The seven-day median proving cost for one general-purpose zk-rollup I monitor settled at $0.041 per transaction last week. The median fee it charged users was $0.038. For forty-one consecutive days, the operator cleared a negative gross margin on every batch it submitted to Ethereum. The anomaly is not the loss. The anomaly is that the loss is invisible. No public dashboard publishes it. No block explorer surfaces it. The sequencer's subsidy appears to users only as a marginally better fee schedule, and the market reads "cheaper" as "healthier."
I have seen this pattern before. In 2020, I built a Python backend to scrape yield-farming data across Uniswap and Compound, tracking more than a thousand daily liquidity-pool entries and computing impermanent-loss scenarios for portfolios above $2 million in simulated value. The lesson from that exercise was not that the yields were fake. It was that the headline number and the underlying economics could diverge for months before anyone noticed. Efficiency hides in the edge cases nobody audits. What follows is the edge case for Layer2 in 2026: a cost line that decides the outcome and that almost nobody publishes.
The consolidation phase we are in โ price chopping within a narrow band, spot volume drifting, funding rates flat โ is precisely the regime where cost structures become visible. In an expansion, emissions and narrative subsidy paper over negative unit economics. In a sideways tape, there is no narrative subsidy left, and the underlying ledger has to clear on its own. That is why I spent the last two weeks rebuilding the cost stack for the rollups I track, from raw blob data through to proving invoices.
The background matters. Ethereum's Dencun upgrade, live since March 2024, introduced proto-danksharding via EIP-4844. It created a dedicated fee market for "blobs" โ data-availability space that rollups use to post compressed transaction data, separate from calldata. The design intent was straightforward: reduce the cost of posting L2 data to L1 so rollups could pass savings to users and, eventually, retain margin.
For a while, it worked exactly as advertised. Blob fees collapsed from dollars to cents, then to fractions of a cent during low-demand windows. Rollup transaction fees followed. Every L2 roadmap slide that had promised "ten times cheaper" was suddenly conservative, and the narrative consolidated around a single sentence: Dencun fixed Layer2 economics.
But Dencun was always a partial story. Posting data to L1 is one line in a rollup's cost stack. It is not the whole stack. For zk-rollups in particular, the dominant marginal cost is not data availability. It is proving. And proving costs did not collapse with blob fees, because they are driven by a different market entirely: prover hardware, circuit design, and the throughput of the proving pipeline.
There is a second reason the narrative hardened so quickly. DA costs are public. Blob fees are observable in real time, granular to the slot, and dramatic in their decline. Proving costs are private. They live inside sequencer infrastructure, they are reported โ if at all โ as blended averages, and they do not produce a satisfying chart. The market optimized the line it could see and inferred that the line it could not see had also improved. That inference is the subject of this analysis.
I want to state my method before I state my findings, because the method is the contribution. A conclusion without a reproducible method is an opinion wearing a lab coat. I reconstructed the cost stack from four sources, weighted by reliability. First, public blob data cross-referenced against the EIP-4844 fee market, to establish the data-availability line with high confidence. Second, on-chain settlement transactions, from which batch cadence, posting frequency, and verification gas can be inferred directly. Third, published proving-throughput benchmarks, reconciled against observed batch latency to estimate the proving line at medium confidence. Fourth, protocol-published figures, treated as the lowest-confidence source and weighted last, because self-reported blended averages are exactly where the constraint hides.
The four cost lines that matter for a zk-rollup are data availability, proving, settlement and verification, and operations. Data availability covers posting state diffs and calldata to L1 or to an external DA layer. Proving covers generating validity proofs for each batch, denominated in prover hardware amortization and electricity. Settlement and verification covers the L1 gas cost of verifying the proof on-chain. Operations covers sequencer infrastructure, monitoring, and the human cost of running the system.
Here is the decomposition I derived for the mid-tier rollup, expressed as a share of total marginal cost per transaction, averaged over the last thirty days.
Table 1 โ Marginal cost decomposition (share of total, 30-day average)
| Cost line | Share of marginal cost | 90-day trend | |---|---|---| | Data availability | 9% | Down sharply | | Proving | 61% | Flat to up | | Settlement / verification | 12% | Flat | | Operations | 18% | Flat |
The table is the entire argument in four rows. DA โ the line everyone optimized for, the line Dencun was built to fix โ is now under a tenth of marginal cost. Proving is nearly two-thirds. The upgrade addressed the smallest line item and left the largest one untouched.
This is not a criticism of EIP-4844. It did what it was designed to do. It is a criticism of the inference the market drew from it โ that cheaper DA equals cheaper rollups equals viable rollups. Those are three different claims, and only the first is true.
