For most of the eighteen months that followed March 13, 2024, the Ethereum blob base fee printed 1 wei. Not one gwei. One wei — the smallest integer the protocol's fake_exponential will return, the floor below which the auction physically cannot fall. Three blobs of 128 KiB each, 393,216 units of blob gas, settled for 393,216 wei: roughly four ten-thousandths of a cent at any ETH price above $2,000.
The market read that number as a victory. Rollup fees collapsed by an order of magnitude within weeks. The modular thesis appeared to have been delivered on schedule. The phrase "Ethereum as the settlement and data availability layer" migrated from conference slides into production architecture. Every metric the industry watches — L2 transactions per second, cost per transfer, sequencer margin — improved in the same direction at the same time.
The number is not a victory. It is a price signal that has been disconnected from scarcity, and the disconnection is structural rather than temporary. When a resource clears at the protocol's minimum for the overwhelming majority of its blocks, the fee market is not clearing demand — it is suppressing the information that would tell builders to ration, and it is accumulating an obligation that will be called later. The invoice exists. It has simply not been delivered.
I have spent a career reading price mechanisms that look like this. In 2022, I built a simulation of Terra's seigniorage loop to answer one narrow question: under what conditions could an algorithmic stablecoin hold a peg without external collateral? The answer was that the system required growth at a rate compounding arithmetic cannot sustain indefinitely. The mechanism worked perfectly — right up to the moment it didn't, and then it failed inside a single afternoon. Blob pricing is not Terra. But it shares one property with it: the elegant mechanism is doing exactly what its specification says, and the specification encodes a terminal condition.
What follows is a teardown of the blob fee market as an economic object rather than a product feature. The proof is in the logic, not the promise.
Context: what EIP-4844 actually shipped
Proto-danksharding, EIP-4844, activated with the Dencun hard fork — Deneb on the consensus side, Cancun on the execution side. It introduced a new resource to Ethereum: the blob. A blob is 4096 field elements of 32 bytes each, which is 131,072 bytes, which is 128 KiB. Blobs carry rollup transaction data, are committed to via KZG polynomial commitments, and are addressable from the execution layer only through a point-evaluation precompile at address 0x0A. The EVM cannot read blob contents. The execution layer sees a commitment and nothing else. The data itself lives in the consensus layer, attached to the beacon block.
The fee mechanism is deliberately separate from the execution fee market. Blob gas has its own base fee, updated per block by an exponential function with a floor of 1 wei. Before Pectra, the target was 3 blobs per block and the maximum 6. Pectra, in May 2025, raised those to 6 and 9 respectively under EIP-7691 and doubled the fee update fraction from 3,338,477 to 5,007,716. Blob retention is 4096 epochs — eighteen days — after which the data is pruned from consensus nodes. This last detail will matter later, and almost nobody quotes it.
Before 4844, rollups posted their data as calldata, competing directly in the same gas auction as every DeFi swap and every NFT mint. Data typically accounted for 80% to 95% of a rollup's L1 cost, which meant rollup users were effectively paying Ethereum's block-space auction twice: once for execution on L2, once for data publication on L1. After 4844, that second payment moved to a separate and much cheaper auction. Rollup fees fell by 10x to 100x. The narrative wrote itself.
That narrative is largely correct as a description of what shipped. It is incomplete as a description of what was priced. A fee floor is not the same thing as abundance, and a separate auction is not the same thing as a functioning market. The distinction between "cheap" and "abundant" is the entire argument of this piece, and the industry has been treating them as synonyms.
The blob market is not an auction. It is an exponential penalty function.
Start with the mechanism, because the mechanism is where the intuition fails.
EIP-1559's execution base fee targets a gas quantity and adjusts by a bounded amount per block — up to 12.5% in either direction. The adjustment is bounded because the mechanism wants to converge smoothly. The blob fee does not work this way. Blob pricing tracks a quantity called excess_blob_gas, which accumulates when the blobs used in a block exceed the target and decays when they fall below it. The blob base fee is then fake_exponential(1, excess_blob_gas, BLOB_BASE_FEE_UPDATE_FRACTION) — an exponential of the accumulated excess, scaled by a constant.
