Hook

One sub-scale ZK rollup provisioned two additional prover clusters last quarter and cut batch proof latency by 61%. Trailing 30-day utilization: 35.0%.

The second number decides whether the first was a cost optimization or a marketing expense. I have the prover logs and the fee-recipient traces; I am not publishing the deployment identifier, because the last time I attached a name to a number, that name's counsel sent a letter I had to answer. The logs are public if you know which endpoints to poll.

The arithmetic: at 35.0% utilization the effective proving cost is $27.20 per batch. Sequencer revenue is $22.00. Data availability adds $2.60. A 5.2% first-pass verification failure rate adds $1.41 in reproving. Net position: negative $9.21 per batch, 31 batches a day, roughly $8,600 a month. The rollup is not failing. It is bleeding in increments too small to trigger an incident response.

Context

Blobs changed the cost stack. In doing so they exposed it.

Before EIP-4844, a ZK rollup's dominant cost line was calldata. Operators priced fees around it, marketed against it, and hid behind it. Post-4844, data availability for blob-posting rollups fell by roughly an order of magnitude. That line item shrank. The others did not.

Proving is off-chain compute. Accelerators, electricity, cooling, redundancy, and the engineers who keep witness generation from silently diverging from the circuit. None of it touches L1 gas. All of it is denominated in dollars.

Settlement is L1 verification gas — fixed per batch, small, and uninteresting.

So the post-blob cost structure is a barbell: one line that scales with throughput and gas price, one that is fixed in dollar terms and scales with provisioned capacity, and one that is negligible. The industry press release said rollups are cheap now. What actually happened is that the variable cost collapsed, the fixed cost stayed, and the fixed cost became visible for the first time.

This matters because fixed costs behave differently under demand shocks. A variable cost that doubles under load is a pricing problem. A fixed cost that sits idle under flat demand is a solvency problem. Hype evaporates; solvency remains.

For scale reference: the operator in this sample sustains roughly 0.7 TPS and 62,000 transactions a day. It is not an outlier in the sub-scale cohort. It is representative of it. Nine general-purpose ZK rollups have shipped mainnet proving across four proving-system generations, and at least five run their own clusters rather than renting. The shared prover markets that would absorb idle capacity are live but thin — liquidity on the proof side is measured in tens of batches per hour across all of them, which is not enough depth to let an operator sell surplus without moving the clearing price. So surplus capacity is stranded. It is paid for either way.

Core

Methodology first, because numbers are only as good as their source. I sampled 4,100 consecutive batches from the operator's public prover endpoint over 132 days, cross-referenced each proof submission against the fee-recipient address on L1, and reconstructed accelerator-hour consumption from the cluster's own timing attestations. Revenue is taken from realized fee transfers, not from the fee schedule. Where I had to model — reprove cost, depreciation — I state the assumption.

Finding one: the denominator.

The marginal cost of a proof is not the accelerator-hour cost. It is the accelerator-hour cost divided by utilization.

One proof takes 34 minutes of wall-clock time on an 8-accelerator cluster. At a contracted $2.10 per accelerator-hour, that is $16.80 per cluster-hour and $9.52 in direct compute per proof. Every operator I have audited quotes me that number. None of them quotes the denominator.

A prover cluster cannot be spun up and down per batch. Witness generation, the setup artifacts, and the aggregation pipeline have to stay warm, and an operator that pauses proving to save money stops finalizing. Capacity is therefore provisioned against peak and against ambition, not against average demand. This operator doubled capacity to halve latency. Throughput did not move. Utilization fell.

At 35.0% utilization, the $9.52 direct cost becomes $27.20. Same proof, same hardware, 2.86x the effective price. Capacity added to improve latency is a marketing expense wearing an engineering costume. I have watched three teams make that investment this year and describe it to their boards as a cost optimization.

Finding two: the reprove tax.

Across the 4,100-batch sample, 213 proofs failed first-pass verification. The failure modes concentrate in precompile edge cases — one keccak path and two range-check boundaries — and each failure costs a full re-prove plus a delay in the aggregation queue. That is a 5.2% multiplier on the proving line, and it appears in no fee model I have reviewed, because fee models are written by people who assume proofs succeed.

Finding three: the depreciation window.

Self-hosted accelerator fleets are not assets on a three-year straight line. Circuit versions change. A cluster that cannot run the current proving system is a depreciated cluster. Effective useful life across the fleets I have reviewed has run 19 to 23 months, not 36. Operators carrying capex on a 36-month schedule are understating per-batch cost by roughly a third.

Finding four: revenue concentration.

The $22.00 per batch is an average, and averages lie in a specific way. The median transaction in the sample paid $0.004. The average was carried by a thin tail of priority-fee payers — 4.1% of transactions produced 38% of batch revenue. When that tail thins, as it does in every consolidation, revenue does not decline proportionally. It falls off a cliff.

One cost I excluded deliberately: verification key upgrades. Two operators I reviewed rotated proving systems mid-cycle, invalidating in-flight proofs, forcing a reprove backlog, and costing one of them 41 hours of finality downtime over a weekend. I have not priced that. It is not a recurring cost. It is the cost that shows up in the quarter you cannot afford it.

The consolidated position, per batch:

• Sequencer revenue: $22.00
• Effective proving cost at 35.0% utilization: $27.20
• Reprove multiplier (5.2%): $1.41
• Blob posting and settlement: $2.60
• Net: –$9.21

Thirty-one batches a day. Eight thousand six hundred dollars a month, before the sequencer's own infrastructure, before the two engineers who own the circuits, before the legal review the operator has now scheduled.

Contrarian

The bulls are not wrong about the curve. They are wrong about the timeline.

Every input in that model improves. Accelerator price-performance has continued to compound. Recursive proof composition and proof aggregation have cut per-transaction proving cost by more than an order of magnitude across four proving-system generations. Shared prover markets let sub-scale operators rent capacity instead of provisioning it, which converts a fixed cost into a variable cost — the single most important structural fix currently available. And utilization is not a constant. It is a function of demand.

Which is what the bearish read keeps missing: the model bends hard at scale. At 70% utilization the same operator's effective proving cost falls to $13.60 and the position turns positive by $5.40 per batch before any fee change. Nothing in the protocol has to improve. Demand has to arrive.

So the operative question is not whether ZK rollups work economically. It is how many of them reach the utilization floor. At current demand, that number is smaller than the number of rollups currently publishing roadmaps. Consolidation is the mechanism here — not a fee switch, not a token, not a governance vote. Fewer operators, higher utilization, and survivors that rent proving rather than own it.

Audits reveal what code conceals. This particular concealment is in a spreadsheet, which is why no auditor has been asked to look.

Takeaway

No rollup publishes its proving utilization. It is the single number that determines whether the operator is solvent at current demand, and it appears in no dashboard, no quarterly report, and no token narrative. Institutional allocators have begun asking about sequencer revenue and DA costs. They are not yet asking the denominator.

Over the next twelve months, the ranking that matters will not be which rollup has the fastest prover. It will be which one is willing to publish the number. Ledger integrity precedes market sentiment, and a cost model that only clears at 70% utilization is not a cost model — it is a forecast wearing a fact's clothing. Precision is the only risk mitigation. The question I would put to any allocator underwriting rollup exposure right now is not where the fee switch is. It is what the operator's utilization rate was last month, and whether anyone on the team can produce that figure without calling an engineer.