I remember the exact moment I stopped believing the data availability story. It was not a panel debate or a thread. It was a number on a dashboard at 2:14 in the morning, Denver time, in a hotel room I had paid for with a speaking fee I felt mildly guilty about accepting.
The number was the blob base fee on Ethereum mainnet. It read 1 wei. The floor. It had read 1 wei for most of that week, most of the preceding month, and โ with brief, violent exceptions during airdrop season โ for most of the two years since Dencun turned data availability into a commodity with a price so close to zero it might as well be a rounding error.
I had a second tab open. A funding announcement. A rollup I will not name, because the point is not the rollup. One hundred million dollars, Series B. The release used the word "modular" four times and "data availability" seven.
Seven times. For a chain that, on its best day that quarter, was posting less than a tenth of the blob capacity Ethereum had already handed it for free.
That is how bull markets work. They do not lie to you outright. They hand you a number and a narrative and bet you will never place them side by side. The chain does not lie. It simply waits to be asked the right question.
So I asked it. And what came back was not a scandal. It was worse. It was arithmetic.
What We Actually Built
To understand why the DA market looks the way it does, you have to go back to the design constraint that produced it, because that constraint was never about storage. It was about verification.
Before March 2024, rollups on Ethereum paid for calldata. Every batch of compressed transactions was written into the same gas market as swaps and liquidations, competing byte-for-byte with DeFi users. A rollup posting a large batch during a congested block was, functionally, a whale buying block space. The cost was brutal and โ this is the part people forget โ variable in a way that made rollup economics impossible to model. You could not price a transaction on an L2 when the L2's own input cost swung by 40x based on whether someone was minting an NFT that afternoon.
EIP-4844, shipped in Dencun on March 13, 2024, separated those markets. It introduced blobs: 128 KiB chunks of data attached to a beacon block, priced by their own independent fee market, and โ critically โ deleted after roughly eighteen days. Not stored forever. Not indexed. Made available, verified, and forgotten.
That deletion is the whole idea. A rollup does not need Ethereum to remember its transaction data forever. It needs Ethereum to guarantee, for a bounded window, that the data existed and was published, so that anyone who cares can reconstruct the state. After that window, the rollup's own state commitment is the truth, and the data is archaeology.
The design was elegant. The pricing was not.
The blob fee market is EIP-1559-style with an exponential adjustment keyed to excess blobs โ how far above target the recent usage has run. Target was set at three blobs per block, maximum six. In 2025's Pectra upgrade, that moved to a target of six and a maximum of nine. The intent was a gentle ramp: start small, let demand find the ceiling, expand as usage grows.
What nobody modeled carefully enough was the shape of the demand curve. Or rather, they modeled it, and then the marketing department modeled it again, and the second model won.
The Arithmetic of an Empty Cathedral
Let me do the math I keep doing, the one that never makes it into a deck.
One blob is 4,096 field elements of 32 bytes: 131,072 bytes, 128 KiB. At the pre-Pectra target of three blobs per 12-second slot, Ethereum made available 384 KiB per slot. That is 32 KiB per second. Over a day: roughly 2.6 GiB. At the six-blob maximum, about 5.3 GiB.
Post-Pectra, with a target of six and a ceiling of nine, the numbers roughly double: around 5.3 GiB per day at target, approaching 7.9 GiB at the ceiling.
Seven gigabytes a day. At the theoretical maximum. For the entire data availability budget of the largest smart contract settlement layer on earth.
A single consumer hard drive holds a full year of Ethereum's maximum DA output with room left for a photo library. That is not a criticism โ it is a design choice, and a defensible one. Ethereum is not selling storage. It is selling a verification guarantee with a very small surface area, deliberately, because every byte of available data is a byte somebody has to check.
But it does mean something uncomfortable for anyone selling DA as a business: the total addressable market, measured in bytes, is small, and the price per byte has collapsed to the floor.
Here is where the honest analyst has to resist the easy version of this argument. The easy version says "nobody uses blobs." That is false, and I want to be precise about why, because precision is the only thing separating analysis from grievance.
