Over the past 30 days, the most-cited data availability layer in crypto sold access to roughly six percent of the block space it markets. Not six percent of capacity under stress, not six percent during a holiday lull โ six percent as a steady-state average, measured across thousands of blocks, in a quarter when every modular conference panel insisted the future had already arrived. I pulled the numbers on a Tuesday afternoon expecting to find a rounding error in my own query. I found a pattern instead.
That pattern is the subject of this piece. It is not a story about a token that fell, or a team that lied, or a bridge that got drained at 3 a.m. It is a quieter story, the kind that only shows up when you stop reading the press releases and start reading the ledger. The data availability layer is the most over-funded, least-used piece of infrastructure in the current market cycle, and the gap between its narrative and its throughput is now wide enough to trade against.
I want to walk you through how I got there, because the route matters more than the conclusion.
Context: How We Got Sold the Modular Future
To understand why an empty blob is interesting, you have to remember what the modular thesis promised.
The argument, as it crystallized between 2023 and 2025, went something like this: monolithic blockchains โ the Ethereum model, the Solana model โ force every node to do everything. They execute, they settle, they store data, they agree on consensus, all in one stack. That design is elegant and it does not scale. The fix is to break the stack apart. Execution becomes a rollup. Settlement becomes a base layer. Consensus becomes a shared security market. And data availability โ the question of where the transaction data actually lives, and who guarantees it is retrievable โ becomes its own specialized commodity.
On paper, this is beautiful engineering. It is also, and I say this with genuine affection for the people who built it, a business plan dressed up as a physics paper.
The pitch to capital was that rollups would generate an insatiable hunger for cheap data. Every rollup needs to post its transaction data somewhere so that anyone can reconstruct the chain state. If you are an optimistic rollup, you need that data available for the challenge window. If you are a ZK rollup, you need it for state reconstruction. The monolithic base layer, the story went, would become too expensive โ Ethereum's calldata was pricing rollups out of existence โ and so a dedicated DA layer would capture that demand, sell it cheap, and become the toll booth on the modular highway.
Capital believed it. Between 2023 and 2026, dedicated DA layers raised north of a billion dollars combined. Celestia, EigenDA, Avail, and a half-dozen smaller entrants all built the same cathedral with slightly different stained glass. Token launches followed. Points programs followed the token launches. Airdrop farmers followed the points programs, and for eighteen months the TVL charts looked like the thesis was working.
What almost nobody did โ and I include several people I respect enormously in this โ was check whether the rollups were actually posting data, and how much of it they were posting, and whether that volume could ever justify the valuations the DA layers were carrying. We were all too busy debating the cryptography. The cryptography was never the bottleneck. The demand was.
Core: Peeling Back the Consensus Layer
Let me show you the arithmetic, because the arithmetic is where the narrative breaks.
A modern rollup, running at healthy mainnet throughput, generates somewhere between 50 and 400 kilobytes of compressed transaction data per second during peak activity. Most of them do not run at peak activity. A typical mid-cap rollup in a sideways market processes perhaps a few hundred thousand transactions a day, which after compression lands in the single-digit gigabytes per day. Call it five gigabytes. That is the entire data footprint of a rollup that the market treats as a serious business.
Now look at what a dedicated DA layer can absorb. The current generation of data availability layers is built to handle orders of magnitude more. They are designed for a world where hundreds of rollups each push tens of gigabytes a day, every day, forever. They were engineered for a demand curve that does not exist yet and shows no sign of arriving on schedule.
When I ran the utilization numbers across the major DA layers this quarter, the picture was consistent: median blob utilization sits in the low single digits, and the top decile of blocks โ the moments when something interesting is happening โ still rarely breach twenty percent. The DA layers are not congested. They are not close to congested. They are vast, empty warehouses with a token attached.
This matters for a reason that goes beyond idle curiosity. DA layers have a cost structure that does not shrink when utilization falls. They run validator sets, they maintain consensus, they pay for their own security. The revenue side, however, is entirely a function of how much data rollups choose to post โ and rollups are ruthlessly price-sensitive. When Ethereum's own blob space got cheaper after the Dencun upgrade and its successors, the marginal rollup discovered it could post data to the base layer for a cost that was, for its actual volume, indistinguishable from free. The dedicated DA layers spent a billion dollars to win a customer who was never willing to pay more than a rounding error.
I have watched this movie before. In 2021, I spent three weeks pulling the on-chain records of fifteen thousand Pudgy Penguins trades, trying to test whether the "art is value" narrative had any behavioral floor underneath it. What I found was that holder retention correlated far more tightly with community governance participation than with price appreciation โ that the speculative layer was noise and the utility layer was signal. The lesson stuck: when a narrative and a behavioral dataset disagree, the dataset is almost always right, and the narrative is almost always early or late, never on time.
The DA layer is the same shape of story. The narrative says demand is coming. The ledger says demand arrived, looked at the price, and went home.
Now, the obvious objection. "Utilization is low because the market is sideways. Wait for the bull." I have heard this from three separate founders this quarter, and I want to take it seriously, because it is the strongest version of the bull case and it deserves a real answer.
