1. The Floor Test
In the second week of February 2026, the Ethereum blob base fee held above its 1-wei reserve price for 61 consecutive hours. For the twenty months prior, that number had been a rounding error โ a fee market that existed on paper, cleared at the protocol minimum, and functioned mainly as a marketing line item for every rollup growth team in the ecosystem. The February print was the first sustained departure from the floor.
I maintain a blob-series model that I rebuild after every network upgrade, and the output is unambiguous. Blob demand is no longer a function of rollup goodwill. It is a function of four independent consumer classes โ optimistic rollups, ZK rollups, non-rollup data publishers, and machine-witness attestation traffic โ and three of those four did not exist at scale when EIP-4844 shipped in March 2024.
This is not a price forecast, and I will not pretend it is one. It is a capacity audit. The audit says the cheap-data-availability era is a subsidy with a term sheet attached, not a permanent property of the settlement layer. Subsidies are useful. They are also, by construction, temporary. The question for anyone holding L2 exposure in this cycle is not whether the subsidy is good. It is what the fee curve looks like on the day it ends.
2. What Dencun Actually Changed
Before March 2024, rollup batch data competed for the same gas as every other transaction on Ethereum. It was calldata, priced by one market. The consequence was structural, not incidental: L2 fees and L1 congestion were positively correlated. When a liquidation cascade or a hyped mint hit mainnet, every rollup's data cost rose with it, and the rollup passed that cost to users within a block or two. Users experienced the phenomenon as L2 fees spiking for no reason. There was a reason. It was the design.
EIP-4844 severed the two markets. Rollup batch data moved into blob gas โ a parallel EIP-1559 controller with its own target, its own maximum per block, and its own base fee. That base fee is burned. It does not accrue to validators beyond a small tip. The intended result was a two-market system: the execution market prices DeFi activity, the blob market prices data availability, and neither sets the price of the other.
The release schedule since then has been a sequence of capacity steps, not a smooth curve.
- Dencun, March 2024. Three blobs per block as the target, six as the maximum. Blob base fee introduced with a 1-wei floor and an exponential excess-blob controller.
- Pectra, May 2025. Target and maximum both raised, doubling the ceiling and re-tuning the controller.
- Fusaka / PeerDAS. Data availability sampling replaces full download-and-store by every node. That is the enabling condition for the next step change in target capacity, because it moves the constraint from storage and bandwidth to probabilistic sampling.
Read that list again and notice what it is not. It is not a supply curve. There is no continuous response function. Blob capacity is a staircase, specified in an EIP, tested on a devnet, and activated at an epoch number. Demand for blobs is a ramp โ sometimes a steep one. That mismatch is the entire thesis of this piece, and everything below follows from it.
3. The Blob Saturation Curve
3.1 Four consumer classes, four elasticities
Aggregate blob demand is not one series. It is four, and they behave differently under price stress.
Class one: optimistic rollups. Base, Arbitrum, and the OP Stack family. Their blob consumption is close to a linear function of L2 transaction count divided by compression ratio, plus a security margin the sequencer sets for batch interval risk. This is the most predictable series in the set and the most commoditized. Migration from calldata to blobs is nearly complete, so growth here now approximates the growth rate of L2 gas consumption itself โ high, but bounded.
Class two: ZK rollups. Historically lower blob consumption per transaction, thanks to better compression from state-diff batching and validity proofs. Their footprint lands elsewhere: proof verification is real L1 execution gas, and it scales with batch frequency, not batch size. A ZK rollup that batches more aggressively to save blobs pays more in verification. There is a genuine optimization frontier here, and most teams are operating well inside it.
Class three: non-rollup publishers. Data availability layers posting commitments, oracle networks anchoring state, and โ the category most analysts underweight โ inscription-style data payloads. Inscription traffic is the shock absorber of the blob market. It is price-elastic in both directions. At near-zero blob fees, posting arbitrary data is free and the volume is enormous. At non-trivial blob fees, it evaporates within hours. Blob utilization is therefore a noisy indicator, and headline utilization numbers overstate structural demand by a wide margin. Separate the elastic classes from the inelastic ones, or you are reading weather as climate.
Class four: machine-witness attestation. This is the new series, and it is the reason I rebuilt the model rather than updating it. As autonomous agent transactions get settled onchain, someone has to attest that a given inference ran on a given model with a given input. Zero-knowledge proof-of-origin schemes produce compact proofs, but a proof still needs somewhere to live, and the commitment structure โ model hash, input digest, output digest, timestamp โ is a data payload. I spent a portion of last year optimizing the computational cost of exactly these proofs, and the constant I kept running into was not proving time. It was publication cost. Attestation traffic is inelastic in a way inscriptions are not: an agent's settlement obligation does not disappear when blob fees triple. It gets passed on, delayed, or batched harder โ but it does not vanish.
3.2 The crossover
My base model rests on three stated assumptions. I am listing them because a model without visible assumptions is an opinion wearing a lab coat.
- Blob target capacity follows the published staircase, with activation dates taken from the EIP schedule rather than from optimistic devnet chatter.
- Space is consumed first by inelastic demand โ rollups and attestations โ then by elastic demand filling whatever remains.
- Rollup blob demand grows with L2 gas consumption at a mild deceleration, reflecting compression improvements roughly offsetting activity growth.
Under those assumptions, the blob base fee sits pinned at 1 wei while utilization stays below target. Then it does not rise smoothly. It rises as a step.