The practical consequence is that the fee is flat at 1 wei for as long as excess is zero, and then it climbs exponentially in the accumulated excess. Pre-Pectra, a single block at maximum load (6 blobs against a target of 3) added 393,216 to the excess, and the update fraction of 3,338,477 meant the fee rose by roughly e^(393216/3338477) - 1, about 12.5%. Sustained over-target load compounds. Post-Pectra, the same block adds the same excess but the larger update fraction softens the per-block step to roughly 8%.
Now run the numbers forward. To move the blob base fee from 1 wei to 1 gwei — a factor of one billion — the exponent must reach roughly 20.7. Pre-Pectra that requires about 69 million units of excess blob gas. At 393,216 excess per full block, that is about 176 consecutive blocks at maximum load. At 12 seconds per slot, approximately 35 minutes of sustained maximum blob usage is enough to move the blob fee from the protocol floor to one gwei.
Thirty-five minutes. That is the entire ramp. The market does not glide toward an equilibrium price that reflects the marginal cost of blobspace. It sits at the floor, and then it spikes, and then it decays back to the floor as the excess bleeds off. There is no price discovery in the normal sense, because the fee is a function of accumulated history rather than instantaneous demand. What the blob market produces is a series of excursions away from zero, not a curve.
I want to be precise about why this matters. A market that clears at its floor most of the time and spikes occasionally sends a very specific signal to builders: do not budget for data cost, because data cost is approximately zero. That signal is rational given the historical data. It is also the signal that guarantees the spikes, because no one builds slack into a system that has never charged them for slack. This is a tragedy of the commons with an exponential penalty clause, and the penalty lands on whoever happens to be consuming blobspace at the moment of the spike.
Capacity: the target is a budget, not a ceiling
Post-Pectra, the target is 6 blobs per block and the maximum is 9. At target, that is 768 KiB per 12-second slot, which is 64 KiB per second, which is roughly 5.3 GiB per day. At the maximum of 9 blobs, it is roughly 7.9 GiB per day. Those numbers sound large until you price them against what a mature rollup ecosystem actually needs.
A simple transfer compressed for rollup data costs somewhere between 10 and 20 bytes. A swap costs more. Call it 30 bytes for a mixed workload. At 5.3 GiB per day, Ethereum's blobspace could theoretically carry on the order of 190 million compressed transactions per day. That is a very large number, and it is the number the bulls cite. It is also irrelevant, because the constraint is not the theoretical maximum — it is the willingness to pay the exponential penalty for exceeding the target.
The practical throughput of Ethereum blobspace is bounded by the fee curve, not by the blob count. If sustained demand exceeds the target, the fee climbs until some marginal consumer exits. That consumer is the one with the thinnest margin, which in a rollup ecosystem is whichever chain is running the least differentiated application mix. So the capacity question is not "how many blobs can Ethereum carry" but "at what price does the marginal rollup migrate to a substitute."
The substitute ceiling: DA is a commodity and its price is set elsewhere
Here is where the simple version of the bull thesis — "blob fees will double, therefore rollups get more expensive" — breaks down, and where I want to correct my own prior framing.
Data availability is not a monopoly. Celestia sells it. EigenDA sells it, secured by restaked ETH. Avail sells it. Near DA sells it. The prices are competitive and the switching cost, while real, is not infinite. If Ethereum blob fees rise far enough, a rollup can migrate its data publication to an alternative DA layer and settle to Ethereum only periodically, accepting a downgrade in security assumptions — the validium trade — in exchange for a lower data bill.
That means Ethereum's blob price has a ceiling, and the ceiling is set by the marginal cost of the cheapest credible alternative. It does not trend to infinity. The correct claim is not that blob fees will explode; it is that blob fees will normalize upward toward the price of the substitute, and the substitute is cheap. Yields are just risk wearing a tuxedo, and DA fees are just security assumptions wearing a price tag. When you compare two DA layers, you are not comparing bytes. You are comparing the cost of the assumptions each one asks you to make.
