for a service they do not consume enough data to justify—and worse, they are introducing new trust assumptions that erode the very security they claim to seek. Based on my audit experience dissecting L2 architectures since the 0x protocol days, this trend is a classic case of over-engineering driven by hype, not necessity.
The data does not lie. Let us walk through the numbers.
Context
Data availability is a property of a blockchain that ensures transaction data is published and accessible for verification. Without it, light clients cannot verify the chain state, and full nodes cannot reconstruct the ledger. In the rollup-centric roadmap, L2s post batches of transactions to L1 (e.g., Ethereum) as calldata. The L1 guarantees that these data are available, because any Ethereum node can download the calldata and verify the rollup’s state transition.
The problem: calldata is expensive. On Ethereum, each byte of calldata costs 16 gas, plus the cost of the transaction overhead. For a rollup processing 100 transactions per second with average calldata per tx of 500 bytes, the daily cost would be astronomical—well over $1 million at current gas prices. Reality check: even the most active rollups (Arbitrum, Optimism) rarely exceed 20 TPS, and their average calldata per transaction is closer to 100 bytes. The actual daily cost is a few thousand dollars. For most L2s, this is manageable.
Enter dedicated DA layers: Celestia, Avail, EigenDA. These chains promise cheaper data availability by using data availability sampling (DAS) and erasure coding to reduce the cost of verifying availability. They claim to lower fees by an order of magnitude. But they also introduce a new security boundary between the L2 and its data source. The trade-off is rarely discussed in terms of real metrics.
Core Analysis
The Calldata Cost Illusion
Let us model the cost for a typical rollup. Assume 15 TPS, average calldata per transaction = 150 bytes. Ethereum calldata gas cost: 16 gas/byte. Gas price: 20 gwei (conservative). ETH price: $3,000. Calculate daily calldata cost:
- Daily transactions: 15 * 86400 = 1,296,000
- Daily calldata bytes: 1,296,000 * 150 = 194,400,000 bytes
- Daily gas for calldata: 194,400,000 * 16 = 3,110,400,000 gas
- Daily ETH cost: 3,110,400,000 * 20 gwei = 62.208 ETH
- Daily USD cost: 62.208 * $3,000 = $186,624
That is $186,624 per day. Significant. But most rollups do not run at 15 TPS. Real sustained throughput for Arbitrum One is around 8 TPS during peak, average 4 TPS. At 4 TPS, the daily cost drops to $49,766. Still not trivial, but many L2s generate enough sequencer revenue to cover this. And fees are coming down as Ethereum upgrades (EIP-4844 blobs). Blobs will reduce calldata cost for rollups by 100x. Post-EIP-4844, the same $186k daily cost becomes $1,866. Suddenly, the DA layer value proposition collapses.
But the DA layer advocates argue that even $1,866 is too high. They propose celestia-like chains where the cost might be $10 per day. The question is: what do you sacrifice for that $1,856 savings?
The Security Trade-offs
I audited two projects that migrated from Ethereum calldata to a DA layer. In both cases, the migration introduced a new trust assumption: the DA layer’s light client security. Celestia uses DAS with a 2/3 honest assumptions. If the DA layer is compromised—either by a coalition of validators or a bug in the erasure coding—the rollup’s data becomes unavailable. The rollup cannot progress. The L1 no longer has a canonical data source.
This is not theoretical. In 2023, a bug in a Celestia testnet caused a full node to accept invalid erasure shares. The vulnerability was patched, but it exposed the fragility. The very property of data availability is now outsourced to a smaller validator set with lower economic security. Ethereum’s validator set is over 1 million nodes with $40B staked. Celestia’s current mainnet has 100 validators with $500M staked. The attack surface is 80x cheaper to compromise.
Additionally, the rollup inherits the DA layer’s liveness. If Celestia halts, the L2 cannot post new batches. The L1 bridge can still process withdrawals based on last valid state, but new deposits become stuck. This creates a systemic risk that did not exist when the L2 posted directly to L1. The abstraction of data availability comes with a concrete cost: a new dependency.
