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

ZK proving has begun competing with trillion-dollar AI knowledge facilities for a similar GPUs, elevating proof prices whilst Cysic reviews a 9% efficiency achieve from enhancing {hardware} use.

Cysic founder and CEO Leo Fan informed crypto.information that GPU use has change into a binding constraint for zero-knowledge proving as a result of proof programs now compete with closely funded AI knowledge facilities for a similar silicon.

“AI fashions are converging. Compute isn’t. Everybody assumed proving prices would fall as a result of chips get cheaper. As an alternative, we’re bidding towards trillion-dollar knowledge centre budgets for a similar silicon. That’s why the {hardware} layer needed to be opened up somewhat than left to a handful of proprietary provers.”

The strain doesn’t come from a scarcity of computing capability alone, in line with Fan. He mentioned the principle downside is an architectural mismatch between zkVM software program and the accelerators used to generate proofs, which leaves a part of the accessible GPU capability unused and raises the price of every proof.

Cysic’s Venus proving engine uncovered that mismatch by lowering the time spent coordinating work between CPUs and GPUs. As reported in April, the corporate recorded an end-to-end proof-time enchancment of greater than 9% towards ZisK 0.16.1 with out changing the underlying {hardware}.

ZK proving prices now matter greater than uncooked pace

Constructed as a hardware-focused extension of Polygon Hermez’s ZisK zkVM, Venus represents proof technology as one related computation graph. Cysic says the design lets the system schedule work throughout the total proving course of as a substitute of dealing with every {hardware} operate as a separate name.

By way of CUDA Graph integration, kernel tuning, and shared-memory adjustments, Venus reduces repeated knowledge transfers and synchronization between the processor and GPU. Fan mentioned the consequence exhibits that present accelerators weren’t being absolutely used, making utilization the sensible bottleneck behind proof prices.

Uncooked proving pace has improved rapidly throughout the trade. Cysic has mentioned ZisK can generate an Ethereum block proof in 7.4 seconds with 24 GPUs and might submit real-time proofs by a single RTX 4090 setup. The claims come from the corporate and haven’t been independently examined underneath a typical benchmark masking power use, proof measurement, safety stage, and complete {hardware} price.

Different builders have additionally crossed Ethereum’s real-time threshold. In November 2025, Succinct reported that SP1 Hypercube proved 99.7% of a 954-block Ethereum pattern in lower than 12 seconds utilizing 16 Nvidia RTX 5090 GPUs. About 95.4% of the pattern was confirmed inside 10 seconds.

The Ethereum Basis defines real-time proving as finishing proofs for a minimum of 99% of mainnet blocks inside 10 seconds. Its framework additionally requires absolutely open-source code, proof sizes under 300 KiB, a minimum of 128-bit safety, {hardware} costing not more than $100,000, and energy use capped at 10 kilowatts.

Power use could also be a extra severe restrict than gear price for residence provers, the Basis mentioned. A proof can arrive earlier than Ethereum’s deadline whereas nonetheless requiring an excessive amount of energy, cooling or capital for an impartial operator.

AI demand is tightening entry to the identical GPUs

Though AI and ZK proving use GPUs in another way, each workloads rely upon Nvidia accelerators starting from shopper RTX playing cards to focus on=”_blank”>Nvidia’s deliberate bond sale sought a minimum of $20 billion to fund AI investments and refinance debt. Bitcoin mining corporations had introduced greater than $70 billion in AI and high-performance computing contracts on the time, illustrating how crypto-linked infrastructure homeowners are additionally redirecting energy and amenities towards AI workloads.

A Bernstein report lined in Might positioned introduced AI infrastructure partnerships at practically $90 billion. The analysts estimated that Bitcoin miners managed greater than 27 gigawatts of deliberate energy capability, in contrast with about 3.7 gigawatts tied to introduced AI agreements, whereas some U.S. grid connections may take as much as 50 months.

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

Actual-time layer-1 proving sits on the entrance of the fee squeeze as a result of it requires GPUs to provide a recent proof for each block, Fan mentioned. Any delay could cause a prover to overlook the community’s time restrict, so operators want spare capability in addition to sufficient {hardware} for regular demand.

ZK-rollups and proof marketplaces observe as a result of GPU-hours feed instantly into working bills and, in some circumstances, person charges. An earlier proving price evaluation estimated that proof technology accounted for 60% to 70% of charges on ZK layer-2 networks, citing L2Beat knowledge.

Based on the identical evaluation, producing a proof for a batch of 4,000 transactions may take two to 5 minutes on an Nvidia A100 and value between $0.04 and $0.17 in cloud computing costs. The figures rely upon the proof system, transaction batch, {hardware} configuration, and cloud charge.

Fan positioned zkML among the many most uncovered purposes as a result of it combines an AI workload with the added expense of proving that the mannequin ran accurately. For personal funds, on-chain video games, and different shopper merchandise, he mentioned the economics usually require proof prices measured in pennies.

“Can price restrict adoption? Sure on the margin,” Fan mentioned, including that non-public funds and gaming would doubtless be deferred first when their economics now not work.

Value strain may additionally have an effect on what number of entities can function provers. Fan mentioned greater than 90% of ZK layer-2 networks depend on a small group of prover providers, though the estimate requires a named dataset and needs to be handled as Cysic’s evaluation.

FPGAs and ZK ASICs supply an alternate {hardware} path

Cysic has responded by growing a number of backends somewhat than relying solely on GPUs. The general public Venus repository contains GPU optimizations, a whole FPGA acceleration backend, and an early ASIC-oriented implementation.

Its FPGA backend accommodates kernels for Goldilocks area arithmetic, NTTs, Poseidon2, Merkle timber, FRI and expression analysis. The code targets AMD UltraScale+ and Versal units with high-bandwidth reminiscence and is out there underneath Apache 2.0 and MIT licences.

Not like an ASIC, an FPGA may be reprogrammed after manufacturing, permitting builders to replace circuits and experiment with new proving programs. Customized ASICs supply much less flexibility however can ship higher efficiency and power effectivity when designed for a secure set of ZK operations.

Fan mentioned shifting to FPGAs and ZK ASICs would take away proof operators from the principle AI {hardware} queue. Specialised units would additionally keep away from paying for GPU capabilities that ZK workloads don’t want, though growth prices and restricted manufacturing volumes stay obstacles.

Opening the software program is one a part of Cysic’s strategy. The corporate additionally proposes a world prover market through which units starting from cell {hardware} to skilled clusters can settle for jobs, with GPU and FPGA backends lowering dependence on one chip class.

Cryptographic verification means a verifier rejects an invalid proof no matter which operator generated it, Fan mentioned. Opening participation, due to this fact, doesn’t change the proof system’s soundness, but it surely will increase publicity to implementation errors in unaudited or unfinished code.

Cysic states within the Venus repository that the mission stays underneath lively growth. Fan mentioned audits and redundant multi-prover configurations can be wanted to restrict implementation threat.

Draft EIP-8025 would let Ethereum validators decide into producing or verifying execution proofs whereas typical block re-execution stays in place. The proposal introduces a proof gossip channel and exterior proof nodes, however its present model doesn’t present incentives for operators that generate and broadcast the proofs.