There is a counter-argument, and it is the strongest one in the bulls' favor. Ethereum's DA has a property no substitute replicates: it inherits Ethereum's security budget and its data is available to the same consensus that secures the execution layer. The optimistic rollup fraud window is 7 days. Blob retention is 18 days. Seven is less than eighteen, so blobs are sufficient for fraud proofs, and a fraud proof that can reach the data it needs is a fraud proof that means something. Alt-DA breaks that property unless the rollup adds a separate availability guarantee, which is exactly the complexity that erodes the cost advantage.
So the honest model has both a floor and a ceiling. The floor is the protocol minimum of 1 wei. The ceiling is the cost of the best substitute. The fee oscillates between them, and the interesting question is not the level but the frequency of excursion. Assume malice, verify everything, trust nothing — including my own prior estimate that fees will double within two years. The more defensible claim is narrower and harder to dispute: the blob fee will spend most of its life at the floor and a small, growing fraction of its life in excursions, and the growth of that fraction is the metric that matters.
Eighteen days is the number nobody quotes
Now the part that marketing omits, and the part that made me write this at all.
Blob data is retained for 4096 epochs. That is eighteen days. After eighteen days, consensus nodes prune it. This is not a bug; it is a deliberate design choice inherited from the danksharding roadmap, which assumed that data availability is a short-horizon guarantee — long enough for fraud proofs, short enough to keep node storage bounded.
For fraud proofs, eighteen days is sufficient. The fraud window is seven. Fine.
For node synchronization, eighteen days is a cliff. A new rollup full node that wants to reconstruct historical state from L1 must read the transaction data that produced that state. If that data was published as blobs and the blobs are more than eighteen days old, the data is gone from consensus. The node cannot reconstruct. It must trust a snapshot, a third-party archive, the sequencer's own database, or some combination of parties whose incentives it has not verified.
This is the same class of problem EIP-4444 addresses for execution history — history expiry, the Portal Network, era files, the whole apparatus of "we stop storing this, and someone else stores it, and we hope the someone else is honest and funded." The blob layer has the same problem with a shorter clock and less discussion.
For ZK rollups with validity proofs, the state is provable regardless of whether the preimage data survives. That is a genuine advantage. But the preimage data is still the thing you need to produce the next proof from a cold start, and "provable" is not the same as "reconstructible." I have made this argument before, in 2021, when I looked at the metadata storage of Bored Ape Yacht Club and found that the "decentralized" art ownership depended on IPFS pinning services with payment thresholds. The community called me a bot. The finding did not change. Static analysis reveals what marketing hides, and what it hides here is that "Ethereum as a DA layer" is a claim with an eighteen-day shelf life. After eighteen days it becomes "Ethereum as a DA layer plus a set of unverified assumptions about who kept the data and why."
Ownership is a ledger entry, not a feeling. Availability is a retention policy, not a promise.
Who actually pays the blob fee
Follow the payment. Blob fees are paid by the transaction sender — in practice, by sequencers. Sequencers are, for most large rollups, either the founding foundation, a small set of operators, or a foundation-controlled multisig. These are known addresses. They are traceable. Their balances are public.
This matters for two reasons.
The first is economic. When blob fees spike, the sequencer's margin compresses before user fees move, because sequencers absorb short-term cost to keep fees stable. That is a rational business decision — user fee volatility is a growth tax — but it means the blob fee is a charge against the treasury, not against the user. A treasury is a finite pool of tokens whose price is itself a function of market sentiment. The blob subsidy is therefore financed by the token, and the token is financed by the narrative that blobspace is cheap. The loop closes.
The second is regulatory. A foundation that pays blob fees is a foundation with a balance sheet and a signing key. The DAO wrapper does not change that. Projects preach decentralization while the wallet that pays the data bill is a multisig with a known signer set, and the compliance structure around that multisig is a shield rather than a fact. This is the pattern I have documented since the Tezos governance transition in 2017: theoretically sound, practically fragile, and the fragility lives in exactly the place the marketing points away from.
If you want to know how decentralized a rollup is, do not read its governance docs. Read its sequencer's fee-paying address and count the signers.
Complexity as camouflage
One more structural note, because it explains why the blob market is discussed so poorly.
