At 03:41 UTC on a Tuesday in February, the blob base fee on Ethereum printed eight wei.
That is nothing. Eight wei is roughly one hundred-billionth of a dollar. You cannot buy a rounding error with it. But I had a single monitor that week dedicated to one number, and that number was excess_blob_gas, and it was climbing. For forty-one minutes it sat above target while a sequencer batch-poster pushed thirty-eight blobs into a nine-blob ceiling. Then it snapped back, the fee returned to its floor of 1 wei, and the network went back to sleep.
Everyone went back to sleep except the four rollup teams whose entire published unit economics assume that floor never moves. I spent the next three weeks chasing that forty-one-minute window โ pulling blob-level data out of an archive node, rebuilding the fee curve by hand, and talking to two sequencer engineers who asked not to be named because their employers still market "sub-cent fees" as a permanent feature rather than a promotional one.
Here's what the math says. Chasing the alpha through the noise is usually a bad use of a Tuesday. This time it wasn't.
Context
The floor everyone is standing on was poured in March 2024. Dencun shipped EIP-4844, and Ethereum stopped making rollups pay for data availability through calldata. Instead of dumping compressed batch data into the execution gas market, rollups started posting blobs โ 128 kilobytes of data attached to a block, priced in their own fee market, verifiable by consensus but not executable by the EVM.
The design was elegant. Blobs got their own EIP-1559-style auction, decoupled from execution gas. Target three blobs per block, maximum six. A blob base fee with a floor of 1 wei. Rollup costs fell off a cliff the week Dencun activated, and every Layer 2 roadmap for the next eighteen months was written on the assumption that data availability would stay free or close to it.
Then Ethereum kept raising the ceiling without ever lowering the floor. Pectra, in May 2025, moved the target to six blobs per block and the maximum to nine. Fusaka, in late 2025, shipped PeerDAS โ data availability sampling โ and pushed the target to fourteen and the max to twenty-one. More room. Same 1-wei floor. From the peak of the 2021 fee wars to the pit of the 2022 bear, the industry learned one lesson: costs that look structural are usually just supply racing demand.
What almost nobody changed was the mechanism. The blob fee market has no gradual ramp. It is not a slope. It is a step function wearing a curve's clothing, and the step is closer than the supply increase suggests.
Core
Blob pricing runs on two variables. excess_blob_gas accumulates whenever a block consumes more than the target and drains whenever it consumes less. The blob base fee is then computed as a fake exponential โ fake_exponential(1 wei, excess_blob_gas, 3,338,477) โ which is Solidity-speak for "this number barely moves, then it moves all at once."
I rebuilt that curve by hand in a spreadsheet, the way I rebuild every mechanism I don't fully trust, and the result is not a gentle upward line. At an excess of 20 million blob gas, the blob fee is around 400 wei. At 60 million, it's roughly 65 million wei. At 100 million, you're into the tens of billions. The blob fee spends most of its life pinned at the floor and compresses most of its upside into a window of hours. There is no "somewhat expensive." There is free, and then there is a wall.
For fourteen months the wall never got hit. Three structural reasons.
First, supply expanded faster than demand. Pectra and Fusaka together took the per-block target from three blobs to fourteen โ a 4.6x increase in available DA capacity in under two years. Each increase reset the congestion clock.
Second, demand was concentrated and polite. The blob market's biggest consumers are a handful of rollups โ Base, Arbitrum, OP Mainnet, and the zk cohort โ plus a thin tail of DA-layer settlement traffic. Rollup sequencers batch on fixed intervals. They don't respond to a fee spike by posting more; they respond by posting less, or by waiting a few blocks. The largest buyers in the blob market are also the most price-elastic, which is the opposite of what a fee market needs to stay healthy. A market where the biggest participants can simply stop participating is a market that clears at the floor forever.
Third, most rollups were never blob-constrained. They're constrained by L1 execution gas for verification, by L2 execution capacity, and by the cost of running the sequencer and prover fleet. Blobs were the line item that made the pitch deck look good. They were rarely the bottleneck.
The real cost stack has three layers, and only one of them is getting cheaper.
When a rollup posts a batch, it pays for at least three things: blob space for the compressed state diff, execution gas for the verification transaction โ a validity proof verification, a fault-proof bond, or a calldata commit โ and the fixed operational cost of the sequencer and prover fleet.
Post-Dencun, the blob line went to near zero and stayed there. The verification line did not. A Groth16 verification on Ethereum mainnet cost roughly 250,000 to 300,000 gas in 2024, and it costs roughly the same in 2026. That number has barely budged in three years, and it's per batch, not per byte. The cheaper your DA gets, the larger the share of your L1 bill goes to a fixed cost you cannot compress.
I watched this play out in a way that should worry anyone still modeling rollups as "cheap DA plus cheap execution." One mid-sized zk rollup I track โ call it Rollup A, since the team asked me not to name them โ cut its blob spend 71% between Q3 2025 and Q1 2026 by improving compression and moving cold state to a competing DA layer. Its total L1 bill fell 19%. The other 52 percentage points vanished into verifier gas and prover machine hours.
Rollups have been selling a data-availability story while running an execution-and-proving business. The story was true for about eighteen months. It is increasingly not.
