Over the last 90 days, the median blob count per Ethereum block moved from 2.1 to 4.7. Over the same window, the blob base fee sat at its 1 wei floor for 93.6% of slots.
Both numbers are correct. Both come from the same chain. Together they describe a market that is not clearing — it is being rationed by an artificial floor while every quarterly deck published this year calls the resource "abundant."
The alpha is in the silenced code. Not in the unlock schedule. Not in the points program. Not in the foundation post about "the modular future." It is in a fee market that was welded into the consensus layer in March 2024 and that almost nobody has plotted on a chart since.
Here is the anomaly in one line: blob consumption nearly doubled while the price of blobs went to zero and stayed there. A resource whose utilization rises 124% and whose price falls to the protocol minimum is not a healthy market. It is a market where the price signal has been disconnected from the quantity signal, and where every business model built on top of it is therefore priced by assumption rather than by data.
Context: what a blob actually is, and why its price is not a price
EIP-4844 shipped with Dencun on March 13, 2024. It introduced a second, parallel fee market to Ethereum: blob gas. Rollups post their transaction batches to L1 as blobs — 4,096 field elements of 32 bytes each, roughly 127 KB of usable payload after encoding — rather than as calldata inside execution blocks. The design intent was clean. Separate the data availability cost from execution cost so that a rollup batch does not compete with a Uniswap swap for block space.
The mechanism is worth over-explaining, because the misunderstanding is now load-bearing across the entire L2 sector.
Blob gas has its own base fee, updated per block by an exponential function. Unlike EIP-1559's execution base fee, which adjusts linearly within a bounded range, the blob base fee moves as fake_exponential(MIN_BLOB_GASPRICE, excess_blob_gas, BLOB_GASPRICE_UPDATE_FRACTION). The floor is 1 wei. The update fraction is 3,338,477. Excess blob gas accumulates when a block exceeds the target blob count and decays when it falls below.
The consequence is a fee curve that is flat, flat, flat — and then vertical. At target, the base fee is unchanged. One blob above target per block, and the base fee rises roughly 12.5% per block. Six blocks of sustained overshoot and you are at 2x. Twenty blocks and you are at 8x. The curve has no memory of how long you have been away; it only cares about the accumulated excess.
Post-Pectra — activated May 7, 2025 via EIP-7691 — the target moved from 3 blobs to 6, and the maximum from 6 to 9. That is 1.18 MB of blob data per 12-second slot at the cap, roughly 98 KB/s, roughly 8.5 GB/day. At target, about 65 KB/s.
That is the whole apparatus. Two numbers, one exponential, one floor. And a governance process that adjusts the numbers every eight to fourteen months.
Now plot utilization against price and the anomaly becomes structural rather than statistical.
Core: the two-sided miss
The blob market is not over-supplied or under-supplied. It is both, alternately, and the alternation is the point.
My monitoring stack runs a simple three-variable capture per slot: blob count, blob base fee, and the ratio of the two against the post-Pectra target of 6. Aggregated over the 90 days to mid-February 2026, the distribution looks like this:
| Metric | Value | |---|---| | Median blobs per block | 4.7 | | 90th percentile blobs per block | 8.2 | | 99th percentile blobs per block | 9.0 (cap) | | Share of slots at 1 wei floor | 93.6% | | Share of slots above target (>6 blobs) | 21.4% | | Median blob base fee when above target | 41 wei | | Peak blob base fee in window | 3,847 wei |
Read the third and fifth rows together. More than a fifth of all slots exceeded target. That should be pushing the base fee up persistently. Instead, the base fee spent 93.6% of the window at the floor, because the exponential decays as fast as it climbs — and the blocks that overshoot are interspersed with blocks that run two or three blobs, bleeding excess back out.
This is the two-sided miss. In the 93.6% of time at the floor, blob space is free at the margin, which means rollups have no incentive to economize, which means demand is structurally inflated relative to what it would be at any positive price. In the <1% of time where the base fee prints above 1,000 wei, DA cost for a rollup posting every block rises by two to three orders of magnitude within an hour, with no hedging instrument available and no forward curve.
A market with a floor price and a step-function ceiling is not a market. It is a rationing scheme with a price display attached.
Scarcity is an algorithm, not a belief system. Right now the algorithm says "abundant" 93.6% of the time and "catastrophic" the rest, and no participant can plan against either state.
Where L2 margins actually come from
Here is the part that gets buried under "L2 fees are down 90%."
An L2's cost structure has three components: L1 data cost (blobs), L1 execution cost (the verification/commitment transactions — proof verification for ZK rollups, batch submission calldata for optimistic ones), and the L2's own execution cost (proving, sequencing, infrastructure). The first two are paid in ETH and are denominated in L1 gas. The third is largely fixed opex.
When blobs sit at 1 wei, L1 data cost approaches zero. The average rollup's DA spend as a share of total L1 cost falls under 8%. That means the headline fee compression users experienced through 2024 and 2025 was not primarily an efficiency gain from better proving systems or denser calldata encoding. It was a DA subsidy paid by Ethereum's own resource underpricing.