The blob fee mechanism is genuinely complex for what it delivers. It has a separate gas type, a separate fee market, a separate accounting variable, an exponential update function, KZG commitments, a point-evaluation precompile, and a retention policy that interacts with the fraud-proof window of every rollup that depends on it. Each of those pieces is defensible in isolation. Together they create a surface area where the interactions are harder to reason about than the components.
I saw the same pattern in Uniswap V4's hooks — a design that turns the DEX into programmable Lego and, in doing so, raises the complexity floor high enough that a large fraction of developers will simply not build on it. Complexity is the camouflage for incompetence when the complexity is accidental. When the complexity is structural, it is camouflage for something else: it makes the system hard to critique because the critique requires understanding the whole stack.
I spent six weeks in 2017 reading Tezos' Coq formal verification proofs. The math held. The governance transition was theoretically sound and practically fragile, and the fragility was invisible to anyone who had not read the proofs. The blob market has the same shape. The math holds. The economics have an edge case, and the edge case is where the whole thing breaks.
In 2024 I submitted a report to EigenLayer's core team describing a vector in which the differentiation matrix could double-slash validators under specific network latency conditions. They acknowledged the theoretical risk and assessed it as low probability given current parameters. They were probably right about the probability. They were answering a different question than the one I asked, which was not "will this happen tomorrow" but "what does the specification permit, and under what conditions does the permission become a reality." If a vulnerability is theoretically possible, it will eventually be exploited. The only open variable is the timeline.
What the bulls got right
I am not going to pretend the bear case is complete, because it is not, and the honest version of this analysis has to concede three things.
First, Dencun delivered. Rollup fees fell by 10x to 100x for real users, in production, on schedule. That is not a narrative; it is a measured outcome. A mechanism that produces a real 10x cost reduction for end users is a success by any reasonable definition, and no amount of fee-market critique changes the user experience.
Second, demand elasticity is real. If blob fees rise, rollups will compress their data harder, batch more aggressively, and push more workloads onto validiums. The system has escape valves, and escape valves mean the catastrophic scenario is less likely than the naive model suggests.
Third, and this is the concession that cuts deepest against my own framing: a 100x increase from 1 wei is still 100 wei. The absolute magnitude of the pain is small. Even at 1 gwei — the level that requires 35 minutes of sustained maximum load to reach — six blobs cost about 0.000786 ETH per block, roughly $2.50 at an ETH price of $3,200. Across a day of full blocks, that is on the order of $18,000. Meaningful for a rollup's treasury, trivial per user. The people who feel the blob fee are sequencer operators and foundation treasuries, not the retail user whose transfer costs a fraction of a cent either way.
So the fee-doubling thesis, stated baldly, is technically true and practically underwhelming. The interesting claim is not that fees rise. It is that the rise is a tax on the entities that subsidize the user experience, and those entities have finite balance sheets, and the subsidy is what makes the current user experience possible. The bill is being paid by someone. The question is whether that someone has been told.
I have been wrong in this specific way before. In 2020 I found an edge case in Yearn's slippage tolerance, reported it to the core team, received credit for it — and then failed to apply the finding to my own portfolio, which took a 15% drawdown on exactly the slippage I had identified. Theoretical correctness is not practical advantage. The gap between them is where careful analysts lose money, and I have lived in that gap.
The metric to watch
Stop watching average rollup fees. They are a lagging indicator and they are dominated by execution cost, not data cost. Watch one number instead: the fraction of slots, on a rolling 30-day basis, in which the blob base fee exceeds 1 wei.
While that fraction is in the low single digits, the subsidy is intact and the "cheap and abundant" conflation is harmless. When it crosses 10%, the subsidy is ending, the marginal consumer is being priced out, and the migration to substitute DA layers begins in earnest. When it stays above 25% for a full quarter, the ceiling has been found and Ethereum's DA layer has become a commodity with a competitor's price.
None of this requires a prediction about price. It requires one public number, measured continuously, against a threshold defined in advance.
Here is the question the bull case has to answer and has not. If the cheapest resource in a system is priced at the protocol floor for years, is it abundant — or is it unpriced, and the invoice is being written to a party that has not yet been told it is liable? The fee market will answer eventually. It always does. The only thing still undetermined is who is holding the bill when it arrives.
