Over the past quarter, Ethereum’s blob count has averaged 1.2 blobs per slot. Each blob carries 128 KB of compressed transaction data. At current usage, that’s roughly 150 KB per second. Arbitrum One, the largest optimistic rollup by TVL, contributes about 40% of that traffic. The remaining 60% is split across Base, Optimism, and a dozen smaller L2s. The total data throughput required by all active rollups combined is less than 1 MB/s. Compare that to the theoretical capacity of a single Celestia light node, which can sample 2 MB/s. The math is trivial. We are building data availability layers that could handle 100 times the current demand. And yet, every new rollup launch touts its own DA solution as a competitive advantage. This is not engineering. This is cargo-culting.
The modular thesis gained traction in 2022 when Celestia published its whitepaper on data availability sampling. The argument was logical: monolithic chains like Ethereum suffered from data bloat, and separating execution from consensus and data would enable scalability. By 2024, every major L2 had either integrated or planned to integrate a dedicated DA layer. EigenDA, Avail, Celestia, and even custom solutions from Near and Polygon became the default stack. The market assumed that data availability was the bottleneck. It was not. It never was. The bottleneck was execution capacity, which rollups solved by batching transactions off-chain. DA, by contrast, was a solved problem at Ethereum’s current throughput. The push for modular DA was a solution in search of a problem.
Let’s examine the numbers. A rollup batch submission to Ethereum costs roughly 0.001 ETH per 100 KB of calldata—around $2.50 at current prices. That same batch submitted to a dedicated DA layer costs about $0.10—a 96% reduction. At first glance, this seems like a clear win. But the batch size for most rollups is significantly smaller. The median rollup submits batches every 10 minutes, with an average payload of 40 KB. That’s 6 batches per hour, 144 per day. The annual saving from switching to a dedicated DA layer would be approximately $12,500 per year. For a protocol managing $500 million in TVL, that is 0.0025% of total value. The cost savings are noise. The real expense is security. By moving data off Ethereum, rollups lose the security guarantees of L1 finality and inherit the weaker trust assumptions of a new, often unproven, consensus set. The trade-off is asymmetric: a marginal cost reduction for a significant increase in trust dependency.
Now consider the security architecture of a typical DA layer. EigenDA relies on a committee of validators who sign attestations that data is available. To avoid Byzantine behavior, the system assumes at least two-thirds of the committee are honest. But who are these validators? In EigenLayer’s current setup, the top 10 operators control over 60% of the restaked ETH. This is a classic principal-agent problem disguised as modular innovation. The committee members have financial incentives to act honestly, but the concentration of power creates a systemic risk. A colluding subset could withhold data or produce false availability proofs. The probability is low, but the impact is catastrophic. Ethereum’s L1, by contrast, has thousands of independent validators with a proven track record of resistance to collusion. The modular solution does not eliminate the trust problem; it simply relocates it to a smaller, more vulnerable set of actors.
Based on my audit experience with 0x protocol’s v2 exchange contracts in 2017, I learned that race conditions often arise from assumptions about external dependencies. The same principle applies here. Rollups that offload DA assume the external layer will always be available and honest. But availability is not a binary property—it is a spectrum with time-dependent guarantees. If a DA layer goes down for five minutes, the rollup must halt or risk invalid state transitions. The fallback mechanism is usually to post data back to Ethereum, but by then the damage is done. During the EigenDA mainnet launch in April 2024, a configuration error caused a 12-minute outage. No funds were lost, but several rollups had to freeze withdrawals. The modular stack introduced a single point of failure that the monolithic design never had. The irony is thick.
Let me be precise about the data demand. According to L2Beat, the total data emitted by all rollups in the past 30 days was 180 GB. Ethereum’s blobs have a theoretical capacity of 1.2 TB per month at current blob count limits. That’s a utilization rate of 15%. Even if all future rollups migrate to blobs, Ethereum can handle at least 5x current demand without any upgrade. The only scenario where dedicated DA makes sense is if rollups generate orders of magnitude more data than today—for example, if every DEX trade produces 50 KB of metadata. But that is not how blockchain applications work. Most transactions are simple token transfers or swaps, each under 500 bytes. The data compression schemes used by zk-rollups already reduce calldata by 90%. The data explosion narrative is a myth propagated by projects selling DA infrastructure.
Now consider the unintended consequences of this architectural choice. By moving data off L1, rollups lose the ability to enforce canonical ordering through Ethereum’s consensus. This creates an opening for MEV extraction at the batch submission level. Proposers on the DA layer can reorder batches before finalizing them, extracting value from pending transactions. This is not theoretical—it has been documented on Celestia where sequencers running on the same infrastructure as validators have frontrun batches. The modular design introduces a new attack surface for MEV that was previously contained within the rollup’s sequencer. The system becomes more complex without a proportional increase in security or throughput.
During the DeFi Summer of 2020, I spent weeks disassembling Uniswap V2’s AMM formula. I noticed that the constant product formula was elegant but fragile under certain market conditions. The same fragility applies to modular stacks. The interdependence of layers creates emergent failure modes that no single team can fully anticipate. For example, if a DA layer’s token price drops by 50%, the cost of posting data might increase dramatically if the fee model is token-denominated. This could make rollup operation unprofitable, triggering a migration back to L1. The transition period is undefined and risky. The modular design is not extensible in the way its proponents claim. It is a collection of tightly coupled components held together by economic incentives that can break under stress.
The contrarian angle is this: the push for dedicated DA is a supply-side phenomenon, not a demand-side necessity. The projects building DA layers need market share, so they market fear about data bloat. The rollups integrating them need differentiation, so they claim modular superiority. The end users—retail traders and application developers—see no difference in transaction fees or speed. In fact, some rollups that moved to dedicated DA actually saw higher fees due to the added overhead of verifying DA attestations. The modular narrative is a feature for VCs, not for users. It allows projects to raise larger rounds by telling a story of infinite scalability. But the technical reality is that Ethereum’s blob space is sufficient for the foreseeable future, and the security guarantees of L1 are irreplaceable.
Take Arbitrum One as a case study. It uses Ethereum’s calldata for DA. Its average transaction cost is $0.02. A comparable rollup using Celestia, like Eclipse, has similar costs. The user experience is identical. The difference is in the trust assumption: Arbitrum inherits Ethereum’s full security, while Eclipse inherits Celestia’s, which is weaker. The market has not priced this difference because retail investors cannot evaluate the cryptographic guarantees. They see lower fees on the surface but ignore the hidden security tax. Standards are just opinions with better PR. The industry will eventually consolidate around the most secure option, not the most modular one.
This leads to a forward-looking judgment. In the next bear market, when TVL drops and fee revenue collapses, the marginal cost savings from dedicated DA will evaporate. Rollups will be forced to justify their architectural choices to token holders. The ones that moved to weaker DA layers will face existential questions about their security model. I predict a wave of “back-to-L1” migrations as protocols seek to simplify their stack and reduce trust dependencies. The modular experiment will be remembered as a detour—a theoretical solution to a problem that never materialized.
The blockchain industry has a pattern of over-engineering. We build complex systems to solve scaling problems that do not exist, then spend years unwinding them. The DA layer is the latest example. The next time a rollup team tells you they need a dedicated DA solution, ask them to show their data usage metrics. If they cannot produce numbers, assume the decision is driven by fundraising, not engineering. The best architecture is the one that minimizes trust assumptions while meeting actual demand. For 99% of rollups, that is still Ethereum L1 calldata or blobs. Everything else is noise.