ZK proving must move beyond GPUs as AI tightens compute supply, Cysic CEO says

ZK proving has begun competing with trillion-dollar AI data centers for the same GPUs, raising proof costs even as Cysic reports a 9% performance gain from improving hardware use.

Cysic founder and CEO Leo Fan told crypto.news that GPU use has become a binding constraint for zero-knowledge proving because proof systems now compete with heavily funded AI data centers for the same silicon.

“AI models are converging. Compute isn’t. Everyone assumed proving costs would fall because chips get cheaper. Instead, we’re bidding against trillion-dollar data centre budgets for the same silicon. That’s why the hardware layer had to be opened up rather than left to a handful of proprietary provers.”

The pressure does not come from a lack of computing capacity alone, according to Fan. He said the main problem is an architectural mismatch between zkVM software and the accelerators used to generate proofs, which leaves part of the available GPU capacity unused and raises the cost of each proof.

Cysic’s Venus proving engine exposed that mismatch by reducing the time spent coordinating work between CPUs and GPUs. As reported in April, the company recorded an end-to-end proof-time improvement of more than 9% against ZisK 0.16.1 without replacing the underlying hardware.

ZK proving costs now matter more than raw speed

Built as a hardware-focused extension of Polygon Hermez’s ZisK zkVM, Venus represents proof generation as one connected computation graph. Cysic says the design lets the system schedule work across the full proving process instead of handling each hardware function as a separate call.

Through CUDA Graph integration, kernel tuning, and shared-memory changes, Venus reduces repeated data transfers and synchronization between the processor and GPU. Fan said the result shows that existing accelerators were not being fully used, making utilization the practical bottleneck behind proof costs.

Raw proving speed has improved quickly across the industry. Cysic has said ZisK can generate an Ethereum block proof in 7.4 seconds with 24 GPUs and can submit real-time proofs through a single RTX 4090 setup. The claims come from the company and have not been independently tested under a common benchmark covering energy use, proof size, security level, and total hardware cost.

Other developers have also crossed Ethereum’s real-time threshold. In November 2025, Succinct reported that SP1 Hypercube proved 99.7% of a 954-block Ethereum sample in less than 12 seconds using 16 Nvidia RTX 5090 GPUs. About 95.4% of the sample was proven within 10 seconds.

The Ethereum Foundation defines real-time proving as completing proofs for at least 99% of mainnet blocks within 10 seconds. Its framework also calls for fully open-source code, proof sizes below 300 KiB, at least 128-bit security, hardware costing no more than $100,000, and power use capped at 10 kilowatts.

Energy use may be a more serious limit than equipment cost for home provers, the Foundation said. A proof can arrive before Ethereum’s deadline while still requiring too much power, cooling or capital for an independent operator.

AI demand is tightening access to the same GPUs

Although AI and ZK proving use GPUs differently, both workloads depend on Nvidia accelerators ranging from consumer RTX cards to data-center H100 systems.

AI training and inference mainly use matrix calculations. ZK proof generation relies heavily on multi-scalar multiplications and number-theoretic transforms, operations that GPUs can process but were not specifically designed to run.

Fan said the mismatch supports the case for ZK-specific hardware because proof systems are entering the same supply queue as AI developers without using the chips in the same way. Since GPU-hours account for much of the proving bill, higher hardware and rental costs pass into the cost of each proof, he added.

Nvidia’s financial results show the scale of demand coming from AI infrastructure. The U.S.-listed chipmaker reported $75.2 billion in data-center revenue for the quarter ending April 26, up 92% from a year earlier. Data-center compute revenue reached $60.4 billion, an annual increase of 77%.

Amazon Web Services has separately agreed to buy one million Nvidia GPUs, with deliveries scheduled to continue through 2027, Reuters reported in March. Nvidia CEO Jensen Huang has estimated a $1 trillion sales opportunity for the Blackwell and Rubin product families through that year.

In June, Nvidia’s planned bond sale sought at least $20 billion to fund AI investments and refinance debt. Bitcoin mining companies had announced more than $70 billion in AI and high-performance computing contracts at the time, illustrating how crypto-linked infrastructure owners are also redirecting power and facilities toward AI workloads.

A Bernstein report covered in May placed announced AI infrastructure partnerships at nearly $90 billion. The analysts estimated that Bitcoin miners controlled more than 27 gigawatts of planned power capacity, compared with about 3.7 gigawatts tied to announced AI agreements, while some U.S. grid connections could take up to 50 months.

ZK-rollups and real-time provers face the pressure first

Real-time layer-1 proving sits at the front of the cost squeeze because it requires GPUs to produce a fresh proof for every block, Fan said. Any delay can cause a prover to miss the network’s time limit, so operators need spare capacity as well as enough hardware for normal demand.

ZK-rollups and proof marketplaces follow because GPU-hours feed directly into operating expenses and, in some cases, user fees. An earlier proving cost analysis estimated that proof generation accounted for 60% to 70% of fees on ZK layer-2 networks, citing L2Beat data.

According to the same analysis, generating a proof for a batch of 4,000 transactions could take two to five minutes on an Nvidia A100 and cost between $0.04 and $0.17 in cloud computing charges. The figures depend on the proof system, transaction batch, hardware configuration, and cloud rate.

Fan placed zkML among the most exposed applications because it combines an AI workload with the added expense of proving that the model ran correctly. For private payments, on-chain games, and other consumer products, he said the economics often require proof costs measured in pennies.

“Can cost limit adoption? Yes at the margin,” Fan said, adding that private payments and gaming would likely be deferred first when their economics no longer work.

Cost pressure could also affect how many entities can operate provers. Fan said more than 90% of ZK layer-2 networks rely on a small group of prover services, although the estimate requires a named dataset and should be treated as Cysic’s assessment.

FPGAs and ZK ASICs offer an alternative hardware path

Cysic has responded by developing multiple backends rather than relying only on GPUs. The public Venus repository includes GPU optimizations, a complete FPGA acceleration backend, and an early ASIC-oriented implementation.

Its FPGA backend contains kernels for Goldilocks field arithmetic, NTTs, Poseidon2, Merkle trees, FRI and expression evaluation. The code targets AMD UltraScale+ and Versal devices with high-bandwidth memory and is available under Apache 2.0 and MIT licences.

Unlike an ASIC, an FPGA can be reprogrammed after production, allowing developers to update circuits and experiment with new proving systems. Custom ASICs offer less flexibility but can deliver better performance and energy efficiency when designed for a stable set of ZK operations.

Fan said moving to FPGAs and ZK ASICs would remove proof operators from the main AI hardware queue. Specialized devices would also avoid paying for GPU functions that ZK workloads do not need, although development costs and limited manufacturing volumes remain obstacles.

Opening the software is one part of Cysic’s approach. The company also proposes a global prover marketplace in which devices ranging from mobile hardware to professional clusters can accept jobs, with GPU and FPGA backends reducing dependence on one chip class.

Cryptographic verification means a verifier rejects an invalid proof regardless of which operator generated it, Fan said. Opening participation, therefore, does not change the proof system’s soundness, but it increases exposure to implementation errors in unaudited or unfinished code.

Cysic states in the Venus repository that the project remains under active development. Fan said audits and redundant multi-prover configurations would be needed to limit implementation risk.

Draft EIP-8025 would let Ethereum validators opt into generating or verifying execution proofs while conventional block re-execution remains in place. The proposal introduces a proof gossip channel and external proof nodes, but its current version does not provide incentives for operators that generate and broadcast the proofs.