I flagged the same accounting illusion in 2017 during the ICO due diligence cycle. Fifteen pre-sales, whitepapers full of "scalable" and "low-cost," and in eleven of them the unit economics only worked because a cost line was assumed to be zero rather than measured. One of them had a reentrancy flaw in the token distribution contract that would have let the first caller drain the vesting schedule; the launch slipped six weeks and the team never disclosed why. The pattern is identical here. A cost that is currently zero is being modeled as permanently zero, and every fee projection downstream inherits the error.
Run the sensitivity. Take a representative optimistic rollup posting 6 blobs per block, 7,200 blocks per day. At the floor, DA cost is effectively nil. At a sustained blob base fee of 400 wei — which is not a tail event, it is roughly the 75th percentile of my above-target distribution — that same rollup pays about 0.0028 ETH per day in blob gas. At 3,800 wei, the 99.9th percentile in my window, it pays about 0.027 ETH per day. Multiply by a year and 400 wei sustained: about 1.0 ETH annually. Cheap in absolute terms — which is exactly why the sensitivity gets ignored, and exactly why it is the wrong frame.
The right frame is margin, not absolute cost. An L2 running at 30% gross margin on fee revenue derived from a DA cost of zero has a margin that is 100% dependent on a parameter set by the Ethereum core developers. That is not a business. That is a position.
The saturation arithmetic
My working position — stated here as an arithmetic exercise, not a prophecy — is that blob space saturates within roughly two years of any given target increase, and that the resulting fee repricing doubles rollup DA cost from its current effective baseline.
Show the work.
Demand side. Post-Pectra capacity is 6 blobs per block at target. Current median consumption is 4.7. That is 78% of target. Three growth vectors are pushing against it simultaneously. First, L2 transaction growth: Base and Arbitrum between them have been adding sequential blockspace demand at a pace that consumed the entire pre-Pectra headroom within eleven months of Dencun. Second, new blob consumers that did not exist in the original accounting: ZK proof aggregation payloads, alt-DA attestation records bridged back to L1, and — since 2025 — attestation hashes for AI inference pipelines that institutions want anchored on a neutral ledger. Third, and least discussed: rollups currently under-post. Several large L2s batch multiple blocks per blob to economize. That economization disappears the moment a competitor posts every block for latency reasons, and when it does, effective demand per rollup rises 3-5x with no change in user count.
Stack 78% current utilization against a 3-5x latent under-posting buffer and a demand growth rate that has historically run 4-6% per month in the L2 cohort, and you do not need aggressive assumptions. You need the current trend to persist for 24 months.
Supply side. Here is where the consensus is wrong in a specific, testable way.
The standard rebuttal is PeerDAS — EIP-7594, shipped in Fusaka — which decouples blob capacity from per-node bandwidth by letting nodes sample rather than download the full blob set. PeerDAS is genuinely excellent engineering. It is also irrelevant to the question being asked.
PeerDAS solves verification bandwidth. It does not solve price discovery. The blob fee market still runs on excess_blob_gas with the same update fraction and the same exponential. Raising the blob target from 6 to 12 to 24 does not lower the equilibrium price of blobs under sustained demand. It moves the choke point. If demand grows to fill 12, the base fee still climbs the same exponential until marginal demand is priced out. More capacity does not make a resource cheaper in equilibrium; it makes the transition to the choke point longer and, because the governor is a governance process running on 8-to-14-month upgrade cycles, the transition is where all the margin gets destroyed.
This is the same intellectual failure that produced the interest rate models in Aave and Compound. A two-segment kink model is a linear approximation of a nonlinear reality, and it works beautifully until utilization crosses the kink, at which point the rate doesn't gently steepen — it dislocates, and everyone who modeled borrowing cost as a smooth function discovers they modeled a step function as a line. The blob fee curve is mathematically superior because it uses an exponential rather than a kink. That does not save it. It has the same defect: the parameters are human, the update cadence is slow, and the participants price against the parameters rather than against the curve.
Correlations are the lie; liquidity is the truth. And liquidity in a resource with a 1 wei floor is a fiction maintained by a governance calendar.
The burst record
One empirical check on the above. If the blob market were genuinely abundant, spikes should be shallow and rare. They are neither.
March 2024, two weeks after Dencun: inscription-driven blob demand pushed the base fee from 1 wei to a peak measured in the hundreds of gwei, before decaying back to the floor within roughly 48 hours. The decay was fast precisely because the exponential is symmetric — which is the design, and which is also why the event left no lasting price signal.
Post-Pectra, the pattern repeated at a higher capacity but with a sharper shape. In my 90-day window there were four discrete burst events where the base fee exceeded 1,000 wei for more than six consecutive blocks. In three of the four, the trigger was not L2 transaction volume at all — it was a single large rollup's batch submission schedule colliding with another's. Two rollups, each behaving rationally in isolation, each posting at maximum batch size at the top of the hour, produce a congestion event neither of them caused and neither of them can see.
