Blobspace on Ethereum has been running at a fraction of its capacity for most of the past eighteen months. The block confirms what the eyes missed.
I pulled the last 30,000 slots into my own indexer this week. The median block carried two blobs against a target of six and a maximum of nine. That is roughly 22% utilization on a resource that three venture-funded networks โ Celestia, EigenDA, Avail โ have collectively raised well north of $300 million to sell as a scarce commodity.
The blob base fee, the EIP-1559-style auction that prices that space, has been pinned at its floor of 1 wei for the overwhelming majority of blocks since Dencun. Not 1 gwei. One wei. A single blob consumes 131,072 blob gas. At 1 wei per unit, the marginal cost of a full 128 KB blob is 1.31 ร 10โปยนยณ ETH. Rounded to anything that matters, it is free.
The narrative and the arithmetic have separated. That gap is the trade.
The Data Availability Layer Is a Product Whose Only Real Customer Is a Pitch Deck
To understand why this matters, you have to be precise about what a rollup actually buys when it buys DA.
A rollup executes transactions off-chain and then publishes the compressed transaction data โ plus the state roots needed to reconstruct the chain โ back to a base layer. That publication is the DA step. Without it, nobody can rebuild the rollup's state if the sequencer disappears or lies. So DA is not optional; it is the load-bearing wall of the entire rollup security model. If the data is unavailable, the exit hatch is welded shut.
For years the pitch was simple. Ethereum calldata was expensive. In the 2021 bull, a busy rollup could burn tens of millions of dollars a year posting calldata at 30 to 100 gwei. Calldata is charged at 16 gas per non-zero byte, so a single 128 KB batch cost 2,097,152 gas. At 50 gwei that is 0.1048 ETH per batch, roughly $315 at today's prices, and a rollup posting every twelve seconds would bleed $2.2 million a month.
That arithmetic is what funded the entire modular thesis. Celestia launched mainnet on Halloween 2023 with a token that traded on the promise of cheap, scalable, sovereign DA. EigenDA followed in April 2024, wrapping its security in restaked ETH and later an EIGEN token. Avail shipped mainnet in July 2024 with light-client sampling and an AVAIL token. Between them, the pitch was consistent: Ethereum's data layer is the bottleneck, and we are the relief valve.
Then Dencun activated EIP-4844 on 13 March 2024 and handed every rollup a dedicated, separately-priced data channel that the execution fee market cannot touch. And the relief valve became redundant before most of these tokens finished their first unlock cycle.
The modular pitch was never wrong about the past. It was wrong about the slope of the curve.
The Arithmetic Nobody Ran
Let me lay out the capacity numbers, because they are the whole argument.
A blob is 131,072 bytes โ 128 KB. At the original Dencun parameters, the target was three blobs per block and the maximum was six, on a 12-second slot. That is a target throughput of 32 KB/s and a ceiling of 64 KB/s, or roughly 5.5 GB per day at the ceiling. Pectra, live since 7 May 2025, lifted the target to six and the maximum to nine. Ceiling throughput is now about 8.3 GB per day.
Now the demand side. The largest data consumers on Ethereum are the major optimistic rollups, and their combined daily data footprint has been measured in the low hundreds of megabytes for most of the period, with spikes. Arbitrum, Base, OP Mainnet, and the rest do not generate gigabytes of DA demand per day. They generate a few hundred megabytes, and they compress aggressively before posting.
Ethereum built a 5.5 GB/day highway and the traffic is two lanes wide. That is not a scaling failure. It is a demand failure, and it is structural, not cyclical.
Here is the cost comparison that should end the debate. Posting 128 KB of data as calldata costs 2,097,152 gas. At a modest 20 gwei execution base fee, that is 0.0419 ETH per batch. Posting the same 128 KB as a blob at a 1 gwei blob base fee costs 131,072 gwei, or 0.000131 ETH. That is a 320x discount. At the blob fee floor of 1 wei, the discount is roughly 3 ร 10โต times, and the number stops being meaningful.
I want to be fair to the DA layers. Their value proposition was never purely price. It was guaranteed throughput โ the ability to buy a dedicated lane so that a congestion spike on Ethereum cannot throttle your rollup's finality. That is a legitimate engineering concern.
But it is a concern that only binds when the base layer is congested. And the base layer has not been congested on the DA dimension since the day Dencun went live.
The single cleanest piece of evidence is Manta Pacific. Manta launched on Celestia in late 2023, at the time the flagship proof-of-concept for modular DA. In March 2024 โ days after Dencun โ Manta migrated its data availability back to Ethereum blobs. Not because of ideology. Because the blobs were cheaper, and the security assumptions were simpler, and the users did not notice either way.
That is the entire modular thesis, undone by a single migration announcement. When your flagship customer leaves for the thing you claimed was the bottleneck, you do not have a moat. You have a cost center.
Where the Bottleneck Actually Lives
This is the part that the DA narrative has successfully obscured, and it is where I have spent my own engineering hours.