Proving cost is a function of proof system, circuit size, and prover utilization. I built a rough model across three proof families โ Groth16, PLONK-style universal setups, and STARK-based systems โ normalized to a common "cost per million constraints" unit so the families can be compared on one axis.
Table 2 โ Normalized proving cost, cost per million constraints (indexed to Groth16 = 1.0)
| Proof system | Relative cost | Setup flexibility | Recursion support | |---|---|---|---| | Groth16 | 1.0 (baseline) | Per-circuit | Limited | | PLONK-style | 1.4 | Universal | Good | | STARK | 3.2 | Transparent | Strong |
The spread is the point. A STARK-based prover can cost more than three times a Groth16 baseline for the same constraint count. That multiplier is not a defect โ it buys transparency and recursion, which matter for security and for compression. But it lands directly on the marginal cost line, and it does not respond to blob fees at all. You can cut DA cost to zero and the proving line would not move by a basis point.
A rollup does not prove every transaction individually. It batches. The economics depend on batch size and proving latency, and those two variables are in tension. If proving is slow, the sequencer must either hold batches longer โ increasing latency and, with it, the risk of stale state and front-running exposure โ or prove smaller batches more often, increasing cost per transaction. There is no free lunch; there is only a trade between latency and unit cost.
I tracked batch cadence for the same rollup across ninety days. The pattern is stable: roughly one batch every few minutes, with proving latency consuming most of the interval. The sequencer is not constrained by L1 posting cost. It is constrained by how fast it can generate proofs.
Table 3 โ Batch cadence and implied cost (90-day window)
| Metric | Observed range | Implication | |---|---|---| | Batch interval | 2โ6 minutes | Proving-bound, not DA-bound | | Proving latency share of interval | 60โ75% | Dominant constraint | | Verification gas per batch | Low, stable | Not the bottleneck | | Blob posting cost per batch | Negligible | Dencun achieved its goal |
This is the structural asymmetry the market has not priced: L2 cost competition is being fought on the DA line, but the war is decided on the proving line. Every roadmap that treats Dencun as the finish line has misidentified the binding constraint.
To understand why DA costs fell so far, it helps to look at how the blob market functions. EIP-4844 introduced a separate fee market with its own supply target: the network targets a set number of blobs per block and adjusts the blob base fee up or down depending on demand relative to that target. When demand is below target โ which it was for long stretches after launch โ the blob base fee decays toward zero, and rollups posting during those windows paid almost nothing. The mechanism is elegant, and it did precisely what it promised. But it also created a false signal. A rollup's DA cost is now a function of when it posts, not of what it does. An operator that batches and posts during quiet windows enjoys near-zero DA cost regardless of its engineering quality. That flatters the blended average and further obscures the proving line, which does not care what time of day it is.
If marginal cost is $0.041 and the charged fee is $0.038, the operator loses $0.003 per transaction before fixed costs. Multiply by observed throughput and the daily subsidy becomes material. I will not publish the operator's absolute numbers โ the point is the structure, not the specific firm โ but the sign is unambiguous and it persisted for six weeks. Where does the money come from? Three sources: treasury runway, the war chest raised at token launch or in private rounds; token emissions, inflation used to pay for activity that would not otherwise occur; and venture funding earmarked for "growth," which is a polite term for subsidized usage. All three are finite. In a bull tape, they are replenished by narrative, and narrative is cheap when prices rise. In a sideways tape, they are not replenished, because the same tape that flattens price also flattens the story. This is why consolidation matters analytically: the sideways market is not a pause in the story; it is the audit.
Table 4 โ Subsidy source quality (my assessment)
| Source | Sustainability | Failure mode | |---|---|---| | Treasury runway | Finite, quantifiable | Depletion date is knowable | | Token emissions | Self-defeating | Mercenary capital exits on schedule | | Venture subsidy | Discretionary | Stops at the next funding winter |
The third row is the one operators understate. Venture subsidy is not a revenue line; it is a grace period. When it ends, the operator must either raise the fee โ and lose the "cheap" positioning that DA cheapening gave it for free โ or keep losing money at scale.
Let me describe the sequence I reconstructed, without naming the operator. From day zero to day twelve, the operator maintained batch cadence and passed through the full DA savings to users, cutting the fee to undercut a competitor. Margin was already negative, but the treasury runway was long and the narrative was favorable. From day thirteen to day twenty-eight, DA fees ticked up modestly as blob demand returned. The operator did not raise user fees โ doing so would have broken the "cheap" positioning โ so the negative margin widened. From day twenty-nine to day forty-one, proving latency drifted upward as circuit complexity grew with state size. The sequencer responded by holding batches marginally longer, which kept unit cost roughly flat but increased latency. Users noticed the latency. Nobody noticed the margin.
At no point did the operator's public communication misrepresent anything. There was simply nothing to see. The subsidy was real, the fees were real, and the loss was structural. This is the failure mode the DA narrative trained the market not to look for.
