I remember the exact moment the number stopped making sense.
It was 1:40 in the morning in Denver, the kind of late-March night where the snow has finally stopped falling but the cold still has opinions. I had a block explorer open on one monitor and a blob-fee chart on the other, and between them a half-finished draft of a review I had agreed to write quietly, without my name attached, for a team that had raised nine figures and had not yet shipped a token. Their landing page used the word sovereignty eleven times. Their new roadmap promised a data availability layer that would, in their own phrasing, unlock the modular century.
The chart on the second monitor was the thing I could not look away from.
Over the preceding thirty days, that network's blobspace โ the actual payload, the actual bytes somebody had paid to have published and made retrievable โ had averaged utilization under two percent of capacity. Not two percent of some theoretical maximum that engineers like to gesture at in conference talks. Two percent of what the team was already provisioning, already paying validators to store, already describing in investor decks as the throughput engine of the next phase of the internet.
I sat with it for a while. Then I did what I always do, which is to go looking for the bug.
There wasn't one. That was the problem.
I am forty-two years old and I have spent the better part of a decade reading other people's code for a living, and the single most important thing that work has taught me is that the dangerous defects are almost never the ones that break the machine. The dangerous ones are the defects that make the machine run exactly as designed, flawlessly, for a purpose nobody said out loud. In 2017 I spent twelve weeks inside a DAO's successor codebase โ a hundred and fifty thousand lines of Solidity, forty-two critical logic flaws, and not one of them a syntax error. Every one of them was a trust assumption wearing a function signature. I submitted the findings to a public repository and watched a community argue for three weeks about whether a bug that had never triggered was still a bug.
It was. And so was the two percent.
A protocol running at two percent of capacity did not fail. It succeeded at something other than the thing it claimed to be doing.
In a flat market, somebody would have asked about that. In the market we are standing in right now โ this particular spring of leverage and re-acceleration, where capital moves faster than comprehension โ the two percent is not a warning. It is a runway. You can raise on a runway. You cannot raise on a warning.
What a Data Availability Layer Actually Sells
Let me back up, because the number only means something if you understand what a data availability layer sells, and almost nobody selling one has an incentive to explain it plainly.
A blockchain is a machine for agreeing on the order of things. That is all it is. Everything else โ the tokens, the governance forums, the philosophical pamphlets โ is downstream of the ordering. And ordering only works if the data behind a state transition was actually published somewhere. If a block producer orders a transaction, takes your money, updates the ledger, and then refuses to tell anyone what was inside the block, you have a ledger that no longer exists for anyone but the producer. This is the data availability problem, and it is genuinely serious. It is the load-bearing wall.
The classic response is to make every node download everything. Elegant, verifiable, and completely unscalable. The modular response, the one that consumed the 2022 and 2023 research cycles, was to split the wall out from the building: let execution live on one chain, settlement on another, consensus somewhere else, and let a specialized layer handle the publication guarantee. Celestia took that thesis further than anyone and gave it a name people could rally around. When I was thirty-eight, freshly bloodied by a market collapse, I locked myself in a Denver apartment for six months and wrote thirty thousand words about it. I called the analysis Sovereignty Through Separation and I believed every paragraph. I still believe the architecture is elegant. That is not the same as believing the market exists.
The mechanism a modern DA layer sells is usually three things stapled together. Erasure coding, which extends a block of data into a larger square so that any sufficient fraction of it can reconstruct the whole. Capacity, measured in megabytes per block or bytes per second. And data availability sampling, which lets a light client download a handful of random chunks from the extended square and conclude, with a quantified probability, that the rest was published. You do not verify the data by downloading it. You verify that someone would have been caught if they hadn't published it. The security of a DA layer is not the security of the data. It is the security of the threat of detection.
That distinction is where the whole industry has been quietly bleeding.
Meanwhile, on the thirteenth of March in 2024, Ethereum activated proto-danksharding. Blobs arrived. A blob is a hundred and twenty-eight kilobytes of data attached to a block, priced by its own fee market, pruned after roughly eighteen days, and โ critically โ never executed by the EVM. It is a publication primitive. It is, in the most literal sense, a data availability layer bolted directly onto the largest and most liquid settlement domain in the world, available to every rollup already living there.
