Over the past quarter, three rollups collectively spent 12,000 ETH on data availability (DA) posting to Ethereum mainnet. Meanwhile, their combined transaction throughput never exceeded 150 TPS. The math is brutal. At current gas prices, each transaction cost roughly $0.80 in DA overhead alone. For comparison, a simple peer-to-peer transfer on L1 costs $0.30. The premium for “scaling” is negative. This is not scaling. This is an architecture tax imposed by a narrative that conflates data availability with security.
Let me be precise. The current rollup landscape operates under a core assumption: that posting all transaction data to a dedicated high-availability layer is necessary for security. This assumption is codified in the canonical rollup design—Ethereum as DA, or Celestia, EigenDA, Avail. The logic is sound in theory: if the sequencer goes rogue, any observer can reconstruct the state from the data. But theory and practice diverge when you examine actual usage patterns.
I have spent the last year auditing rollup contracts for a tier-1 L2. What I found is consistent across six different implementations: the average rollup batch contains fewer than 50 transactions. Of those, 80% are trivial transfers or mints. The cryptographic commitment posted to L1 is a single Merkle root—256 bits. The data being “availed” is mostly noise. The DA layer is not a security guarantee; it is a performance tax on low-activity chains.
This is the unintended consequence of designing for the worst-case threat model while ignoring the average-case economics. The worst case is a sequencer withholding data. The solution is expensive storage. The average case is a sequencer posting data that no one ever needs to read. The cost is borne by every user, every transaction, every day.
Consider the numbers. A typical rollup batch of 50 transactions, each consuming 500 bytes, generates 25 KB of data. Posting 25 KB to Ethereum calldata costs roughly 0.003 ETH at current base fees—about $0.60 per batch. For 50 transactions, that is $0.012 per transaction. Double that with blobs. Now add compression. Most rollups achieve 4x compression, dropping per-transaction DA cost to $0.003. Still, that is 10% of the total transaction fee for a typical $0.03 user payment. On a chain processing 10 million transactions per month (optimistic estimate), DA overhead alone consumes $30,000 per month. For what? The probability that a user will ever need to force-withdraw that data is astronomically low.
The narrative that “DA is essential” was forged during the 2021 congestion era, when L1 gas prices spiked to 500 gwei. It made economic sense then to offload calldata to cheaper layers. But we are not in 2021. Today, Ethereum base fees hover around 5-20 gwei. The cost differential between L1 calldata and a dedicated DA layer (EigenDA, Celestia) is shrinking. Meanwhile, the complexity tax—extra trust assumptions, additional bridge risk, latency overhead—remains constant. The DA layer is a solution to a problem that no longer exists at the current scale.
Now, the contrarian angle: most rollups do not generate enough data to need dedicated DA. The threshold where dedicated DA becomes economically beneficial is around 500 TPS sustained. Below that, L1 calldata is cheaper and simpler. Above that, you have a scaling problem. Today, no rollup sustains 500 TPS for more than a few minutes. Ethereum itself averages 15 TPS. The entire ecosystem of rollups combined might hit 200 TPS on a good day. Dedicated DA is a premature optimization.
But there is a deeper blind spot: the security model of dedicated DA layers is often weaker than advertised. I audited the smart contract for one such DA validator set in 2023. The slashing conditions were intentionally vague to allow for “social recovery.” In practice, the DA committee could collude to censor data without losing capital. The whitepaper promised economic finality; the code delivered an honor system. This is the unintended consequence of trusting a new consensus group: you inherit their governance risks.
Let me illustrate with a concrete example. Project X, a prominent rollup, migrated from L1 calldata to EigenDA in Q3 2023. They reported a 40% reduction in fees. What was not reported was the 2-hour latency increase for data confirmations. Their sequencer now waits for EigenDA attestations before finalizing a batch. That extra 2 hours exposes users to frontrunning via cross-domain arbitration. The L1-based rollup had zero latency for DA—data was available as soon as the transaction landed. The dedicated layer introduced a new timing risk. Performance optimization without considering timing constraints is a classic engineering fallacy.
Based on my experience dissecting 0x protocol’s order matching in 2017, I recognized a similar pattern: a system designed for peak throughput under ideal conditions, but fragile under adversarial latency. The solution then was to batch orders with a deadline. The solution now is to batch transactions with a timeout. But rollups cannot simply add deadlines—they need deterministic finality. The tension remains unresolved.
So what is the alternative? It is not to abandon scalability. It is to accept that most applications do not need full data availability. Why post every transaction to a globally replicated ledger when 99% of users trust the sequencer? The market is already voting with its feet: perp DEXs with order books, like dYdX, moved to their own app-chains with minimal DA. NFT marketplaces batch trades off-chain. Even Uniswap’s L2 deployments use optimistic rollups but rely on forced inclusion mechanisms for security, not constant DA.
There is a simpler model: validium. Execute transactions off-chain, post only state roots and fraud proofs to L1. No data, just cryptographic commitments. The trade-off is that users must trust the data committee—or exit via a forced withdrawal based on the state root. This is the architecture of Immutable X, Sorare, and many gaming chains. It works. It is cheap. It is secure enough for non-financial use cases. And it proves that the DA layer is not a universal requirement.
The current hype around modular DA—Celestia, Avail, EigenDA—reminds me of the 2018 “scaling trilemma” obsession. We spent years debating whether sharding or DAGs were the answer. The answer turned out to be: it depends on the application. The same is true for DA. Dedicated DA is a tool, not a solution. Most rollups today would be better off staying on L1 calldata or moving to validium. The data shows it.
What will happen next? I predict a consolidation wave. As the market realizes that DA layers are overcapitalized relative to actual demand, we will see mergers and pivots. EigenDA will likely reposition as a general-purpose sequencing layer, not a DA layer. Celestia will focus on sovereign rollups where DA and execution are tightly coupled. The survivors will be those that offer measurable performance gains—not abstract security guarantees.
The key insight: data availability is a spectrum, not a binary. Just as decentralization is a spectrum, so is DA. A rollup processing 100 TPS does not need the same DA as one processing 10,000 TPS. The industry must stop treating DA as a checkbox and start treating it as a scaling dimension to optimize per use case. Otherwise, we are just paying for security theater.
I will leave you with a forward-looking question: In a world where AI agents execute billions of microtransactions daily, will they want to post every step to a global DA layer? Or will they settle for local verification with periodic audits? The answer determines the architecture of the next decade. My bet is on minimal DA with trust-minimized exit paths. The code is waiting for someone to write it.