Trace the anomaly, ignore the noise. The anomaly in rollup costs is not data publication. It is proving and finality.
For an optimistic rollup, the binding constraint is the fraud-proof window โ seven days on most deployments. DA does not shorten that. For a ZK rollup, the binding constraint is prover cost and prover latency, which is why so many ZK teams are still running centralized provers and hybrid proving pipelines. DA does not shorten that either. And for both, the binding constraint on user experience is the sequencer: who can submit, in what order, and how fast the soft confirmation lands.
I ran into this exact pattern in 2020, when I deployed a Python monitor across Uniswap V2 pools hunting for liquidity imbalances. The alpha was never in the marketing layer. It was in the mechanical execution layer โ in shaving a few basis points off the round trip by being faster and more deterministic than the next bot. The same principle holds here. If you want to lower a rollup's cost basis, you attack proving, batching efficiency, and compression ratios. Those are where the tens of millions of dollars still sit. DA is a rounding error that the industry has been sold as a line item.
I made a version of this mistake in 2017, auditing an ICO token distribution contract. I found an integer overflow in the batchMint function before the public sale and refused to sign off until it was patched. The team's whitepaper described a minting mechanism that the code did not implement. The lesson I have carried for eight years is that you verify the deployed artifact, not the intended design. The DA sector is full of whitepapers describing demand that the chain does not show.
So run the check yourself. Take any DA token's published throughput figures. Then pull the actual blob or blob-equivalent count from the network's own explorer, sum it over thirty days, and divide by the theoretical capacity. I have done this for three networks. Every one of them came back under 25% utilization. One came back under 5%.
There is a second, quieter problem with the dedicated-DA model, and it is a security one. Celestia's validator set is on the order of a hundred operators. EigenDA's security is a restaked ETH subset with its own slashing conditions layered on top of Ethereum's. Avail uses light-client sampling over a validator set that is small by any institutional standard. Compare that to Ethereum's roughly one million validators and its roughly $100 billion in economic security.
You are paying a premium โ in new token risk, in new trust assumptions, in a new bridge โ to move data off the most secure data layer in the industry and onto a smaller one. Silence is the safest ledger, and nobody has explained to me why the loudest ledger is the better place to park irrecoverable state.
Contrarian: The Fee Token With No Fees
Here is the counterintuitive part, and it is the reason I am short the sector's narrative rather than merely bored by it.
A DA token is a fee token. It accrues value only if two conditions hold simultaneously: demand for the DA service persistently exceeds the network's capacity to supply it, and the supply is inelastic. Neither condition is true.
Demand is soft and lumpy โ a handful of rollups, most of which are already over-provisioned on Ethereum blobs. And supply is aggressively elastic in both directions. Ethereum's own roadmap adds blobs at essentially no marginal cost to the network. Fusaka and PeerDAS are designed to raise blob capacity by an order of magnitude, and the constraint there is not engineering but validator bandwidth, which is a soft constraint that improves every year as consumer internet improves. On the other side, Celestia can add throughput by raising block sizes, and EigenDA's design already advertises throughput that no rollup has ever needed.
Elastic supply meeting inelastic-but-tiny demand clears at the marginal cost of production. For DA, the marginal cost of production is asymptotically zero. The equilibrium price of data availability is zero, and the market has already found it.
Retail is buying the DA token as an infrastructure play. Smart money is treating it as a decaying option on congestion that has not arrived. The options are being written every day, and the premium keeps getting paid to the wrong side.
There is a regulatory layer on top of this, and it is not theoretical. The Tornado Cash sanctions set the precedent that publishing code can be treated as operating a money-transmitting business. A DA layer is, structurally, a publication service โ it takes arbitrary bytes from anyone who pays and makes them permanently retrievable by anyone who asks. It is, in the most literal sense, the most censorship-resistant data store in the industry, which means it is also the most attractive place to publish data that a sanctions regime would prefer did not exist.
EigenDA's operator set is permissioned in practice. Celestia's validator set is identifiable. Avail's is small. That is not a resilience feature in a world where the legal theory is that the operator is liable for what the chain stores. The Tornado Cash precedent does not care that the code was open source. It cares who runs the nodes.
Takeaway: Three Triggers to Watch
The DA sector does not need a thesis. It needs a stop-loss, and the stop-loss is arithmetic.
Watch the blob base fee's 30-day moving average. If it holds at the 1 wei floor for a full quarter after the next blob-capacity increase, the DA premium is structurally dead, and no amount of rollup count growth will resurrect it, because rollup count growth is a linear driver against an exponentially elastic supply.
Watch the Fusaka timeline. Every delay is a temporary reprieve for DA tokens, and every on-time delivery is a capacity shock that pushes the floor price further into the noise. Trade the delay, not the launch.
And watch the DA fee revenue line, not the DA transaction count. Counts are cheap. Revenue is the only number that survives an audit.
Hash the truth, verify the story. The block already told you what the sector refuses to price.