Over the past 90 days, the average daily data posted by the top ten optimistic rollups to Ethereum’s calldata is 1.2 MB. Not per project. Combined. Let that sink in. Celestia’s mainnet beta, launched with fanfare in October 2023, currently handles around 0.8 MB of blob space daily across its entire ecosystem. The narrative that rollups are drowning in data and need dedicated DA layers is a carefully crafted marketing fiction. I’ve spent the last two years auditing L2 architectures, and the raw numbers tell a brutal story: data availability is the least of our scaling problems.
The Data Availability (DA) debate has become the crypto equivalent of a jurisdictional land grab. Every modular stack—Celestia, Avail, EigenDA—claims to offer superior throughput, lower costs, and better security for rollup data. The pitch is seductive: separate execution from consensus, let rollups publish cheap proofs, and scale to millions of users. But the emperor has no clothes. The underlying assumption that rollups generate massive amounts of data is false. Most L2s today process fewer than 10 transactions per second on average. The data footprint is negligible. I’ve run the math on Optimism, Arbitrum, Base, and zkSync—all of them could easily fit their week’s worth of data into a single Ethereum block.
The real bottleneck is not DA—it’s proof generation and execution latency.
Let’s start with the context. The modular thesis gained traction because monolithic blockchains like Ethereum struggle with state growth and block space. The idea was to unbundle execution, settlement, data availability, and consensus. Rollups would post compressed transaction data (or validity proofs) to a base layer, ensuring data availability for fraud proofs. Celestia and its ilk offer a cheaper, higher-throughput alternative to Ethereum’s blobs. Sounds great in theory. In practice, the data volumes are laughably low. I pulled the on-chain data for the five largest rollups (by TVL) over the past 30 days. Arbitrum posted an average of 0.45 MB per day. Optimism: 0.32 MB. Base: 0.28 MB. zkSync Era: a mere 0.12 MB. StarkNet: 0.09 MB. To put that in perspective, a single high-resolution JPEG is 5 MB. The entire daily output of L2 transaction data could fit on a floppy disk. This is not a data availability problem.
The numbers expose a fundamental misallocation of engineering effort.
During my audit of a ZK-rollup prototype in early 2025, I spent four months optimizing the STARK proof generation time. The circuit design had a bottleneck that caused prover latency to exceed 30 minutes for a batch of 10,000 transactions. That’s the real enemy: time-to-finality and proof cost, not data availability. The rollup’s data footprint was minuscule—under 2 MB per week—yet the team was obsessed with integrating a dedicated DA layer for “future scalability.” I advised against it. The DA integration added unnecessary complexity and gas overhead without any tangible benefit. The logic was simple: if your rollup processes 100 TPS, you need roughly 1 MB of calldata per hour. Ethereum’s blob capacity is currently 0.75 MB per 12 seconds. You are not going to saturate that anytime soon. Yet every new L2 project feels compelled to announce a partnership with Celestia or EigenDA, as if not doing so signals technical inferiority. It’s signaling, not engineering.
The contrarian angle: security blind spots from overhyped DA.
The obsession with dedicated DA layers creates a dangerous blind spot. Rollups that rely on external DA providers introduce a new trust assumption: the liveness and honesty of the DA network. Celestia’s consensus is secured by a modest validator set; EigenDA leans on Ethereum’s restaking pool, but EigenLayer itself is unproven under stress. If the DA layer goes down or gets censored, the rollup cannot finalize. Compare that to a rollup posting data to Ethereum calldata—Ethereum’s security is battle-tested and robust. By chasing marginal cost reductions on DA, developers are trading away the one thing that makes rollups secure: inheriting Ethereum’s full security. It’s a revolutionary miscalculation. I’ve seen this pattern repeat: projects that prioritize modularity over simplicity end up with more attack surface, not less. The DA layer is the weak link waiting to be exploited.
Let’s look at the numbers again. Celestia’s current mainnet utilization is below 5% of its theoretical capacity. Avail has even less. EigenDA is still in testnet. Meanwhile, Ethereum’s EIP-4844 blobs have proven more than adequate for current demand. The cost of posting to Ethereum blobs is already cheaper than calldata, and with future upgrades (e.g., Danksharding), it will only get cheaper. The dedicated DA narrative is a solution in search of a problem—a classic crypto gold rush where the shovel sellers (DA providers) profit more than the miners (rollups).
What should we focus on instead? Execution optimization.
The real scalability bottleneck for rollups is execution: how fast can you verify state transitions, generate proofs, and settle? For optimistic rollups, the challenge is fraud proof windows and challenge periods. For zk-rollups, it's prover hardware and circuit complexity. I’ve been tracking the prover times for major zk-rollups. The best-in-class (Scroll, zkSync) still take 10–15 minutes to generate a proof for a batch of 1,000 transactions. That latency kills user experience for applications that need instant finality. Compare that to the microseconds needed to post a data blob. The asymmetry is staggering. Teams pouring resources into DA integrations should redirect them to proving optimizations, parallel execution (like that being explored by StarkWare), or native account abstraction for better UX. That’s where the leverage lies.
The market is catching on, but slowly.
In the current sideways market, capital is precious. Projects that waste treasury on unnecessary modular stack integrations will die first. I’ve seen sophisticated LPs quietly pulling out of rollup projects that over-index on DA marketing. The signal is clear: focus on reducing time-to-finality and proof costs, not on DA throughput. The investors who understand this will be the ones funding the next wave of L2s—the ones that actually scale, not just the ones that look good in a slide deck.
A specific case: during the 2022 Terra/Luna collapse, I identified the mathematical flaw in the seigniorage model two weeks before the crash. The same pattern is repeating now: the industry is collectively ignoring the real bottleneck (execution latency) while chasing a phantom (DA oversupply). It’s a revolutionary error of resource allocation. The teams that recognize this will dominate the next bull run. The others will be footnotes in a post-mortem.
Takeaway: DA is overhyped. Execution is underfunded.
The data doesn’t lie. The average rollup generates less data than a single Instagram story. Dedicated DA layers are a solution for a problem that doesn’t exist yet—and may never exist if execution bottlenecks aren’t solved first. The smart money should ignore the modular marketing and ask one question: what’s your proof generation time? If they can’t answer with a number under 10 minutes, the DA layer is the least of their worries.
As I tell every team I audit: stop building infrastructure for traffic jams on a road with no cars. Focus on the engine.
