The billion-dollar bottleneck keeping your crypto transactions expensive | Opinion

You tried to bridge $50 from Ethereum (ETH) to a ZK-rollup, such as zkSync, expecting to pay a few pennies. Instead, a fee between $0.15 and $0.50 hits your wallet. That’s tough to swallow when you know that optimistic rollups vs ZK-rollups are often three times cheaper, and the Dencun upgrade slashed data costs by over 90%. The bottleneck isn’t network congestion. It’s the proof itself.

Generating the zero-knowledge proof for a transaction batch is an intensive process. It involves trillions of mathematical operations, specifically elliptic curve multiplications, running on hardware not designed for the task. According to data from L2Beat, the proving process alone accounts for a staggering 60-70% of all fees on ZK-L2s.

This has created a centralized market of “prover farms” that is estimated to be worth over $97 million. This is the billion-dollar bottleneck. It’s a centralized, power-hungry system that relies on mismatched hardware, and it’s the reason why ZK scaling still feels more like a promise than a reality. 

The solution is twofold. We need domain-specific ASICs and open prover markets. With these in place, sub-cent transactions can become the default, not the exception. This is not a fantasy. It’s an engineering reality waiting to be embraced.

Where your fees really go

A ZK-rollup transaction has several costs. L2 execution is nearly free. Posting data to the mainnet is also cheap now, thanks to blobs, costing a tenth of a cent per transaction. The real beast is the proving. A single proof for a batch of 4,000 transactions can take anywhere from two to five minutes to generate on a high-end A100 GPU. 

This costs between $0.04 and $0.17 in cloud computing fees alone, based on benchmarks from Brevis. When you amortize that cost, it comes out to about one to four cents per transaction. But under heavy load, it can easily spike to over ten cents.

ZK proving is not the same as the matrix math used in AI. It relies on multi-scalar multiplications (MSMs) and number-theoretic transforms (NTTs) over elliptic curves. These operations are a heavy lift for GPUs, which are designed for parallel floating-point operations. A single Groth16 proof requires around 10¹² field operations.

According to tests by Ingonyama, about 80% of the GPU’s cycles are idle during this process. The hardware just isn’t a good fit. In Q1 2025, zkSync Era’s report showed that proving accounted for 65% of their $2.3 million in revenue. 

That’s $1.5 million that went straight to centralized clusters. Meanwhile, rollup teams pay AWS $1,000 to $5,000 per month per rig, and that cost scales linearly with volume. It’s no wonder that the total value locked in top ZK L2s is stuck at $3.3 billion, while optimistic rollups have surpassed $40 billion.

But optimistic rollups (Base, Arbitrum, Optimism, etc.) are not a complete solution either, as they come with tremendous tradeoffs. Optimistic rollups require at least a seven-day withdrawal delay so that watchful validators can challenge invalid state transitions. In the fast-paced world of DeFi, this wait time is simply too long for most use cases. These networks also require users to trust these watchful validators. While this may be fine for low-value transactions like web3 gaming or social media, it leaves much to be desired when it comes to real-world assets and other high-powered financial use-cases.

Therefore, the path to truly scalable, cheap, and secure transactions lies not in avoiding proof generation but in revolutionizing it.

This is the core inefficiency: we pay a web2 premium, in both cost and centralization, to generate trust for a decentralized network. The proving fee isn’t just an expense; it’s the economic anchor holding back ZK adoption. Until proof generation becomes as cheap as execution, ZK-rollups will remain trapped in a system where their greatest strength, cryptographic security, is also their most costly bottleneck. And that cost doesn’t just inflate fees; it builds the centralized trap we must now escape.

The centralization trap

Over 90% of ZK-rollups outsource to a handful of “prover-as-a-service” companies. This creates a massive single point of failure. In March 2025, Blast operators froze attacker transactions and 12,000 innocent users for 48 hours. Centralized provers see plaintext batches, extract MEV, and refuse transactions. dYdX uses a 3-of-5 multisig, meaning three insiders could hijack the chain. This mirrors hardware dependency issues in other blockchain systems, like Solana’s consensus mechanism.

This centralization creates fragility. Top provers have 99.2% uptime, but a single AWS outage can cascade. In Q2 2025, Starknet lost 20% throughput due to one node failure. These providers capture 80% margins, leaving rollups rent-strapped, while Ethereum’s L1 provides security. It’s an architectural contradiction. We have “decentralized” L2s dependent on Web2 data centers. As Vitalik Buterin said at Devcon 2024, “Outsource proving, and you’ve scaled nothing, just trusted a new oracle.”

The fix is in

To solve this problem, we need a one-two punch. We need silicon that is built for ZK, and we need a marketplace to distribute it. First, the hardware. GPUs are like race cars on a highway. They are great for the tensor operations used in AI, but they are terrible for the curve operations used in ZK. This is where domain-specific ASICs come in.

The winners of the 2025 ZPrize competition were able to clock STARK proofs at 6-8 seconds on FPGAs. That’s 10-100 times faster per watt than a GPU, because they hardwired the MSMs and NTTs. Cysic is already taping out ASICs that can generate proofs for billion-gate circuits in under two seconds, and they do it with a 50x reduction in energy consumption. This is the same evolution that we saw with Bitcoin (BTC) mining. We went from CPUs to ASICs, but this time it’s for truth, not for hashes.

Second, the markets. We need to ditch the SaaS model and treat provers as a commodity. Succinct’s 2025 mainnet exemplifies this. Rollups post jobs to Ethereum-based auctions, and specialized rigs bid on them in seconds. Brevis ProverNet achieved 20x CPU speeds via GPU clusters with slashing for lazy nodes.

The result is 40% cost reduction, equitable value distribution, and censorship resistance. Anyone can prove, Ethereum verifies. ZPrize 2025 recursive aggregation verified proofs in under 200 milliseconds on-chain. A $10,000 DeFi bot can run private ZK-ML on an L2 for a thousandth of a cent per transaction.

Unlocking the floodgates

Cheaper proving will lead to predictable, sub-$0.01 fees at 15,000 TPS, per the zkSync Atlas roadmap. This unlocks pay-per-pixel NFTs, real-time gaming economies, and AI agents settling trades without frontrunning. L2s become neutral infrastructure, not prover fiefdoms. 

Developers focus on VM optimization instead of managing massive server farms. This sparks tenfold app innovation. ZK TVL grows from $28 billion in 2025 to over $100 billion. The real opportunity is systemic. 

When proving becomes cheap and decentralized, the entire value proposition crystallizes. Users get speed and cost. Developers get a platform that doesn’t require trusting a handful of operators. Ethereum gets a scaling solution that maintains security guarantees without compromise.

The fight is here

The biggest barrier to ZK adoption isn’t crypto or gaming, it’s engineering economics. Centralized infrastructure is choking the scalable truth. The war isn’t over a consensus. It’s in hardware and markets for verifiable compute. Ship decentralized, accelerated provers, and we deliver cheap, trustless transactions for billions of people. 

We don’t need prover overlords. We need a global bazaar for proofs, where anyone contributes, everyone verifies, and the open web is secure.