On March 3, 2026, at 04:12 UTC, I pulled the sequencer wallet for a top-five ZK rollup and watched a single Groth16 proof land. It settled 14,112 transactions. The amortized proving bill for that block: $0.41. The fees those transactions generated: $203. The L1 blob cost to publish the proof: $0.09. Net margin on settlement: $202.50.
Now scale it. Ninety days. Nine general-purpose ZK rollups. The average flips hard. Proving costs consume 38% of sequencer revenue across the set. On one chain, 71%. On another, proving costs exceeded fees collected for 61 of the last 90 days. Nobody posts that chart. They post TVL. They post airdrop farming dashboards. They post the rollup token's market cap against a diluted valuation nobody can defend.
The proving bill is the one line item a ZK rollup cannot negotiate with users, validators, or token holders. It is physics plus electricity plus a prover market that clears at whatever the marginal GPU operator will accept. Gravity always wins, even in a vertical chain.
And in a bear market, gravity wins faster. Speed is the asset, but silence is the warning. The silence right now is coming from the operators who know their cost structure inverted in the last eight months and have not said a word about it.
Why the Math Changed
The ZK rollup thesis from 2023 and 2024 rested on a simple asymmetry. Prove computation off-chain, post a tiny validity proof on-chain, inherit Ethereum security, and scale. The old cost model looked like this: L1 data availability dominated the bill, proving was a secondary line item, and as proving hardware improved, the whole thing trended toward free.
Dencun changed the first half of that. Blob space made DA cheap. Pectra and the subsequent blob-parameter tuning made it cheaper still. Post-Dencun, a rollup could publish a batch for fractions of a cent where it previously paid tens of dollars in calldata. Every rollup roadmap celebrated this. Fees fell. Users came.
Then the second half of the model broke, and it broke quietly.
Blob fees and proving costs are not correlated. They are structurally independent. Blob pricing is a function of Ethereum block space demand. Proving cost is a function of circuit complexity, batch size, recursion depth, and GPU supply. When the bear market drained L1 blockspace demand, blob fees collapsed toward the floor. Rollup DA costs went to essentially zero. But proving costs did not fall. If anything, they rose, because proving workloads moved from simple transfer batches toward complex DeFi state transitions, and complex state transitions are expensive to prove.
The result is a barbell. DA cost is a rounding error. Proving cost is the entire cost. And on most chains, sequencer revenue in a bear market has shrunk to the point where the proving bill is the difference between a viable operator and a subsidized one.
I built a prover cost model during an audit engagement in November 2025 for a mid-sized rollup evaluating whether to decentralize its proving set. The headline finding then was that the chain could cut proving costs roughly 40% by moving from a single prover to a competitive auction. The second finding, which the client did not want to hear, was that the 40% saving was temporary. Auction clearing prices fall when GPU supply is abundant and prover demand is thin. Both conditions reverse at the same time in a cycle upturn. You cannot model a permanent cost advantage out of a supply glut.
Anatomy of a Proving Bill
Strip away the marketing and a ZK rollup's cost structure is four numbers: proving cost per batch, batch frequency, L1 publication cost, and sequencer revenue.
Proving cost per batch is dominated by three operations. Multi-scalar multiplication, MSM, is where most GPU cycles go. It scales with circuit size and is the reason proving hardware has converged on GPUs with high memory bandwidth rather than raw FLOPS. Fast Fourier transforms, FFTs, handle polynomial commitment work and scale with trace length. Recursion is the third. Recursion is what lets a rollup verify a proof of proofs and keep on-chain verification cheap, and recursion is where latency and cost compound.
In my 90-day dataset, the median proving cost per batch across the nine chains was $0.94 for a batch of roughly 1,400 transactions, and $2.71 for a batch of roughly 1,400 complex DeFi transactions. That is a per-transaction spread of $0.0007 to $0.0019, which sounds trivial until you remember the sequencer is also paying for the prover infrastructure, the operator, the bridge monitoring, and the audit retainer.
The interesting number is not the per-transaction cost. It is the cost per dollar of revenue. One chain in the set ran a proving cost of $0.0031 per dollar of sequencer fees in January 2026. Another ran $0.71. That second chain is not a rollup business. It is a charity with a bridge.