The one genuine cost lever zk-rollups have is recursion โ aggregating many proofs into a single proof, so that the expensive verification step scales sublinearly with the number of underlying batches. Recursion is why STARK-based systems, despite a higher base cost per constraint, can be competitive at scale: they compress more aggressively. But recursion is a scale play. It pays off when batch volume is high and steady. In a sideways tape, volume is neither, so the recursion dividend is deferred. The operator pays the higher base proving cost now and books the compression benefit later โ if the volume arrives. In a consolidation regime, every architectural bet that assumes scale is a financing decision, not an engineering one.
It is worth separating zk-rollups from optimistic rollups here, because the cost stacks differ. Optimistic rollups do not prove every batch; they assume validity and rely on fraud proofs, which pushes cost onto data availability and the challenge window. For them, Dencun was closer to a complete fix, because their dominant marginal cost genuinely was DA. This distinction explains a divergence I observed in the data: optimistic rollups passed through fee reductions faster and more durably than zk-rollups after Dencun. That was not a management decision. It was a structural consequence of where each architecture spends its marginal dollar. The same upgrade that healed one architecture's economics merely exposed another's.
Table 5 โ Architecture comparison, post-Dencun marginal cost
| Architecture | Dominant marginal cost | Dencun effect | |---|---|---| | Optimistic rollup | Data availability | Large, durable reduction | | zk-rollup | Proving | Small, transient reduction |
I want to be careful with this table. "Transient" does not mean "reversed." The DA reduction is real for zk-rollups too; it is simply too small a share of the stack to change the sign of the margin. The bottleneck moved from DA to proving, and proving is where the money now goes.
I owe the reader an honest accounting of uncertainty. The DA line is high confidence: blob data and the EIP-4844 fee market are public and verifiable. The settlement line is high confidence: verification gas is on-chain. The operations line is low confidence: it is an estimate of infrastructure and human cost, and it will vary by operator. The proving line is medium confidence: I derived it from published throughput benchmarks reconciled against observed batch latency, and it is sensitive to assumptions about prover utilization. The conclusion I draw โ that proving dominates marginal cost โ is robust across a wide band of those assumptions. The exact share, sixty-one percent, is not. Readers should treat the direction as reliable and the magnitude as approximate. That is the same discipline I applied in 2022, when I audited failing lending protocols and separated technical debt from asset backing: the sign of the finding was certain, the precision was not.
Here is where I have to push against the prevailing framing, including framing that flatters my own thesis. The market's favorite explanation for L2 margin pressure is "liquidity fragmentation" โ the idea that too many rollups are splitting a fixed pool of liquidity and that consolidation into a few winners will restore health. I do not buy it, and the data does not support it.
Liquidity fragmentation is not a market condition; it is a product narrative. It is the story that justifies funding the next rollup, the next interop layer, the next unified-liquidity primitive. The claim is that fragmentation is the disease and the product is the cure. But the disease, if you read the cost stack, is proving economics โ and no interop layer fixes a proving-cost line. Bridging liquidity across ten chains does not reduce the cost of generating a validity proof on any of them. You can unify every pool in the ecosystem and the proving invoice would be unchanged.
Correlation is not causation, and the correlation between "more rollups" and "thin liquidity" is a coincidence of timing, not a causal mechanism. Rollups launched in a wave; liquidity thinned in a bear-and-sideways regime; the two facts share a calendar, not a cause. If fragmentation were the true constraint, merging liquidity would restore margins. It would not, because the margin is consumed on the proving line, which is orthogonal to how liquidity is distributed.
The blind spot cuts the other way too. The DA-cheapening narrative was so satisfying that it suppressed scrutiny of the proving line for two years. Everyone optimized the line that was easy to measure โ blob fees are public, granular, and dramatic โ and ignored the line that was hard to measure, because proving costs live inside private infrastructure and are reported, if at all, as blended averages that conceal the constraint. The metric that is easy to watch is not the metric that decides the outcome. That is the same trap I documented in 2021, when reported NFT volume was easy to watch and unique-buyer concentration was not, and a $5 million gap opened between headline volume and actual distinct buyers.
The forward signal for the next several weeks is not price. It is the proving-cost curve against the fee curve. Watch two things. First, whether any zk-rollup publishes a transparent proving-cost-per-transaction figure โ the ones that do will be the ones whose margins can survive the sideways tape. Second, whether the fee compression that followed Dencun reverses as DA demand returns; if fees tick up while proving costs stay flat, margins briefly heal and the market will misread a reprieve as a turnaround.
The question worth sitting with is not which L2 wins. It is whether the proving line can fall faster than the fee line before the treasury that subsidizes the gap runs dry. Everything else is commentary.