Ethereum's initial parameters targeted three blobs per block with a maximum of six. At twelve-second slots, that works out to roughly thirty to thirty-two kilobytes per second of guaranteed throughput, with a ceiling around twice that. In May 2025, the Pectra upgrade doubled the target to six and raised the ceiling to nine. Further expansion was scheduled behind PeerDAS and the Fusaka line of work, with sampling pushed down to the networking layer so that the capacity could grow without every node paying the bandwidth bill.
Read those numbers again and then read the marketing material of a dedicated DA layer. The dedicated layer will show you design targets in tens of megabytes per second. Terabytes per day. It will show you a chart with Ethereum's line flat and their line vertical. It will not show you the price.
Because the price is the whole story.
The Price Floor Problem, Which Nobody Puts on the Landing Page
Here is the arithmetic that took me three columns of a legal pad to work through, and that I now suspect is the actual reason the two percent number exists.
Blob gas on Ethereum has its own market with its own minimum. That minimum is one wei. A single blob consumes 131,072 blob gas. Multiply one by 131,072, convert through the prevailing gas price, and you land at an amount so small that a rollup posting a batch every few minutes pays, in a quiet week, less than the cost of a cup of coffee. Even during genuine contention, when blob base fees have climbed by orders of magnitude, the absolute cost stays in the range that a mid-sized rollup can absorb out of pocket.
Now put yourself in the position of a dedicated DA layer trying to sell blockspace.
Your product is a publication guarantee. So is Ethereum's. Ethereum's is backed by the largest validator set in the industry, secured by the most expensive asset in the industry, and integrated into the settlement domain where your customers already operate. And Ethereum has deliberately priced it near zero, because Ethereum's own roadmap wants the data to be cheap.
The competitor of a dedicated DA layer was never another DA layer. It was Ethereum deciding that availability is a public utility rather than a product.
I have watched a dozen pitch decks that compare Celestia to EigenDA to Avail, arguing about sampling parameters, finality times, and fault-proof windows. All of it is real engineering and most of it is good engineering. None of it answers the question the customer actually asks, which is: why would I move my data off the cheapest settlement domain in the world in order to pay somebody else to hold it?
The honest answer โ and I have given this answer to founders who did not enjoy hearing it โ is safety margin. A rollup running on Ethereum blobs is exposed to blob fee spikes during congestion. A rollup on a dedicated layer gets a separate market with different dynamics, a bigger block, and a sovereign escape hatch. That is a legitimate argument. It is also an argument worth a few basis points of cost, not an argument that sustains a nine-figure treasury.
And here is the deeper cut. The blobspace on Ethereum is itself chronically underused. Not just the dedicated DA layers. Ethereum. The demand that everyone assumed would arrive โ the throttling wall of rollup data that would force modular separation โ never showed up at anywhere near the projected volumes. In the weeks after proto-danksharding went live, blob fees collapsed to the floor and stayed there for long stretches, punctuated by brief frenzies of what was essentially speculative data inscription. A capacity constraint that the entire 2022 research cycle was built on turned out to be a capacity surplus.
I wrote about separation and sovereignty for thirty thousand words. The market read it and said: publish it here, it is cheaper.
That is not a betrayal. That is information.
The Two Percent Is Not Organic Demand
Back to the two percent.
When I finally sat down with the raw usage data instead of the throughput charts, the composition of that two percent was more interesting than its size. Because the two percent was not uniform. It came in bursts, and the bursts correlated with something that had nothing to do with rollup economics.
The bursts correlated with incentive programs.
A project announces a points season. Points are redeemable for a token that does not yet exist. To earn points, you deploy a rollup onto the DA layer, or you post data through the DA layer's documented path, or you run a node that registers activity. Within a week, usage spikes. Within a month, usage compresses back to baseline. The data being published during the spike is, in a meaningful fraction of cases, not the output of any production system. It is the minimum viable payload required to have a signed record of participation on a chain somebody else is watching.
I spent four years of my life explaining why this pattern is a subsidy wearing the costume of demand. In 2020, working with a team of four on a governance module audit, we found a reward distribution algorithm that quietly favored the earliest addresses โ a design that contradicted the protocol's own manifesto in the most literal way possible, not through malice but through an ordering assumption that nobody had interrogated. I wrote five thousand words about it and it traveled further than anything I had written before. The mechanism there was subtler than what I was now looking at. The mechanism here is not subtle at all.