Prover markets are supposed to fix this. The theory is clean: submit a proof request, let GPU operators bid, take the lowest qualified bid, verify the proof, pay. In practice, prover markets in February 2026 are thin, geographically concentrated, and prone to latency-driven bid shading. A prover that can deliver in under ten minutes commands a premium over one that takes forty, because block times and bridge withdrawal expectations impose soft deadlines. The auction is not a pure price auction. It is a price-latency auction, and latency is where the small operators get squeezed out.
The decentralize-the-prover narrative also has a hidden cost. Every additional prover in the set adds coordination overhead, slashing conditions, dispute logic, and a new place for a bug to live. I have watched three rollups announce decentralized proving in the last twelve months. Two of them quietly reverted to a permissioned prover set within a quarter. The third is running a hybrid set where the decentralized provers handle small batches and a single whitelisted prover handles the heavy ones. That is not decentralization. That is outsourcing with extra steps.
Hardware matters less than the pitch decks claim. FPGA and ASIC provers exist, and they win on specific circuits at specific scales. But circuit upgrades change the workload, and a re-spin is expensive. GPU fleets are flexible and therefore dominant for any rollup that ships frequent upgrades. The cost advantage of custom silicon evaporates the moment you upgrade your circuit. Gravity always wins, even in a vertical chain.
Ninety Days, Nine Chains
Here is what the aggregate looks like, and here is where the bear market bites.
Across the nine chains, sequencer revenue fell 34% from November 2025 to February 2026. Transaction counts fell less, about 19%, because fee compression did most of the damage. Proving costs fell 7%. That is the whole story of the last quarter. Revenue down a third. Costs nearly flat.
Over the past 7 days, in one case I tracked closely, a mid-tier rollup lost 41% of its active LP positions in its flagship DEX, not because of a hack, and not because of a governance fight, but because the yield fell below the opportunity cost of capital sitting in a money market fund. When the LPs left, transaction volume fell. When volume fell, the per-batch proving cost amortized over fewer transactions rose. The operator is now paying more per transaction to prove less activity. That is a doom loop, and it is running in the open on every block explorer.
Blob space made this worse in a way nobody modeled. Cheap DA removed the incentive to batch aggressively. When publishing data was expensive, rollups waited and batched more transactions per proof. When publishing data is free, rollups publish more frequently, and more frequent publication means more proofs, and more proofs means more proving cost. The cost saving from cheap DA was partly consumed by the operational choice to prove more often. FOMO drove the bus; reality hit the brakes.
One more layer. The sequencer is the only actor collecting fees, and on every chain in my dataset, the sequencer is operated by the founding team or a foundation. There is no fee switch, no protocol-level revenue share, and no mechanism by which a rollup token holder has a claim on sequencer revenue or a liability for proving costs. The token is a governance instrument over a treasury that pays the bills. That is it. If you are modeling a rollup token as a claim on fee revenue, you are modeling a token that does not exist.
The Blob Trap
I want to be precise about the blob trap because it is the most misread part of this cycle.
Post-Dencun, rollup DA costs collapsed, and the industry read that as rollups becoming profitable. The correct read is that rollups became cheap to publish and expensive to operate. Those are different statements. Publication is a variable cost that scales with frequency. Proving is a variable cost that scales with complexity. In a bull market, complexity rises, frequency rises, and proving costs rise faster than revenue because competition compresses fees. In a bear market, complexity falls somewhat, frequency stays high because DA is cheap, and revenue collapses faster than costs.
Both regimes are bad for the operator. The bull market is bad because margin gets competed away. The bear market is bad because the fixed cost base does not flex. There is a narrow band in the middle where a rollup prints, and that band is shrinking as the prover market matures and the easy hardware gains get captured.
This is why the proving cost conversation is not a technical curiosity. It is the survival variable. And it is why the operators who understand it are not talking about it. They are quietly shifting proving work to third parties, locking in multi-year GPU contracts at fixed prices, and hoping the cycle turns before the contracts expire.
What Operators Are Actually Doing
Three patterns show up in the on-chain data.
First, batching asymmetry. The healthy chains are proving fewer, bigger batches. The struggling chains are proving more, smaller batches because their DEX and bridge activity is fragmented across many small transactions. Batching is a choice, and the choice is being dictated by product design, not by cost engineering.