Blob usage is not uniformly empty. It is violently Pareto-distributed. A handful of high-throughput rollups โ one in particular, the consumer-facing L2 with the social app and the payments volume โ routinely consume a double-digit percentage of available blobspace during peak hours. On a busy day, a single chain can absorb a meaningful share of the target. Aggregate utilization figures in the 20โ40% range during active periods are not unusual, and during airdrop weeks the market has genuinely spiked, with blob fees jumping by orders of magnitude for a few hours before collapsing back to 1 wei.
So the cathedral is not empty. It has three or four congregants, and they are the ones who built it.
Which means the actual market for third-party data availability โ Celestia, EigenDA, Avail, and the rest of the 2024 cohort โ is not the high-throughput chains. Those chains post to Ethereum blobs because their security model depends on Ethereum's consensus, and because the marginal cost is already near zero. Why would they pay a third party for a weaker trust assumption at the same price?
The third-party DA market is the long tail. The dozens, then hundreds, of app-chains, rollups, validiums, and sovereign L2s that need cheap bytes and do not need Ethereum-grade guarantees. That is the customer. That is the entire thesis.
And the long tail is broke.
The Long Tail Is Broke
I spent six months in 2022 โ the bad winter, the one where I isolated myself in Denver and stopped answering consulting emails โ writing a 30,000-word analysis of Celestia's modular architecture. I called it "Sovereignty Through Separation." I believed every word of it. I still believe most of it. The insight that execution, settlement, consensus, and data availability are separable concerns is genuinely important, and data availability sampling is one of the few cryptographic ideas of the last decade that changes what a phone can verify.
But writing 30,000 words about a protocol gives you an unusual vantage point on its economics. You stop seeing the architecture and start seeing the cost structure.
Here is the cost structure of DA. Marginal cost per byte, for a sufficiently large operator, approaches the cost of bandwidth and disk โ which is to say, it approaches zero. There is no scarcity rent, because there is no scarce input. The scarce input in a rollup is not data availability. It is sequencing โ the right to order transactions and capture the MEV that ordering produces.
So we built a market where the product is cheap, the marginal cost is near zero, the differentiation is a trust assumption most buyers cannot evaluate, and at least one major competitor (EigenDA) launched by giving it away free to establish share. Then we tokenized it and expected the token to accrue value.
I have seen this movie. I audited it in 2020.
When my team and I went through Compound's governance module that summer โ four of us, remote, arguing over a screen share at odd hours โ we found a reward distribution curve that quietly routed disproportionate emissions to the earliest depositors. It was not a bug in the code. The code did exactly what it was told. It was a bug in the thesis: the assumption that a subsidy attracts users rather than attracting the appearance of users. I wrote 5,000 words about it and called it "The Hypocrisy of Decentralized Centralization," and it traveled further than anything I have written before or since, mostly because people recognized the shape of it.
Liquidity mining APY is not yield. It is a project buying its own TVL chart with its own token. Turn off the faucet and watch the number, not the community. The community stays. The number does not.
DA incentive programs have the same silhouette. Points. Blob credits. Subsidized posting. They produce usage charts that look like adoption and behave like a promotional rate. The tell is always the same: ask what happens to the number when the subsidy stops, and watch whether anyone answers with a mechanism instead of a roadmap.
What Data Availability Actually Sells
Here is the contrarian turn, and it is not the one the DA bears make.
The bear case is that DA layers are failing. The bear case is wrong. DA layers are succeeding so completely that they have become invisible โ and invisible infrastructure is structurally incapable of supporting a token.
Think about what TCP/IP is worth. It carries essentially all human commerce conducted over networks. It has no price. Not because it failed, but because it won so thoroughly that charging for it became impossible. The value migrated upward to the applications and downward to the physical layer. The protocol in the middle became a public good, funded by everyone and owned by no one.
DA is on that trajectory, and it got there in under two years, which is fast even by crypto standards. The reason is not competition. The reason is that the thing DA layers actually sell is not bytes โ it is the ability to verify without downloading, and almost nobody buying bytes cares about that.
Data availability sampling lets a light client โ a phone, a browser tab โ probabilistically confirm that a block's data was published by sampling a few dozen random chunks. That is a profound capability. It is the difference between a chain that only a data center can check and a chain that a person can check. It is, genuinely, sovereignty.
But a rollup operator buying DA is not buying sovereignty. A rollup operator is buying a cheap place to dump a batch and a security story for its bridge. The sampling is a feature they mention in docs and never use. The buyer and the beneficiary are different people, and the buyer sets the price.