When a data availability layer's utilization curve is a series of spikes that decay back to a floor, you are not measuring demand. You are measuring the half-life of a subsidy.
Stop the points. Stop the grants. Stop the ecosystem fund that pays teams to migrate. The bytes stop moving, and they stop moving within one programmatic interval, because the entities generating them were never solving an availability problem. They were farming a claim on a future balance sheet.
I say this as someone who believes deeply in the idea of open, permissionless publication infrastructure. I also say it as someone who has now watched the same shape recur across four separate cycles: liquidity mining that subsidized TVL and evaporated at the withdrawal of the tap; the NFT volume that was wallet-to-wallet circularity rather than collecting; the DA utilization that is airdrop-adjacent rather than rollup-driven. The industry has a remarkable capacity to mistake a subsidy for a signal, and an equally remarkable capacity to build an entire valuation model on the mistake before the tap is turned off.
The uncomfortable corollary is this: if you strip the subsidized bytes out of the utilization charts, the surviving residue is not two percent. In several of the datasets I examined, it is under one percent, and it is concentrated in a handful of counterparties who are mostly there for reasons of ideology or integration lock-in rather than raw cost.
The Security Budget Question Nobody Underwrites
The other thing the landing pages will not show you is what actually secures the guarantee.
For a layer using an independent consensus โ the Celestia model โ the answer is a validator set with a native staking asset, and the honest question is whether the value securing the chain is proportional to the value being published on it. This is the same question that has haunted every proof-of-stake system since the beginning, and it has no clean answer. A chain securing two billion dollars of stake while guaranteeing availability for ten billion dollars of rollup value is running a security budget deficit that only becomes visible during an attack, which is exactly when you least want to discover it.
For a layer using restaked security โ the EigenDA model โ the answer is more sophisticated and correspondingly more fragile. You borrow security from an existing asset pool by letting that pool commit to additional slashing conditions. The pitch is that the marginal cost of securing a new service approaches zero, because the capital is already there. I have read the papers. The elegant part is real.
The part that gives me the 2017 shivers is the correlation structure.
Restaking does not create security. It rehypothecates it, and rehypothecation always looks like free money until the day the correlation matrix turns out to have been diagonal only in the spreadsheet.
When I audited that DAO's successor codebase, the flaws that mattered were never in the arithmetic. They were in the assumptions stacked above the arithmetic โ the unstated belief that two independent-looking conditions could not both be true at the same time. Restaking has the same shape. The same capital secures several services. In normal conditions, those services fail independently, and the shared pool absorbs the loss cleanly. In a genuine stress event โ a deep market drawdown, a cascading liquidation, an oracle failure that touches several venues at once โ the services fail together, the slashing conditions trigger together, and the pool that was supposed to absorb one shock is asked to absorb four. The marginal cost of securing an additional service was never zero. It was deferred.
This is not an argument that restaking is broken. It is an argument that the risk premium attached to restaked availability should not be the risk premium attached to the underlying asset, and that most of the pricing I have seen does not reflect the difference.
There is a second, quieter problem, and it lives inside the sampling parameters.
Data availability sampling offers probabilistic security. A light client samples a fixed number of chunks from the erasure-coded square. If the samples succeed, the client concludes that the data was published with high probability. That word โ probability โ is doing an enormous amount of work. The security of the whole construction rests on assumptions about the number of honest samples, the independence of the sampling, and the adversary's inability to predict which chunks a given client will request. I have seen three separate implementation write-ups where the sampling procedure was described as random in the prose and implemented, in practice, with a seed derived from something an adversary could partially observe. None of those were catastrophic. All of them were the kind of thing that becomes catastrophic only after somebody reads the code with hostile intent.
The DA layer's real product, in the end, is not bytes. It is a quantified guarantee about a piece of data. And a quantified guarantee is only as good as the adversarial model handwritten in the appendix, which is the part of the paper nobody screenshots.
