On June 4, the Ethereum blob base fee printed its floor value — 1 wei — for the sixty-third consecutive day. At that price, the entire daily data-availability budget of the network, sustained at target utilization of six blobs per 12-second slot, costs roughly $0.0000142. Not per gigabyte. Not per megabyte. In total. Every blob posted that day — 43,200 of them at target, 128 kilobytes each, 5.7 gigabytes of data — cost less than two ten-thousandths of a cent combined.
The anomaly is not that data is cheap. The anomaly is that the scarcest resource in the modular stack has no price. Ethereum spends block space, validator bandwidth, and cryptographic verification on KZG commitments for rollup data, and the market clears at the minimum representable unit of account. Something that costs nothing to consume tells you nothing about how much of it you should consume. That is a failure of price discovery, not a gift.
I have been running a blob-fee monitor since the week EIP-4844 went live on March 13, 2024. It is a descendant of a Python script I wrote during DeFi Summer 2020 to track oracle lag between Uniswap and SushiSwap — same discipline, different data source. The output of that monitor over the last 90 days is what follows.
Context: what the blob market measures, and what it does not
Blobs are the data structure introduced by EIP-4844. Each carries 131,072 blob gas — a fixed unit independent of execution gas — and is pruned from consensus after 4,096 epochs, approximately 18.2 days. Rollups post compressed batch data as blobs and rely on the availability guarantee for that window. The fee for blob space runs in its own market, separate from the execution fee market, with its own base fee, its own excess variable, and its own update rule.
That update rule is the whole story, so it deserves precision. The protocol tracks a variable called excess_blob_gas. If the parent block's blob usage plus the accumulated excess is above target, excess increases by the difference. If usage is at or below target, excess resets to zero — immediately, in one block. The blob base fee is then fake_exponential(MIN_BLOB_BASE_FEE, excess_blob_gas, BLOB_BASE_FEE_UPDATE_FRACTION), with the floor set at 1 wei and the update fraction at 3,338,477.
Read that again. Congestion memory in the blob market is asymmetric. It compounds upward exponentially, and it forgets downward instantly. Scarcity here is an algorithm, not a belief system — and the algorithm has no hysteresis in the direction that matters for planning.
EIP-7691, shipped in Pectra on May 7, 2025, raised the target from three blobs per block to six, and the maximum from six to nine. Capacity at target went from 21,600 blobs per day (2.83 GB) to 43,200 (5.66 GB). At the maximum it is 8.49 GB per day. The received wisdom is that this doubled the runway. The number that matters is not capacity. It is the ratio of sustained utilization to target, because that ratio is the only input that moves the excess variable, and the excess variable is the only thing that moves the fee.
My methodology here is deliberately narrow. I use three sources: blob-inclusion telemetry from my own execution client, cross-checked against public dashboards that aggregate blob count and base fee per slot, plus sequencer batch-posting transactions decoded by rollup and by poster address. I am not interested in aggregate transaction counts. I am interested in who is buying blob space, how much they are buying relative to target, and what they do the moment the price moves.
The evidence chain: a knife-edge disguised as a floor
The first finding is that utilization and price have decoupled almost entirely within the current capacity envelope. Post-Pectra, seven-day blob utilization has oscillated between roughly 30% and 65% of the new target. In that entire range, the base fee sits at 1 wei. This is not a bug. It is exactly what the update rule produces: as long as aggregate usage stays below the six-blob target, excess never accumulates, and the fee never leaves the floor. A single large poster can consume 90% of target indefinitely and pay nothing for the privilege.
The second finding is the shape of the regime change when utilization does exceed target. At the nine-blob maximum, excess grows by 393,216 blob gas per block, and the per-block fee multiplier is e^(393,216 / 3,338,477), or approximately 1.125. Twelve and a half percent per block. Compounded every twelve seconds.
Run the arithmetic forward from the floor. From 1 wei, the blob base fee crosses 100,000 wei in under twenty minutes of sustained maximum utilization. It crosses one billion wei in roughly thirty-five minutes. It crosses ten billion wei — the point at which a rollup posting 300 transactions per blob re-crosses one cent of data-availability cost per transaction — in about thirty-nine minutes.
That is the finding. The distance between free data and expensive data on Ethereum is under forty minutes of sustained congestion. Not a quarter. Not a halving cycle. Forty minutes. There is no gradual ramp, no early warning in the price series, no opportunity to renegotiate batch sizes over weeks. The fee market is a step function wearing the costume of a curve.
