For most of the past quarter, the Ethereum blob base fee has printed its floor of 1 wei. Whole weeks pass without a single block clearing the minimum. Rollup operators read that number and concluded that data availability is free. They then rebuilt their cost models, their sequencer budgets, and their token emissions around it.
They are reading the wrong number.
The floor is not evidence of abundance. It is an artifact of an asymmetric curve. Blob pricing moves only when demand crosses a network-defined target, and the network keeps moving that target upward. Below the target, price sits at its floor. Above it, price compounds exponentially, with no soft shoulder in between. A fee market with a floor and a cliff is not a fee market. It is a switch.
I have spent four years auditing rollup economics out of Berlin. The same omission appears in nearly every model I open: the data-availability line item is written as a constant. Collateral is a lie; math is the only truth.
The mechanism deserves more precision than it gets. EIP-4844, activated in March 2024, gave Ethereum a second, independent fee market for blobs — 131,072 bytes of data each — priced by their own base fee rather than by the gas market that settles execution. Blob fees follow an EIP-1559-style rule with two critical differences: the minimum base fee is 1 wei, and the exponential update fraction is fixed at 3,338,477.
When Dencun landed, rollup fees collapsed by roughly an order of magnitude within weeks. That collapse was real. It also produced a narrative that data availability had been solved — a narrative that mistook a price floor for a permanent state.
Before Dencun, rollups posted state diffs as calldata, competing directly with ordinary transactions for the same blockspace. A busy rollup could spend seven figures a year on that alone, and the cost was visible to every user as a line item. Blobs decoupled the two markets. That decoupling is the design's greatest strength and the source of the blindness: once data availability stopped competing with execution for the same gas, it also stopped appearing on the dashboards operators watch.
In May 2025, Pectra raised the blob target from three per block to six, with a maximum of nine. Headroom doubled overnight. The floor held, and the narrative hardened. Fusaka and its PeerDAS work promise another multiple on top. Every one of those upgrades is a genuine engineering success. None of them changes the shape of the curve.
Here is the arithmetic. The blob base fee is computed as fake_exponential(1, excess_blob_gas, 3338477), where excess_blob_gas accumulates the gap between blobs actually used and the per-block target. Each blob posted above target multiplies the base fee by roughly e^(131072/3338477), or 1.0400.
Four percent per blob. That sounds gentle until it compounds. A chain sustained at one blob above target doubles the blob base fee in about eighteen blocks — under four minutes. Sustained at full capacity, three blobs above target, it doubles roughly every six blocks, a little over a minute. From the floor to a materially expensive fee takes a couple of hours at most. It does not take a quarter.
There is no tip mechanism on the blob portion of the fee. Blob gas is base-fee-only; there is no mempool auction in which a large rollup can outbid a small one for inclusion. The only levers are volume and timing: post fewer blobs, or post them later. Both are cost decisions with second-order consequences.
The exit is not instant either. When demand falls, excess_blob_gas drains at the target rate, so a fee that took two hundred blocks to build takes a comparable number to unwind — assuming demand actually returns to target. It usually does not, because the applications that generated the load do not stop running, and their demand is reflexive. Cheap data does not satisfy appetite. It expands it.
The second-order consequence is where most rollups are exposed. Data availability is not a smooth cost; it is a packing problem with a hard edge. A blob is atomic — 131,072 bytes, billed whether you fill it or not. A rollup packing 60 percent of each blob is paying a 66 percent premium on its effective DA cost. At floor pricing, nobody notices. At spike pricing, that inefficiency is the difference between a real margin and a subsidized one.
Then there is the sequencing decision. When blob fees climb, sequencers either pay or defer posting. Deferring is rational in isolation and dangerous in aggregate. For optimistic rollups, delayed batch submission stretches the interval between user transaction and L1-finalized state, which feeds directly into challenge-period assumptions and bridge withdrawal times. For ZK rollups, proof aggregation pipelines are usually tuned to a fixed submission cadence; when that cadence breaks, proving costs and DA costs interact in ways most teams have never simulated. In my own audit work — including a four-month review of a ZK-rollup aggregation layer for a Berlin venture studio — I found compression inefficiencies that looked trivial at floor prices and catastrophic under load. My recommendation delayed their mainnet by three weeks. The team was unhappy. The alternative was a network that congested precisely when congestion was most expensive. I do not trust; I verify the hash.
And the demand side is not static. Every cheap-DA application built since 2024 — high-frequency social, onchain order books, AI-agent logging — was constructed on the premise that posting to L1 costs nothing. Those applications do not have elastic demand. They have reflexive demand: the cheaper the DA, the more of it they generate. The curve will not be tested by steady growth. It will be tested by a step function.
Worth naming who sets the target. The blob count is not a governance vote. There is no onchain proposal, no token-weighted ballot, no quorum, no turnout. It is an EIP, decided by client teams and researchers. A design fact that deserves more scrutiny than it receives from the same DAOs that publish manifestos about decentralization.
The metric worth tracking is not the average blob fee. It is the share of blocks in which used blob gas exceeds target. That share is the only variable that predicts a repricing, and it is currently near zero — which is exactly what makes it informative. A market that has never been stress-tested has no observable elasticity, no realized volatility, and no hedged participants. Absence of stress is not evidence of resilience; it is absence of evidence.
The bulls are not wrong about everything, and pretending otherwise would be lazy. They are right that DA has become a genuine commodity: Celestia, EigenDA, and Avail create real substitution pressure, and that competition does cap the ceiling on any single spike. They are right that capacity expansions are engineering rather than marketing, and that PeerDAS is a real multiple rather than a slide. They are right that the 2024 pivot worked, and that the applications cheap DA enabled are legitimate.
What they have not done is hedge the convexity. Competition caps the price of alternatives; it does not cap the cost of migration. A rollup mid-flight on Ethereum blobs cannot switch DA layers in a week — the proving system, the bridge contracts, and the data-retrieval assumptions are all coupled to the blob commitment scheme. Substitution is real over eighteen months and nearly absent over eighteen hours. The window in which you need the alternative is precisely the window in which you cannot use it. Between the lines of bytecode lies the trap.
The forward question for 2026 is not whether blobs are cheap today. It is which rollups have modeled the cost of the day they are not — and which ones have simply priced a floor as if it were a law. The curve is not cruel. It is indifferent. It will reprice the moment demand crosses a line that a small group of engineers drew, and it will not announce the crossing. The proof is complete; the doubt is obsolete.