The Amortization Illusion
There is one more piece of arithmetic worth putting on the table, because it is the argument I hear most often from the technically serious people defending the space, and it deserves a serious answer. The argument goes: availability is not expensive per byte if you amortize the erasure coding across the square, if you prune aggressively, if you let the data get cold. Cheap DA is simply a matter of running the redundancy factor and the retention window correctly.
That argument is true for the wrong kind of data.
Erasure coding with a two-times extension in each dimension gives you four times redundancy. To publish two megabytes of usable data, you store and serve eight. That is the price of availability, and it is a fine price for cold data โ the historical batch records, the state diffs from six months ago, the archive nobody reads but everyone wants to know exists. For cold data, a dedicated DA layer is genuinely efficient, and Ethereum's eighteen-day pruning window makes the comparison flattering on the margins.
But rollups do not only post cold data. The data that makes a rollup usable โ the recent state, the batch that a sequencer needs to process the next withdrawal, the proof inputs that a fault-proof game will need to reference within the challenge window โ is hot data. Hot data is measured in retrieval latency, not in stored bytes. And every DA layer I have examined, without exception, prices retrieval the way it prices storage: a bandwidth-shaped cost structure on top of a capacity-shaped cost structure on top of a sampling-shaped assurance that the bytes are there at all.
Dedicated DA makes cold data cheaper and hot data no cheaper at all, and almost every rollup's demand curve lives on the hot side.
This is why the migration arguments so often fail at the pilot stage. The pilot posts historical batches to the new layer, the cost comparison looks excellent, and the board approves. Then someone tries to route live sequencer traffic through the same path, the latency budget blows out, and the project quietly keeps a hot path on the settlement domain while the cold data lives elsewhere. The DA layer gets a line item and a press release. It does not get the throughput.
I have now watched this exact cycle three times with three different teams. The pattern is stable enough that I have stopped treating it as an execution problem and started treating it as a product problem.
The Turn Nobody Priced: AI Does Not Want Availability
Here is where the story inverts, and this is the part I actually came to write.
In 2026 I led a six-month open-source initiative to build a verifiable training dataset on-chain. Three researchers, one protocol, one very stubborn question: could a blockchain function as a provenance layer for machine learning, guaranteeing that a model was trained on data whose origin, licensing, and transformation history were auditable? At forty-two, I took the project on with genuine fear. I have spent my career arguing that blockchains preserve human intent. Artificial intelligence is the first technology I have encountered that can convincingly simulate intent without possessing any.
The question I kept circling was whether the thing I had spent my life on could serve as a truth layer for a system designed to produce plausible untruths.
Six months in, the answer is more precise than yes or no, and it has direct implications for every DA layer currently repricing itself as AI infrastructure.
What we actually needed, technically, was three things. We needed commitments โ a Merkle root or a polynomial commitment that pinned a specific dataset to a specific version at a specific time, so that no party could later swap the training corpus and claim it was always that way. We needed a provenance chain โ a signed, ordered record of who contributed which shard, under what license, with what preprocessing applied, so that a biased or poisoned upload could be traced to its source and excluded retrospectively. And we needed retrievability for auditors โ a way for a third party, months later, to obtain the actual bytes and re-verify the commitments.
The first two are signature problems. The third is an availability problem, but a very particular one.
AI does not need a data availability layer. It needs a provenance layer, and provenance is a cryptographic commitment, not a bandwidth guarantee.
Our commitments were on the order of hundreds of bytes per dataset version. Our provenance records were a few megabytes over six months. Our retrievability requirement covered a few terabytes of cold, immutable, rarely-read shards that will be read perhaps a dozen times in their lifetime, by auditors, under no latency pressure whatsoever.
We did not use a dedicated DA layer. We used a content-addressed store with an availability committee, a challenge mechanism, and the commitments anchored on a settlement chain. The total cost was a rounding error. If we had routed the same workload through a throughput-optimized DA layer, we would have paid for megabytes per second we would never consume, in exchange for a sampling guarantee we did not need, to solve a latency problem we did not have.
Every DA layer I have seen pivot toward AI in the last four quarters has made the same category error. They are selling capacity to a market that wants attestation. They are selling bandwidth to a market that wants memory.