The third finding concerns who absorbs the shock. I decoded sequencer batch-posting transactions across the major rollups over the sample window and grouped them by compression efficiency — bytes of blob data per transaction settled. The distribution is wide, and it is not correlated with the size of the chain. Some mid-cap rollups are posting at roughly 90 to 140 bytes per transaction. Others, with larger share of activity, are running at 400 bytes and above, which suggests they are not applying dictionary compression, state-diff batching, or calldata-to-blob re-encoding aggressively. Why would they? At 1 wei there is no economic pressure to improve compression. The floor price has removed the incentive to invest in the one engineering discipline that determines who survives the step function. The alpha isn't in the headline; it's in the silenced code — specifically, in the compressor that nobody bothered to optimize because the invoice read zero.
During my 2017 due-diligence work on pre-sale token distributions, I learned that the cost of a design flaw is not paid when the flaw is written. It is paid at the moment of maximum load, when the fix is most expensive. The reentrancy bug I found in one distribution contract had been dormant for months. It was only dangerous on the day the contract was busy. Blob compression ratios are that bug, at protocol scale.
The fourth piece of evidence is behavioral, and it is the one that makes the rest actionable. During the four blob-fee spikes in my sample window — each lasting between forty minutes and three hours — batch-posting behavior changed in a measurable, repeatable way. Posters with alternative data-availability paths switched within two to six blocks. Posters without them throttled batch submission and let the sequencer queue absorb the delay. Blob counts fell back below target, excess reset to zero in a single block, and the fee returned to 1 wei within roughly ninety minutes of the spike peak.
This is a system that resets perfectly and learns nothing. Each cycle leaves the participants exactly as they were: the flexible ones re-optimized, the inflexible ones unchanged, and the aggregate compression ratio of the network flat. Which brings me to the part of this analysis that most observers get backwards.
The contrarian read: cheap data did not cause the adoption
The dominant narrative since Dencun is a clean causal chain: cheaper data availability produced cheaper rollup fees, which produced more activity, which validated the modular thesis. The correlation is real. The causation is inverted, or at least underwritten by a transfer that nobody is pricing.
Look at the order of operations. Average rollup fees fell by an order of magnitude within days of the 2024 hard fork, before any measurable change in application-level demand. Then, when Pectra doubled blob capacity again in May 2025, several rollups barely moved their user-facing fees at all. If data cost were the binding constraint on adoption, a doubling of the supply of the cheapest input should have produced a second step down. It did not, because for most rollups the marginal cost of a transaction was never the constraint. Bridge friction, fragmented liquidity, and app-level utility were. The fee cut was absorbed as margin, not passed through as growth.
Which means the current state is not a productivity gain. It is a transfer. Every blob posted at 1 wei is a burn that did not happen, an issuance cost borne by the settlement layer, and a subsidy delivered to sequencer margins. That transfer is invisible in the metric everyone watches — blob count — and fully visible in the metric almost nobody watches: data-availability cost as a share of sequencer revenue.
Correlations are the lie; liquidity is the truth. The relevant question is not how many blobs are being posted. It is how much of the fee floor is structural, and how quickly the structure flips.
There is a second blind spot, and it is worse. Every dashboard tracks blob count against target. Almost none tracks the variance of blob count within the sample. A seven-day utilization of 50% sounds comfortable. A utilization that alternates between 15% and 95% on a six-hour cycle is a different object entirely — it is a system whose fee will be repriced violently the first time the peaks stop being offset by the troughs. Distribution, not mean. I learned that lesson the expensive way in May 2022, when the on-chain flows out of Anchor were unremarkable in aggregate and catastrophic in sequence. Averages are how you miss a bank run.
The signal to watch, and the position it implies
In a tape like this one — sideways, directionless, where chop is the only reliable feature — the useful move is not to forecast the level. It is to identify which assets have a cost structure that survives the step function. Two numbers tell you almost everything. First, the seven-day moving average of the blob base fee: when it departs from 1 wei and holds above the floor for five consecutive sessions, the regime has changed and the market has not yet repriced it. Second, the share of a rollup's sequencer revenue consumed by data-availability cost, computed at a hypothetical fee of ten billion wei rather than today's. If that stress number exceeds low single digits, the chain is running a leverage it does not disclose.
Watch the compression ratios, not the transaction counts. Watch the posters, not the TVL. Due diligence is the only hedge against chaos, and the ledger remembers what the marketing forgets — including which rollups were engineered for a world where data was free, and which ones quietly built for the world that is coming.