That is a coordination failure, not a capacity failure. And it is invisible to anyone measuring the sector through TVL or daily active addresses. The ledger remembers what the marketing forgets: in each of those bursts, DA cost for every rollup posting during the window rose 100-1,000x for a period of 15 to 90 minutes, and none of them had a hedge.
What the L2 revenue data actually shows
Aggregate the fee revenue of the six largest rollups against their L1 data spend and a pattern emerges that contradicts the sector's own narrative.
During floor periods — 93.6% of the time — L2 gross margin runs 88-94%. During burst periods, it compresses to 55-70% for the affected hours. Weighted annually, the margin looks healthy. Unweighted by time, it looks fragile. The distinction matters because fee schedules are set on the weighted number and stress events are governed by the unweighted one.
There is a second-order effect that almost nobody has priced. Rollup fee schedules are updated on weekly or monthly cadences. They are, by construction, lagging indicators of a leading cost. When DA cost is 0%, that lag is free. When DA cost is 40% of the marginal transaction, the lag becomes a short position against the blob market that the rollup is running involuntarily, at a size proportional to its batch cadence.
I have seen this exact structure once before at scale. In May 2022 I was watching Anchor Protocol's deposit flow drain in real time, hours before the mainstream narrative caught up, and the signal was not the price — it was the mismatch between a yield promise set on a fixed cadence and a cost of capital that repriced continuously. My fund exited stablecoin exposure entirely that week and preserved about 90% of capital while peers took the full drawdown. The lesson generalized: whenever a system quotes a price on a slow clock against a cost on a fast clock, the gap is a short option, and someone always ends up short it.
Rollups are now short a blob option they did not knowingly sell.
Due diligence is the only hedge against chaos. There is currently no forward curve on blob gas. There is no hedging venue, no futures market, no protocol-level instrument. The only available hedge is knowing your exposure number before the burst, and virtually no L2 publishes it.
Contrarian: the fee pass-through argument is weaker than it looks, and the real risk is elsewhere
Here is where I will argue against my own read.
The obvious inference from everything above is that blob repricing passes through to users and L2 fees double. That inference requires three links, and the middle one is weak.
Link one: blob costs rise. Plausible, defended above. Link two: rollups pass the increase to users. This is where it breaks. Rollups are not profit-maximizing price setters in a competitive market — they are distribution-maximizing subsidizers in a land-grab market. Base has run at negative net margin during promotional periods. Arbitrum's fee schedule has been set by governance, not by cost recovery. When DA costs rise, the first response from every large L2 will be to absorb it and defend share, not to reprice. Pass-through will be partial, delayed, and selective by cohort.
Link three: absorbed cost translates to visible stress. It will not, not immediately. It translates to a quiet compression of the treasury runway that funds the subsidies — which shows up 12-18 months later as a fee increase nobody attributes to blobs, or as a foundation grant program that stops, or as an incentive budget that gets cut. The repricing will be architecturally real and narratively invisible.
Which means the analysts watching user-facing fees will see nothing, and the analysts watching L2 treasury composition against DA spend will see everything. The signal is not the fee. The signal is the ratio of annualized DA spend to annualized fee revenue, plotted against the L1 gas price regime.
There is a second blind spot in the consensus view. The standard escape hatch is "rollups can migrate to alt-DA." Technically true. Celestia and EigenDA both offer blob-equivalent capacity at prices that undercut L1 even at the L1 floor, and migration is a config change plus a proof-system update. But the switching cost is not technical — it is the settlement-security narrative, and that narrative is exactly the thing L2 tokens are priced on. A rollup that moves DA off Ethereum has, by its own marketing, downgraded from "Ethereum L2" to "Ethereum-anchored validium," and there is a measurable discount in the market for that label. So the migration will be partial: the largest L2s stay on L1 blobs for the label and eat the cost, the mid-tier migrates and takes the discount, and the small ones leave and disappear from the conversation.
The net effect is a barbell that nobody has modeled: blob demand becomes less price-elastic at the top of the market, which means when the top of the market grows, the base fee climbs further before demand responds. That makes bursts sharper, not softer.
The third blind spot is my own. If PeerDAS matures and blob parameters are raised aggressively on a short cadence — 6 to 12 to 24 within, say, 18 months — the saturation timeline extends and the repricing slips beyond my two-year window. I will say plainly that this is the single assumption that would falsify my position, and it is observable. Watch the EIP tracker, not the blog posts.
Takeaway: the number to watch next week
Forget TVL. Forget daily actives. Plot one series: the 30-day rolling median blob base fee as a percentage of the 1 wei floor, alongside the 30-day rolling share of slots exceeding the post-Pectra target of 6.
Right now those two series are diverging — utilization up 124%, price flat at the floor. In a functioning market they converge. When they converge, they will converge violently, because an exponential fee curve does not do gentle.
The analyst who has that chart on their desk in week one of the divergence will understand what happens next. Everyone else will read about it in a quarterly deck three months after the fact, under a heading that calls the resource "abundant."
Watch the code, not the deck.