The reason is the controller. The blob fee market uses the same shape as the execution market: below target, the fee decays toward the floor; above target, the fee increases exponentially in the size of the excess. There is no gentle middle. The market transitions from free to price-discovery inside a narrow utilization band, and it does so faster than the smoothing constants on most institutional dashboards can represent. Every rollup fee schedule published to users is calibrated to a regime defined by the floor price. When the floor breaks, the passthrough is mechanical, and it happens on the timescale of a batch interval.
3.3 The passthrough ledger
A rollup's fee to the end user decomposes into four components. Write them down separately, because the marketing does not.
- Sequencer execution cost. Off-chain hardware and bandwidth, amortized. Small, roughly constant, and falling with better software.
- Blob cost for data availability. Post-Dencun, this is the line item everyone watches. It has been approximately zero.
- L1 execution gas for proof verification and state root submission. This has not been approximately zero. For a well-designed rollup it has been the dominant term in the cost stack for over a year.
- Sequencer margin. The difference between what the user pays and items one through three.
Item three is the detail that gets lost in every all-time-low fee headline. The reason optimistic rollups advertise sub-cent transfers is not that data availability became free โ it is that they batch aggressively enough to amortize a fixed verification cost across thousands of transactions. That is an engineering achievement, and it is nearly exhausted. You cannot amortize a constant to zero.
3.4 Why cheapest is not a moat
L2s have competed on a single ordinal metric for two years: cheapest. That is a deflationary race on one cost line, and the line they were racing on was being subsidized by a component none of them control. A rollup cannot control blob demand from inscriptions, from DA layers, or from agent attestation. It can only decide how much of the blob cost to absorb versus pass through.
Absorption is a balance-sheet decision, and it is procyclical. In a bull market, sequencer revenue is high and rollups absorb the cost to protect the growth narrative. In a down market, sequencer revenue falls, absorption stops, and the fee schedule resets to reflect the true cost stack. The subsidy is therefore correlated with the same cycle that makes users care about fee levels in the first place. That arrangement is unstable, and it is invisible on a chart of average L2 transaction fees, because the chart shows the subsidy, not the cost.
3.5 The condition and the trigger
I will not give a date. Dates are how analysts get credit for being early and blame for being wrong. I will give a condition and a trigger, both observable.
The condition: sustained blob utilization above target for more than 72 hours, observed in two separate weeks, with the elastic publisher class excluded from the measurement.
The trigger: a capacity increase announced but not yet activated. When both are true, the market has entered the regime where the effective fee ceiling is set by a committee calendar, and the marginal data-availability buyer is bidding against everyone else's batched data with no new supply arriving until an epoch number passes.
My expectation is that the floor test becomes routine before the next capacity step is scheduled, not after. Staircases and ramps cross at the top of the stair, not the bottom. And the ramp has a new passenger โ class four โ that was not in anyone's 2024 model.
4. The Contrarian Angle: The Decoupling Thesis
The consensus reading is that blob fee normalization is bullish for L2 tokens, because it means real demand. Follow the rent and it falls apart.
Blob fees are burned. They do not accrue to L2 token holders. Sequencer revenue is captured by the sequencer operator, which in most rollups today is a single entity. Value accrual to the L2 token, where it exists at all, comes from governance control over a fee switch that no major rollup has fully flipped, or from staking and sequencing decentralization that most have not shipped.
So the transmission chain from rising blob demand to L2 token appreciation has two broken links. The data-availability fee leaves the system entirely, burned at layer one. The remaining user fee is retained at the sequencer. The token is a claim on a future governance decision, not on present cash flow. That is not a criticism of any specific team; it is a description of the architecture as deployed.
There is a second decoupling, and it is regulatory rather than technical. Jurisdictional competition for licensed intermediaries has become explicit industrial policy across Asia. Hong Kong's licensed platform regime has moved faster than Singapore's on retail access; Singapore's has moved faster on institutional custody and fund structuring. Both are optimizing for the flow of licensed intermediation, because that is where fee density concentrates. The effect on an L2 is distribution, not utility: a token listed on a licensed venue reaches a new buyer base, but the venue's licensing status does nothing to the protocol's cost structure. Conflating the two shows up in valuation models as a permanent multiple expansion that the cash flow line never confirms.
And there is a third decoupling, quieter, that deserves a direct comparison. Blob pricing looks like a market. It has a target, a maximum, and a controller constant, all chosen by people. Anyone who has read the interest rate curves on Aave or Compound will recognize the shape: those APRs are not discovered by supply and demand, they are drawn by governance and then described as if they were markets. The data-availability layer has the identical property. Calling it market pricing for data availability is a convention, not a description. Once you accept that, the question stops being what the market will do and becomes what the people who set the parameters will do, and when. That is a far more tractable question.
5. Takeaway
Exit strategies are written in ice, not in hope.
The practical version for this cycle: treat every current rollup fee schedule as a promotional rate with a disclosed but unmodelled expiry. Model L2 exposure on the cost stack โ verification gas plus blob cost at a non-floor price โ rather than on the advertised number, because only one of those two is stable. Watch two series, not one: inelastic blob consumption with publishers stripped out, and the gap between announced capacity and activated capacity.
The question I would leave with anyone allocating into this cycle is not whether rollups will be cheap. It is who is still standing when being cheap stops being a choice they get to make.