Blobs expire in eighteen days, and that redefines what "DA" even means.
Here's the fact that gets left out of every "blobs are cheap" thread: blob data is pruned after 4,096 epochs โ roughly eighteen days. A node is not required to retain it. After that window closes, the data exists only if someone else kept a copy.
I spent a week in January trying to sync a rollup node from scratch using only L1 blob data, as an experiment. I couldn't. The historical blobs I needed were gone, and the archival providers I queried had gaps in exactly the blocks where congestion had been highest โ because those were also the blocks where indexers had dropped data under load.
This is not a bug in EIP-4844. It's the design. Blobs were built as a settlement layer, not a storage layer, and the 18-day window was a deliberate trade to keep node hardware requirements flat. But it means that when a rollup says "your data is on Ethereum," what it actually means is "your data was on Ethereum for eighteen days and is now on our servers and whoever else bothered to keep it."
That is a trust assumption, and it is exactly the trust assumption rollups were supposed to eliminate. Breaking silos, one block at a time, was the promise. The silo, it turns out, just moved into an S3 bucket.
The competing DA layers know this, and it's why Celestia and EigenDA sales decks over the past year have led with retention guarantees rather than price. It's also why the "Ethereum is the DA layer" argument has gotten quieter in engineering circles even as it's gotten louder on social media.
The new buyer does not care what blobs cost.
Through most of 2025 I ran an AI-agent trading bot โ badly, on purpose, documenting the whole mess in a series I called Chaos Cooking. The bot was dumb, but it was prolific. It generated somewhere between 40,000 and 90,000 on-chain actions a day in simulation, and the only thing that stopped it was running out of testnet funds.
I tell that story because it is the shape of what's coming. Nearly every blob-demand model I've read assumes the buyer is a human-driven rollup with a batch interval and a cost committee. That buyer is price-elastic. It throttles. The next wave of consumers โ autonomous agents settling microtransactions, oracle networks pushing signed attestations, inference markets logging results on-chain โ has no throttle. An agent paid per action does not stop because the blob fee went from 1 wei to 10 million wei. It pays, or it stops existing, and the ones that stop existing get replaced by ones with better economics.
Price-inelastic demand meeting a step-function fee curve is how you get vertical spikes. Unlike the 2021 fee spikes, which were driven by humans who eventually got bored, this demand doesn't get bored.
And here's the part that surprises people: the RWA wave that was supposed to be the blob market's savior hasn't arrived, and probably won't. The institutional tokenization platforms โ the ones with real balance sheets โ run permissioned environments with their own settlement rails. They were never going to post custody ledgers to public blobs, because they don't need your chain to do what a private database does better. Stablecoin payment flows have the same problem. The volume is real, but the biggest issuers are structuring it as regulatory partnerships rather than open public-blockchain traffic. If you're waiting for traditional finance to fill the blob pool, you're waiting for something with no incentive to show up.
Contrarian
Here's the angle I haven't seen anyone take seriously.
The consensus worry is that blob saturation makes rollups expensive. True, but it's the wrong worry, because rollups can eventually pass cost down. The subtler risk is that a step-function fee curve makes rollup pricing unforecastable, and you cannot build a subscription business on an unforecastable input.
Think about what a step function does to a fixed-fee product. A rollup that wants to sell a predictable monthly rate to a DeFi protocol has to reserve against the worst case โ blobs at a thousand times the floor for three days. Either it charges the worst case and loses every price-sensitive customer, or it doesn't and eats a catastrophic quarter. There is no middle path, because there is no middle price. And there's no hedging market to bridge the gap. The only blob-fee derivatives I could find are thin, retail-facing, and listed on venues with less depth than the underlying.
Second blind spot: everyone watches the blob fee, and nobody watches the blob count. The fee is a lagging derivative of the thing that actually matters. excess_blob_gas is the leading indicator, and any archive node will hand it to you in real time. My 41-minute window never pushed the fee above 8 wei, but it pushed the excess above target for the first sustained stretch since Fusaka. If you're building on an L2 and want to know what your costs look like in six months, the blob fee chart is close to useless. The excess chart isn't.
Third: the Fusaka supply increase may have bought less time than it appears. Going from a target of six to fourteen blobs reads like a 2.3x buffer. But agent-driven demand doesn't scale linearly with human activity. It scales with the number of deployed agents, and that number compounds. My half-broken bot pushed 90,000 actions a day. Ask what a thousand production agents do to a fourteen-blob target.
Takeaway
Watch two numbers this quarter, and neither of them is price.
Watch excess_blob_gas on Ethereum mainnet over a rolling seven-day window. A sustained drift above zero is the tell, and it will show up in the data weeks before it shows up in anyone's fee announcement.
And watch zk verifier gas on your favorite rollup's settlement contract. That's the line item blobs were supposed to make irrelevant and didn't.
The 1-wei era was real, and it lasted longer than anyone expected. It was never permanent. It was a supply expansion racing a demand curve that hadn't arrived โ and that demand curve is now being written by software that doesn't know how to wait.
So the question isn't whether rollup fees go up. It's whether the teams still selling sub-cent transactions have a plan for the week their DA bill multiplies by a thousand โ and whether anyone in this ecosystem has built the instrument that would let them survive it. From where I'm sitting, the answer to both is still no.