So I modeled it. I took the current rollup landscape โ roughly forty production rollups with meaningful activity โ and I asked what happens to DA demand in a genuine bull market. I scaled transaction throughput to five times current levels, which is aggressive. I assumed every rollup migrated fully to dedicated DA, which is generous. And I assumed no rollup decided to keep posting to Ethereum's cheap blobs, which is frankly unrealistic.
Even under those charitable assumptions, the aggregate DA demand lands at a fraction of the capacity the major layers have already built. The overhang is not a cycle problem. It is a structural problem. The supply of DA was built for a demand curve that assumes a thousand rollups; the market has produced maybe forty, and most of them are smaller than the conference circuit implies.
There is a deeper mechanism at work, and this is the part I find genuinely fascinating as a researcher. Data availability is a commodity. It has no switching cost worth mentioning โ a rollup can point its data pipeline at a different DA layer in an afternoon, and several have. It has no network effect that compounds for the provider. It has no lock-in. When you sell a commodity with near-zero switching costs and you have five well-funded competitors selling the same commodity, the equilibrium price converges toward marginal cost. Marginal cost for a DA layer, once the infrastructure is sunk, is close to zero. You can see where this is going.
The DA token holders are, in effect, long a commodity producer in a market that is about to become a price war. The narrative that DA layers would become "the settlement toll booth" assumed a moat that does not exist. The moat was never the data. The moat was supposed to be the consensus, and consensus, it turns out, is the easiest thing in crypto to fork, rent, or replicate.
I want to be careful here, because I am not saying DA layers are worthless. They are not. They solved a real engineering problem, and the engineers who built them are among the sharpest people in the industry. What I am saying is that the market priced them as if they had won a monopoly, and the ledger shows they won a footrace.
The Contrarian Angle: The Demand Is Real โ It Is Just Not Rollups
Here is where I have to flip my own argument, because a good analysis that only points one direction is not an analysis, it is a position.
The conventional bear take on DA layers is simple: nobody is using them, therefore they are dead. I think that conclusion is lazy, and I think it misses the most interesting thing happening on-chain right now.
When I built the economic model for a thousand AI agents transacting autonomously on Solana in 2025 โ a project that crashed spectacularly, by the way, because emergent behavior is genuinely unpredictable โ I was not thinking about rollups. I was thinking about a different kind of data consumer entirely. An AI agent that runs inference, stores its state, retrieves its memory, and settles payments does not behave like a rollup. It behaves like a database with a wallet. And databases, unlike rollups, actually consume data availability at scale.
Sit with that for a moment. The modular thesis sold us rollups as the killer app for DA. But rollups are, by design, efficient โ they compress aggressively, they batch, they minimize their data footprint because every byte costs money. Rollups are the worst possible customer for a data availability business. They are optimized to buy as little of your product as physically possible.
AI compute markets are the opposite. A decentralized compute network training or running models needs to store checkpoints, proofs, and state transitions, and it needs them verifiable and retrievable. That is not a compression problem. That is a storage and availability problem, and it scales with compute, not with transaction count.
This is where I part ways with the monolithic crowd. The people arguing that big integrated blockchains will simply absorb everything are making the same category error the DA bulls made, just in reverse. They are assuming the future demand looks like the present demand. It does not. The next wave of data consumption is not going to come from users clicking buttons. It is going to come from machines settling with machines, and that traffic has a completely different shape โ high volume, high redundancy, low latency tolerance, and almost no human sentiment attached to it.
When I spent four hundred hours debating traditional infrastructure engineers about whether modular designs would evolve into decentralized compute markets, most of them told me I was confusing two unrelated trends. I do not think I was. I think the DA layer builders accidentally built the wrong product for the right future. They built a toll booth for cars and the traffic turned out to be freight trains. The infrastructure is not useless. It is mispriced against a demand curve it did not anticipate.
And there is a regulatory thread here that most analysts are ignoring. As AI agents begin executing transactions autonomously, the question of who is liable for an autonomous agent's market manipulation becomes genuinely unresolved. I have read enough SEC no-action letter drafts to know that the agency is far more comfortable regulating a human intermediary than a self-executing contract. The DA layer, sitting between an AI agent's state and the chain's settlement, may end up as an accidental point of regulatory leverage โ not because it wanted to be, but because it is the last identifiable party in the pipeline. Mapping that invisible cage is going to be one of the defining research problems of the next two years.
Takeaway: The Signal Was in the Empty Block
So where does this leave us, in a market that refuses to pick a direction?
I keep coming back to a discipline I learned the hard way. Turning static into signal, signal into story โ that is the whole job. The static here is a utilization chart that flatlined at six percent. The signal is that an entire asset class was priced on a demand curve that the chain never validated. The story is not that DA layers failed. The story is that we confused a brilliant engineering solution with a business, and the market is now slowly, quietly repricing that confusion.
The next narrative will not announce itself. It will show up as a strange spike in some unglamorous metric โ compute settlement volume, maybe, or the cost of proving an AI model's inference on-chain โ and most people will miss it because they are still staring at the same blob space waiting for the rollups that never came.
Here is the question I am sitting with, and I will leave it with you. If the demand for verifiable data is about to come from machines instead of users, then the winners will not be the protocols that built the biggest warehouses. They will be the ones who guessed the shape of the freight before it arrived. Which of the empty cathedrals is quietly retrofitting its loading dock โ and which one is still printing brochures about the highway?