I learned this lesson in a different medium. In 2021, consulting on a generative art collection, I spent three months staring at on-chain data for a thousand algorithmically produced works, trying to work out whether a token could hold an artist's moral rights. What I concluded โ in a manifesto I called "Algorithmic Authenticity" โ was that the chain preserves provenance perfectly and intent not at all. It records that a thing happened. It does not record why it mattered. You have to build that layer yourself, by hand, with people who care.
DA is the same shape. It guarantees that data was published. It cannot guarantee that anyone wanted it.
The Subsidy Reflex, and the Lightning Parallel
There is a pattern in this industry that I have watched for seven years and that I now treat as close to a law: when a piece of infrastructure is genuinely good and genuinely unmonetizable, the response is never to accept the economics. The response is to subsidize the demand side until the chart looks like product-market fit.
I watched the Lightning Network do this for seven years. The pitch was beautiful โ instant Bitcoin payments, global, permissionless. The engineering was real. And the routing layer remained a permanent tax on anyone who tried to use it at scale: liquidity that had to be actively managed, channels that went stale, payment failure rates that were tolerable in a demo and intolerable at a checkout counter. The technology did not fail. The operational burden did. Someone had to sit there and manage channel liquidity, and that someone was never going to be a merchant, and so Lightning became a niche โ genuinely useful for a specific set of custodial and remittance flows, permanently peripheral to everything else.
The lesson is not "Lightning bad." The lesson is that an infrastructure layer whose operational cost falls on the user is an infrastructure layer that will be routed around.
DA has a version of this coming. Not routing failure โ something subtler. As the number of DA providers multiplies and the price stays at the floor, the differentiation collapses onto trust assumptions and social consensus. "Who do you trust to have published the data?" is a question that most teams answer with vibes, and vibes do not survive a bear market. The chains that will still be posting to a third-party DA layer in three years are the ones whose operators can articulate, in a sentence, exactly which trust assumption they are accepting and why it is cheaper than the alternative. That is a small set. It is not a $100 million set.
And the $100 million is the actual story. Not the raise itself โ raises are weather. The story is what the raise is for. If you are a DA layer with a hundred million dollars and a product whose marginal cost is zero and whose price is at the floor, your only remaining lever is to buy demand. You fund ecosystem grants that pay teams to post to you. You run points programs that reward bytes. You become, functionally, a subsidizer of your own utilization chart.
Which is fine. It is a legitimate strategy. But it is a strategy that produces a number, not a moat, and I have audited enough of those to know what the number looks like eighteen months after the faucet closes.
The Question Nobody Is Asking
I still have the notebook. The 2017 one, from the twelve weeks I spent inside 150,000 lines of Solidity looking for the 42 flaws that were not syntax errors but trust errors โ places where the code assumed something about human behavior that turned out to be false. I kept it because the flaw pattern repeats across domains, and it repeats here.
The DA thesis assumed that blockchains would specialize, that data availability would become a purchased service, and that the resulting market would be large enough to sustain multiple competing providers with token-based value capture. The first assumption was right. The second was right. The third was never tested against a price floor of 1 wei.
What is genuinely unresolved โ and this is where I would put my attention rather than on another rollup raise โ is whether DA finds its real customer somewhere other than rollups. There is one candidate that does not require a subsidy: verifiable data provenance for machine learning. Training corpora whose lineage can be attested, whose inclusion can be proven, whose exclusion can be audited. That is a use case where the buyer, the beneficiary, and the regulator are all the same person, and where "the data was published and can be checked" is not a documentation footnote but the entire product.
I spent six months of 2026 building a prototype of exactly that with three researchers, and I will tell you honestly: the DA layer was the easy part. The hard part was everything above it โ the incentives, the governance, the human willingness to be held accountable for a dataset. The bytes were free. The conscience was expensive.
So the question I would put to the $100 million rollup, and to every DA layer with a treasury and a points program and a chart that goes up and to the right, is not "how much data are you making available." It is this: when the subsidy stops, who is still paying, and what exactly are they buying?
Zero is a price, not an absence. The floor is not a failure of the market. It is the market telling you, with perfect clarity, what it thinks the bytes are worth. The interesting question was never whether the cathedral is full.
It is whether anyone would still walk in if you stopped paying for the pews.