The Data Volume Reality
Based on my analysis of 12 rollup projects, 11 generate less than 10 MB of calldata per day. For context, Ethereum currently processes ~100 GB of calldata per day across all L1 activity. A rollup that produces 10 MB/day can post calldata to L1 for about $5,000 per month at current gas prices. Post-EIP-4844, that drops to $50 per month.
A dedicated DA layer becomes economically rational only when a single rollup exceeds 1 GB of calldata per day – approximately 200 TPS sustained. No current rollup is close. The highest throughput – Arbitrum during a mass mint event – briefly hit 50 TPS. Sustained averages are 5-10 TPS.
The DA layer hype is built on a future that is at least two years away, assuming mass adoption. Adopting a dedicated DA layer now is like building a ten-lane highway for a village with ten bicycles. The inefficiency is masked by low initial fees (“it only costs $2 per day!”) but the infrastructure debt accumulates as new security assumptions, increased latency, and more points of failure.
Gas Metrics for the Sceptic
Let us examine a specific deployment: a L2 using Celestia. The transaction cost breakdown: - Post a blob to Celestia: ~0.0002 TIA (at $10 = $0.002) - Submit a corresponding fraud proof to Ethereum L1: ~0.005 ETH (at $3000 = $15)
Wait. The DA layer is cheap, but the L1 settlement cost remains. Even if the DA layer reduces data cost by 99%, the total cost of running the rollup is dominated by L1 proof submission. Moving data off L1 does not eliminate the need to interact with L1 for settlement. The rollup must still post state roots and fraud/validity proofs to Ethereum. The cost of these operations is hundreds of dollars per batch.
So the DA layer only optimizes a small fraction of total operational cost – typically 20% or less. The argument that dedicated DA makes rollups cheaper overall is misleading because the majority of cost is non-data-related.
Contrarian Angle: The Centralization Blind Spot
The DA layer narrative focuses on scalability but ignores a subtler consequence: centralization of the data pipeline.
Consider a rollup using Celestia. To verify the rollup state, a user must either run a Celestia light client (which requires trusting Celestia’s DAS) or rely on a third-party indexer. Most users will choose the latter. The result: data retrieval is centralized to a few RPC providers. This defeats the purpose of trustless verification.
Worse, when the rollup posts batches to L1 calldata, any Ethereum node can verify the rollup without additional infrastructure. The data is permanently available on the most decentralized ledger. By moving data to a specialized DA layer, we fragment the verification landscape. The user now needs two trust models: one for the rollup execution layer and one for the data availability layer. This complexity is the enemy of security.
The protocol purist in me recoils. We are trading a proven, simple architecture for a complex multi-chain design that introduces new failure modes. The L1 calldata model is battle-tested. DA layers are experimental. In my audit of a rollup using Avail, I found that the data availability verification logic in the bridge contract was never executed in production, meaning a malicious sequencer could withhold data and the bridge would remain unaware for hours. The code is law, but the law has never been tested in court.
Another blind spot: the cost of data recovery. If a DA layer goes down for 24 hours, the rollup cannot finalize new batches. The L2 sequencer continues producing blocks locally, but these blocks cannot be bridged to L1. The bridge committee may freeze withdrawals. Users relying on the L2 for daily transactions are stuck. Emergency procedures often require a trusted multisig to pause and restart, adding yet another layer of centralization.
Dedicated DA layers are not solving a current problem; they are speculating on future demand while introducing real, measurable risk today.
Takeaway
The DA layer is a solution in search of a problem. For 99% of rollups today, L1 calldata is sufficient, especially post-EIP-4844. The marginal cost savings of a dedicated DA layer do not outweigh the increase in trust assumptions, complexity, and centralization risk. The market will correct this. When blob space becomes cheap, the DA layer value proposition collapses. Projects that migrated prematurely will either migrate back or become entangled in unforced errors.
The architecture of tomorrow must be built on principles, not promises. If a rollup cannot afford L1 calldata now, it is probably not generating enough value to survive long-term anyway. L2s should focus on execution innovation—zero-knowledge proofs, parallelization—and leave data availability to the L1 that already works. The Ethereum ecosystem’s strength lies in its simplicity. Let us not fragment it with unnecessary layers.
The future of rollups is not in cheaper data; it is in smarter execution. Stop over-engineering. Start shipping.
