ZKsync’s Airbender Prime prover generated proofs for 7,200 consecutive Ethereum mainnet blocks on a single graphics card, finishing 99% of them in under ten seconds. That is a real milestone. Real-time proving, long treated as the finish line for zero-knowledge rollups, is now achievable on hardware an individual can realistically run. The announcement also makes clear what the milestone is not: a fast prover is not a correct prover, and the next hard problem in zero-knowledge Ethereum is proving the prover itself.
The numbers ZKsync published on October 13 are specific. One NVIDIA RTX PRO 6000 Blackwell GPU proved 99% of a consecutive batch of 7,200 mainnet blocks in under 10 seconds each, including witness preparation, and finished every block within 12 seconds. Airbender Prime is built on the open RISC-V instruction set and uses hash-based STARKs that combine GKR sumcheck, WHIR polynomial commitments and BLAKE2s hashing; its layer-1 proofs move to Keccak-256 commitments instead of BN254 pairings. It ships with a standalone verifier, so a third party can check a proof without trusting ZKsync’s own numbers. ZKsync describes it, to the team’s knowledge, as the first Ethereum prover to demonstrate real-time proving on a single GPU.
A throughput benchmark is exactly the kind of problem that is easy to measure and, in relative terms, easy to optimize. Speed has numbers attached: blocks per second, seconds per block, GPUs required. Correctness has no such dashboard. A prover can be fast on a friendly workload and still diverge from Ethereum’s rules on some input, and no timing chart will reveal that divergence. That is why Airbender Prime matters, and why it should be read as a starting point rather than a conclusion.
Ethereum co-founder Vitalik Buterin framed the shift in an October 15 post, describing the zkEVM “battle for speed” as “definitively won” and arguing the priority should now “shift fully to security.” He then set out three conditions for what formally verifying a prover means. First, prove that the arithmetization, the translation of Ethereum’s rules into the constraints a proof system checks, is equivalent to the Ethereum specification. Second, set a hard upper bound on prover cost relative to the gas in a block. Third, verify the security of the proof system itself. None of these produces a headline benchmark, and that is the point.
The three conditions share a property: they are mathematical and slow. Matching a proof system’s constraints to a living specification means auditing every opcode, every edge case and every hard fork. Bounding prover cost is an adversarial exercise that assumes the worst block rather than the average one. Verifying the proof system means scrutinizing cryptographic assumptions that are far harder to check than a timing chart. Teams that do this work first ship security fixes with no headline benchmark attached, which is why the shift is easy to miss from the outside.
For readers evaluating zero-knowledge rollups, the practical step is to separate the two races. Ask what a prover has demonstrated on throughput, then ask what has been independently verified about its correctness. Quantum-resistance claims deserve the same treatment: a hash-based proof system is only post-quantum-ready if its cryptographic assumptions actually hold, which is something to verify rather than repeat. The next observable milestones are not faster benchmarks but formal-verification results, published audits and independent reproduction of security arguments. Speed is no longer the differentiator. Provable correctness is.
Source: BlockchainReporter.