The pitch deck reads like a fairy tale: infinite scalability, negligible fees, and the holy grail of Ethereum’s future. But the code tells a different story. Over the past 90 days, I have dissected the on-chain cost structures of three major ZK-rollup operators — zkSync Era, Scroll, and Polygon zkEVM. The numbers are stark. Proving costs for a single batch of transactions on these networks range from $1,200 to $8,000, depending on the computational complexity of the circuit. When you divide that by the number of transactions per batch — often 500 to 2,000 — the per-transaction cost contributed by proof generation alone is between $0.60 and $4.00. Compare that to the current user fee of $0.02 to $0.10 per transaction. The delta is not just a subsidy; it is a hemorrhage. The operators are bleeding capital, and the market has not priced this in.
Context
The Layer 2 narrative has dominated crypto discourse since 2021. The thesis: Ethereum’s base layer cannot scale; rollups will inherit the users and the volume. ZK-rollups, in particular, were hailed as the ultimate solution — offering trustless finality, fast withdrawals, and security derived from the main chain. The industry rallied around this vision. Venture capitalists poured billions into zkEVM projects, and users flocked to the promise of cheap, fast transactions. The hype cycle peaked in early 2024 with the launch of several mainnets, each touting “Ethereum-equivalent” execution environments.
But there is a structural flaw that the marketing silences: the cost of generating zero-knowledge proofs does not scale linearly with transaction volume. It scales with the size of the circuit and the frequency of aggregation. Most operators currently run their provers on expensive cloud GPUs or custom hardware — think NVIDIA A100 clusters or FPGA arrays. The electricity, compute time, and hardware depreciation are real. And they are not reflected in the transaction fees users pay. The operators are effectively burning their treasury to attract users, hoping to achieve a network effect that will eventually justify raising fees or finding cheaper proving methods.
Core: The Systematic Teardown of Proving Economics
I spent three weeks working with a research team to model the proving cost for a hypothetical rollup processing 10 million transactions per month. We used conservative estimates: a medium-sized circuit (2^18 constraints), batch size of 1,000 transactions, and a proving time of 15 minutes per batch on a single A100 GPU. The cloud rental cost for that GPU is roughly $3.50 per hour. At 15 minutes per batch, that is $0.875 per batch. But that is the compute cost only. The real killer is the prover’s memory bandwidth and the cost of multiple GPUs for parallel proving. In practice, proving requires at least 4 GPUs to maintain a reasonable latency, bringing the per-batch compute cost to $3.50. Then add storage for the setup parameters, network overhead, and developer salaries. The total operational cost per batch balloons to $5.00 to $6.00.
Now multiply that by 300 batches per month (assuming one batch every 2.5 minutes on a fast rollup). That is $1,500 to $1,800 per month in operational costs for a single sequencer-prover setup. But the user fees from those 300,000 transactions (300 batches * 1,000 txs) at an average fee of $0.05 per transaction yield only $15,000 in revenue. The operator’s cost is 12% of revenue — and we haven’t included the cost of posting data to Ethereum (L1 calldata or blob fees). When you add L1 data availability costs, which can be $0.05 per transaction during peak L1 congestion, the total cost per transaction spikes to $0.10 to $0.15, while revenue per transaction remains at $0.05. The operator is losing $0.05 to $0.10 per transaction.
This is not a viable long-term model. It works only if (a) the operator has a large token treasury to subsidize usage, (b) they expect to eventually monetize via MEV or token price appreciation, or (c) they achieve massive scale to negotiate cheaper hardware costs.
I have audited the tokenomics of three ZK-rollup projects. Two of them allocate less than 15% of their token supply to operational subsidies. At the current burn rate, they will exhaust that allocation within 18 to 24 months. The third project has a more sophisticated model: it uses a portion of sequencer revenue to repurchase and burn tokens, theoretically creating deflationary pressure. But this only works if the transaction volume grows exponentially — which is the very assumption that is being tested.
The data from the past six months does not support exponential growth. Daily transaction counts on most ZK-rollups have plateaued between 50,000 and 200,000, well below the millions needed to achieve meaningful hardware economies of scale. The user growth is being driven by airdrop farming and liquidity mining incentives, not organic demand. Once the incentives attenuate, volume will collapse, and the operator will be left with idle hardware and a massive overhead.
Contrarian: What the Bulls Got Right
I am not here to dismiss the entire ZK-rollup thesis. There are legitimate technical advantages that bulls have correctly identified. First, ZK-proofs offer trustless finality — users do not need to wait for the challenge period that plagues optimistic rollups. This matters for high-frequency trading and cross-chain composability. Second, the aggregation of proofs into a single submission to L1 drastically reduces congestion on Ethereum. During the NFT mint mania of March 2024, zkSync Era processed over 1.5 million transactions in a single day while paying only $2,000 in L1 fees — a 100x reduction compared to doing the same volume on L1. That is real efficiency.
Third, and most importantly, the bulls are correct that proving costs will eventually decline. The hardware is improving: custom ASICs for ZK-proving are in development at companies like Ingonyama and Cysic. Estimates suggest that a dedicated ASIC could reduce proving cost by a factor of 10 to 20 compared to GPUs within two years. If that happens, the per-transaction proving cost could drop to $0.01 to $0.05, making the economics sustainable even at current fee levels. The question is timing — and whether the operators have enough runway to survive until then.
Takeaway: The Responsibility of Transparency
The industry owes its users an honest accounting of these subsidies. The average L2 user sees a fee of $0.02 and assumes that is the true cost. It is not. The real cost is hidden in treasury burns, VC capital, and eventual dilution. I have written about this before — complexity hides the body. The body here is the unsustainable proving expenditure.
Read the code, not the pitch deck. Look at the sequencer’s balance sheet. If a rollup operator is losing money on every transaction, the only question is when the music stops. Either the proving costs come down faster than the incentives wear off, or the project collapses into a death spiral of rising fees and declining volume. The data will tell us which scenario unfolds. Until then, the smart money is hedged, the sidelines are watching, and the protocols are bleeding.
Trust nothing. Verify everything.