Gas is back. Fees are up. TVL is mooning. And every L2 marketing page is screaming about how their rollup is the cheapest lane to Ethereum.
But nobody wants to talk about what the proving side of the equation looks like.
I spent last week running the actual cost models on ZK proving for the top five ZK rollups. Not the whitepaper math. Not the "we expect costs to decrease 10x by 2027" slide deck math. The real numbers: circuit complexity, witness generation time, hardware rental rates, and current ETH gas prices.
The result? Most of these operators are running at a loss on every single batch they post. And they know it. They're just betting that you'll keep depositing while they figure out how to make the tech cheaper than the competition's smoke and mirrors.
Let me show you the ledger.
The Context: Everyone's a Victim of Their Own Marketing
The L2 narrative in 2025 is a race to the bottom on transaction fees. Arbitrum charges a few cents. Base charges less. ZK Sync charges almost nothing. And the newest entrants are literally paying users to transact.
Smart money sees through this. We don't chase the cheapest gas fee; we chase the most sustainable business model. And right now, the ZK ecosystem has a structural cost problem that no amount of token incentives can fix.
The core issue is simple: proving a ZK rollup batch is computationally expensive. Unlike optimistic rollups that assume validity and let anyone challenge, ZK rollups must prove every single transaction mathematically before posting to Ethereum. That proof generation requires massive parallel processing power, specialized hardware, and electricity that isn't free.
I've been tracking the cost curves since 2021 when the first production ZK rollups hit mainnet. The narrative back then was that hardware would get cheaper, circuits would get more efficient, and proving costs would plummet. The first part happened. The second part is still a work in progress.
Let's talk about what that means in USD terms today.
The Core: Where the Money Actually Goes
I broke down the cost structure for a typical ZK rollup batch containing 1,000 transactions. The proving time on modern GPU clusters runs between 15 to 45 minutes depending on circuit complexity. The hardware cost, amortized over a 3-year lifecycle, is roughly $280 per hour for a competitive proving cluster.
Do the quick math. That's $70 to $210 in pure hardware cost per batch.
Now add the Ethereum posting fee. In a bull market where gas hovers around 20-40 gwei for complex calldata, posting a ZK proof and state diff costs between $50 and $150 per batch.
Total cost per batch: $120 to $360.
Total revenue per batch from user fees: This is the killer. At current L2 fee rates, a batch of 1,000 transactions generates maybe $15 to $40 in revenue. That's a 90% loss on every single batch.
The subsidy gap is the entire L2 business model, and it's not sustainable.
I pulled the on-chain data for one major ZK rollup's batch submissions over the last 90 days. They posted 2,300 batches. At an average loss of $200 per batch, that's $460,000 burned in three months. Just on proving and posting. Not counting team salaries, marketing budgets, or the token incentives they're paying users.
Where does that money come from? The treasury. And what's in the treasury? Your money from the token sale. And what happens when the treasury runs dry?
This isn't a hypothetical. I've seen this play out in the 2020 DeFi summer when yield farms burned through their incentive budgets and collapsed. The mechanics are identical. You're watching a subsidized service that will either need to raise fees dramatically or die when the venture capital taps run out.
Let me put this in perspective with what I learned during the 2017 ICO fire sale. Back then, I shorted utility tokens that had no revenue model. The ones that survived were the ones that actually generated fees. The ones that died were the ones that burned through their war chests on marketing and incentives. Today's ZK rollups are the utility tokens of 2025, and their revenue per unit of work is dangerously close to zero.
Now, here's the counter-intuitive part. The price of ETH matters more than the price of the L2's token. In a bull market, ETH gas prices rise, which means posting costs rise. Proving costs in fiat terms also rise as the ETH price rises, because hardware rental and electricity costs are priced in stablecoins or fiat.
The L2's token could pump 100% on a narrative catalyst, but the underlying business still loses money on every batch. That's not a business. That's a Ponzi-like structure dressed in zero-knowledge clothing.
Yield is the rent you pay for holding someone else's risk. And right now, the yield on your L2 position is the rent you're paying for the privilege of subsidizing someone else's proving costs.
The Contrarian View: What If The Math Changes?
Before you write this off as another bearish L2 hit piece, let me steelman the bull case.
The hardware curve is real. I've been renting GPU clusters since 2021, and the cost per teraflop has dropped roughly 60%. If that trajectory continues, proving costs could drop by another 2-3x in the next 18 months.
There are also algorithmic improvements. The recursive proof aggregation techniques that were theoretical in 2023 are now being implemented. Some of the newer ZK provers claim 5-10x speedups on witness generation. I've tested a few of these on my own infrastructure, and the claims are partially true.
But here's the problem. The revenue side isn't growing at the same rate. User fees are collapsing, not because of efficiency gains, but because of competitive pressure. Every new L2 launches with cheaper fees as the selling point. The market is in a fee war, and nobody wants to be the first to raise prices.
This is the classic adoption trap. You can't build a sustainable business when your primary metric is being cheaper than your competitors, especially when your competitors are also losing money. That's not a market. That's a subsidy war with VCs writing the checks.
My Take: What Happens Next
Bull markets mask structural flaws. I've watched it happen with Terra/Luna in 2022, with the NFT floor sweep craze in 2021, and with the ICO mania in 2017.
The moment the gas price drops or the token price falls below the cost of proving, the music stops. And it stops fast.
We don't need to debate whether ZK tech works. The proofs are valid. The tech is real. What's not real is the economic model attached to it.
The endgame is one of three scenarios:
First, ZK proving costs drop 10x through hardware and algorithmic advances. This is the optimistic scenario. It's possible, but it requires a technical breakthrough, not an incremental improvement. I'm skeptical of teams that promise this timeline because I've been promising cost reductions to my own clients for years. The curve is real, but the slope is not as steep as the marketing suggests.
Second, L2 operators finally raise fees to sustainable levels. This will happen when the treasury pressure becomes unbearable. The moment users see fees rise 5-10x, the narrative shifts from "ETH's scalability solution" to "another expensive chain." The market will likely consolidate to the two or three L2s with the deepest pockets and the most patient investors.
Third, the subsidy model continues until the VCs refuse to fund the next round. Look at the token vesting schedules. Most L2s have heavy unlocks coming in the next 12-18 months. If the price is already under pressure from proving costs and dilution, the incentive to keep the service running drops to zero. We're seeing early signs of this in the smaller ZK projects that have already cut back on user rewards.
My professional bias is toward the second scenario. The market will correct toward sustainability, but only after a painful consolidation phase.
The smart money is positioned for that correction. They're not buying the narrative; they're buying the future fees that will come from consolidation. The question you should be asking isn't whether ZK tech works. It's whether the specific L2 you're holding has the balance sheet to survive the subsidy drought.
I'm running the same models I used during the 2025 AI-agent trading protocol pilot, where I learned that human oversight and strict risk limits are more important than raw execution speed. The same principle applies to L2 investing. The narrative is the execution engine. The economics are the risk limits. If the economics look like a loss on every batch, your risk limit should be zero exposure.
Yield is the rent you pay for holding someone else's risk. And right now, the yield on your L2 position is the rent you're paying for the privilege of subsidizing someone else's proving costs.
Do the math on your own portfolio. Smart money already has.
The question isn't whether ZK rollups can scale Ethereum. The question is whether you can scale your exit before the subsidy runs dry.