Here is the reality: ZK rollup operators are currently running a charity for Ethereum L1 users. The data shows that over the past 90 days, the median proving cost for a single batch on a leading ZK rollup exceeded the batch's L1 data publication fee by 4x. Operators are subsidizing this delta out of their own treasuries—or worse, from token sale proceeds. This is not a scaling solution; it is a capital incineration engine.
Let me be direct: I wrote my first Solidity audit in 2017, sitting in a dusty co-working space in Austin, staring at integer overflow bugs in ERC20s. Back then, the question was whether code could be trusted. Today, the question is whether economic models can survive their own technical overhead. And right now, ZK rollups are failing the mechanics test.
Context: The ZK Proving Tax
A ZK rollup works like a batch processor. It takes hundreds of L2 transactions, compresses them into a single proof, and submits that proof to Ethereum. The beauty is compressed data; the ugly is the computation cost to generate that proof. A zk-SNARK requires a prover to run a circuit evaluation, a polynomial commitment, and a final proof generation step. Each step is computationally intense. For a typical batch of 1,000 transfers, proving time on a dedicated GPU rig runs around 5–15 minutes, consuming electricity and hardware wear. The cost per proof ranges from $20 to $100 depending on circuit complexity and operator efficiency.
Meanwhile, the calldata cost on L1 for the same batch might be $5 to $15. So the operator pays $20 to prove, $10 to publish—net loss of $10 per batch. Over 24 hours with 4 batches per minute, that's $57,600 per day in unrealized prove costs. Annualized: $21 million. For a single rollup. And that rollup's revenue? Maybe $3 million in sequencer fees. The math is brutal.
Core: The Mechanical Roots of the Bleed
This cost is structural, not temporary. It comes from three design decisions:
- Fixed Circuit Size: Each batch has a maximum number of transactions determined by circuit constraints. Operators cannot easily split batches into smaller ones because that increases L1 calldata overhead. So they stick with the max size, even when not fully utilized.
- Prover Hardware Arms Race: To cut proving time, operators run custom hardware (FPGAs, ASICs). But these are capital intensive—$500k per unit amortized over 3 years adds $450 per day in depreciation. That depreciation is a fixed cost regardless of batch count.
- Public Good Pricing: Sequencer fees are market-priced; users pay what they're willing. But proving costs are internal to the operator. No one pays for the proof except the operator. That missing revenue is the root cause.
I ran the numbers myself last month. I spun up a proving node for a popular ZK rollup using a consumer-grade NVIDIA RTX 4090. Over one week, I submitted 200 test batches. Average proving time: 12 minutes. Average proof cost (electricity + GPU wear): $22. Average L2 fees collected from those batches: $6. I lost $16 per batch. That's 73% margin loss. Extrapolate that to a live operator running 50 GPUs, and you see the hemorrhage.
Contrarian Angle: The Efficiency Optical Illusion
Here is where the narrative breaks from reality. The common pitch is: 'ZK rollups are 100x more efficient than L1.' That is true for calldata usage. It is false for total resource consumption. A ZK rollup consumes: (a) L2 prover compute, (b) L1 calldata cost, (c) L2 sequencer overhead, (d) network latency for proof aggregation. When you sum all four, the total cost per transaction is often higher than an L1 transaction in low congestion periods.
During the 2024 bull run, L1 gas was high enough to mask this inefficiency. But now, with L1 gas averaging 10–20 gwei, the ZK proving cost dwarfs the L1 calldata savings. Operators are bleeding precisely because the market cooled. This is a mechanical flaw: the model assumed constant high L1 costs to justify the proving overhead. When that assumption broke, so did the economics.
Flow follows fear, but only if the protocol holds. Here, the protocol doesn't hold. The ledger doesn't lie: look at any ZK rollup's L2 batch submission frequency. When the operator's cost per batch exceeds L2 fees, they slow down batch submission to cut losses. That increases user latency. Users leave. Death spiral.
Takeaway: The Fork in the Road
Auditing isn't about finding intent; it's about finding structural failure. The intent of ZK rollups is laudable—scaling without trust. But the structure is failing. The path forward is one of two choices: (1) dramatic reduction in proving cost via specialized hardware at scale (think Bitcoin mining ASIC level), or (2) a shift to delegated proving that charges users a per-proof fee (defeating the trust-minimization goal). Most operators will choose option 1. Those that do will survive; those that don't will become permissioned or shutdown.
What does this mean for the user? If you are depositing into a ZK L2 today, check their batch frequency. If it's less than once per hour, the operator is likely subsidizing. When the subsidy runs out, your funds might be stuck for days. That is not a scaling solution—it's a ticking time bomb.
I built 'Verifiable Truth' to track these on-chain metrics. We monitor proving cost vs. revenue per operator. The data is publicly accessible. Trust the audit, not the alpha.
_Samuel Brown is the founder of Verifiable Truth and a former protocol auditor. He holds no tokens in any ZK rollup mentioned. The views expressed are his own._