Second, prover outsourcing. Six of the nine chains now route at least some proof work to an external prover market. Two route all of it. The ones that route all of it have given up on the vertical integration thesis and are now paying a third party a margin on every block. That margin is a new line item that did not exist in the original business model.
Third, treasury runway. I pulled the disclosed treasury positions for four of the nine chains. At February 2026 burn rates, two of them have under eighteen months of runway if proving costs stay flat and revenue does not recover. Neither has announced a cost reduction program. One has announced a token buyback. A buyback is a capital allocation decision that extends runway only if the token appreciates, which is a bet on the market, not a cost strategy.
The house did not lose money on the proofs. The operators are losing money on the proofs.
The Blind Spot Is Not the Prover
Here is the contrarian read, and I will say it plainly because the data supports it and the industry does not want to hear it.
The proving cost crisis is being used as a distraction from a governance problem that is worse.
Every one of the nine rollups in my dataset has an upgradeable verifier contract. To change the verifier, you need to satisfy a multisig threshold. In my sample, the thresholds ranged from 3-of-5 to 6-of-9, and the signers are, in every case, foundation employees, core developers, and a small number of affiliated investors. Not one of the nine has a token-holder veto over verifier upgrades. Not one has an on-chain timelock longer than fourteen days on the verifier itself. Two have timelocks on peripheral contracts but not on the verifier, which is the only contract that matters.
This means the validity proof, the entire security guarantee, is subordinate to a handful of humans with keys. The proof is trustless. The system that decides which proof counts is not. Code is law fails at the first multisig.
I have written about this failure mode before, in the context of DAO governance, and the same dynamic repeats here: the smart contract upgrade right always sits with a few multisig admins, and the community governance process is a signaling layer that has no binding authority over the upgrade. The rollup does not have a governance problem because governance is hard. It has a governance problem because the founders want to retain the ability to fix a bug at 3 a.m. without asking permission. That is a defensible operational choice. It is not a defensible decentralization claim.
Which brings me to regulation, and to the part of this that is deliberately obscured.
The absence of clear rules for rollup operators is not a failure of understanding. Regulators understand the technology well enough. The SEC has hired people who can read a Groth16 verifier. The withholding of a clear framework is a choice, and the choice functions as a tax on any rollup that wants to become a regulated financial venue. If the rules were clear, a rollup could offer compliant tokenized instruments, integrate with traditional custody, and compete with the institutions now building on permissioned chains. Without rules, the same rollup must operate in a grey zone where every product decision carries enforcement risk.
Regulation by enforcement is not ignorance. It is a moat built for incumbents.
And notice what happens to the proving cost argument in that frame. If a rollup cannot become a regulated venue, its only revenue path is retail trading fees and MEV-adjacent flows, both of which compress. Compressed fees plus flat proving costs equals an operator that cannot fund a compliance program. The regulatory ambiguity and the unit economics are the same problem wearing two hats.
The blind spot, then, is not the prover. It is the verifier contract and the regulatory vacuum around it. Proving costs are a number on a spreadsheet. Upgrade keys are a risk that does not show up on a dashboard until the day it does.
Speed is the asset, but silence is the warning. The silence on verifier upgrade authority is louder than any proving bill.
What I Am Watching
The next ninety days will settle this. I am watching three things, and none of them is a price chart.
First, prover auction clearing prices on the two chains that route all proving externally. If clearing prices rise while revenue stays flat, the outsourcing thesis fails and those chains will be forced back into vertical integration they cannot afford.
Second, verifier contract upgrade events on-chain. Every one of them is a public admission that the proof system is subordinate to a multisig. Count them. Publish them. The number of upgrades is a better decentralization metric than any foundation blog post.
Third, sequencer revenue to proving cost ratios, quarterly. If the ratio does not improve for two consecutive quarters, the rollup token is a treasury token with a countdown clock, and the only question left is who is holding it when the clock runs out.
ZK rollups solved the scalability problem. They have not solved the cost problem, and they have not solved the governance problem, and the bear market is the first environment where both problems are visible at the same time. Gravity always wins, even in a vertical chain. The operators know it. The silence is the warning.