The pivot is not wrong because AI will not need verifiable data. The pivot is wrong because verifiable data is a much smaller, much colder, much more signature-shaped problem than the capacity charts suggest. And a smaller problem supports a smaller business. Which is fine, if you stop valuing it like a large one.
The Lightning Test
There is a test I apply to any infrastructure thesis before I write a word about it, and I borrowed it from the most instructive failure in this industry. I call it the Lightning test.
The Lightning Network has been functionally half-dead for seven years. Not dead โ functional, in production, capable of moving real value for real people in real seconds. And yet the routing failure rate on any non-trivial payment remains high enough to break the user experience, the channel management burden remains a full-time job for anyone who wants reliable liquidity, and the population of people who will tolerate that burden in exchange for the benefit is, seven years in, still small and still heavily self-selected for ideological commitment.
The technology did not fail. The complexity budget did. The cost of using it exceeded the value of using it for everyone except the people who were already convinced.
I apply that test to DA layers and the shape is uncomfortably familiar. The technology works. The erasure coding is sound. The sampling is clever. The validator sets are real. And the population of teams who will tolerate running a second consensus client, managing a second fee market, integrating a second finality assumption, and re-underwriting a second security model โ in exchange for a marginal cost improvement over a domain they are already operating in โ is small, and heavily self-selected for teams that raised a token against the migration thesis.
The pragmatic test is not whether a DA layer can do the job better. It is whether anyone who is not being paid to care will care.
And here is where I have to be careful, because the honest answer is not the cynical one. I do not think the DA thesis is dead. I think it was mispriced and mis-sold, and those are different failure modes with different remedies. The mispricing came from building a business on a capacity constraint that the settlement layer dissolved from underneath it. The mis-selling came from describing the product as throughput when the product is a guarantee.
Because nobody actually wants bytes. Nobody has ever wanted bytes. What people want is to know that a specific piece of data was published at a specific time, that it cannot be retroactively withheld, and that if it were withheld, someone would have been provably caught. That is a small market. It is also a defensible one, and it is not currently served well by anybody, because everybody is busy selling megabytes.
The blind spot is not that DA layers are useless. The blind spot is that the industry measured its demand in the wrong unit for three years, and built the wrong capacity to serve it, and is now raising on the resulting charts. The teams that survive will be the ones who stop counting bytes and start counting guarantees โ who narrow to the counterparties with an actual withholding problem, price for cold immutable retrieval instead of hot throughput, and accept that they are selling insurance rather than bandwidth.
Insurance is a smaller business than bandwidth. It is also a business where the underwriting actually means something, which is more than can be said for a network that is ninety-eight percent empty and calling it capacity.
What I Am Watching Next
The other night, I closed the block explorer and left the blob chart open. It was still sitting near the floor. On the other monitor, the roadmap draft still said sovereignty eleven times, because the eleventh instance was in the footer and nobody had looked at the footer.
I have stopped being angry about that. Anger is a young man's tool and I have spent it. What I have instead is a specific and durable curiosity about a question that this cycle will answer whether it wants to or not: when the subsidy taps close and the points seasons end and the token unlocks land, what is the residual demand for a publication guarantee that costs more than free?
If the answer is small, then the DA layer was never an infrastructure business. It was a well-engineered answer to a question the settlement layer had already answered more cheaply, and the industry mistook the elegance of the question for the existence of the market.
If the answer is small but nonzero, and concentrated in a few dozen counterparties with a genuine withholding threat โ a bridge, a cross-domain proof system, a regulated data attestation, a provenance chain for models that will be audited by someone with subpoena power โ then the DA layer becomes what it should have been all along. Not a throughput engine. A notary.
Notaries are not exciting. They do not raise nine figures. They do not put the word sovereignty eleven times on a landing page.
They just do the one job that nobody else can do, and they do it for the people who genuinely need it, and they are still there in the winter.
I know which version I would rather underwrite. I am less sure the market agrees with me yet, and I suspect the disagreement will be resolved by an earnings report rather than an argument.
So here is the question I am carrying into the rest of this year, and the one I would put to every founder currently raising against a utilization chart: if you had to price your layer as though the bytes were free โ because, increasingly, they are โ what would you be selling, and to whom, and would it survive the withdrawal of the subsidy that is currently the only thing standing between your network and its own two percent?