GPU shortages driven by AI are squeezing zero-knowledge proving — and one prover startup says the answer is to move beyond GPUs. Cysic founder and CEO Leo Fan told crypto.news that ZK proof systems are now directly competing with trillion-dollar AI data-center budgets for the same Nvidia silicon, pushing up proof costs and creating a binding constraint on ZK throughput. “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,” Fan said. He argued that the hardware layer for proving must be opened up rather than left to a handful of proprietary provers. Architectural mismatch, not just capacity Fan emphasized that the pressure is not only from raw GPU demand but from an architectural mismatch between zkVM software and the accelerators used to generate proofs. ZK workloads use operations such as multi‑scalar multiplications and number‑theoretic transforms — not the matrix-heavy workloads GPUs were optimized for — and current prover stacks leave part of GPU compute idle, raising the cost per proof. Cysic’s Venus: better utilization, same hardware Cysic’s Venus proving engine — a hardware-focused extension of Polygon Hermez’s ZisK zkVM — exposed that mismatch by treating proof generation as a single connected computation graph and scheduling work across the whole proving pipeline. By reducing CPU/GPU coordination overhead through CUDA Graph integration, kernel tuning and shared‑memory improvements, Venus reduced repeated data transfers and synchronization, delivering a reported end‑to‑end proof‑time gain of more than 9% over ZisK 0.16.1 without changing the underlying hardware. Industry gains, but independent benchmarking is limited Proof speeds have been improving fast industry‑wide. Cysic says ZisK can produce an Ethereum block proof in 7.4 seconds using 24 GPUs and that real‑time proofs are possible on a single RTX 4090; those claims are from the company and have not been independently benchmarked against standardized metrics (energy, proof size, security level, total hardware cost). Other teams have crossed Ethereum’s “real‑time” threshold too: in November 2025 Succinct reported its SP1 Hypercube proved 99.7% of a 954‑block Ethereum sample in under 12 seconds using 16 Nvidia RTX 5090 GPUs (about 95.4% of the sample was proven within 10 seconds). What “real‑time” means 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, total hardware cost no greater than $100,000, and power use capped at 10 kW. The Foundation notes energy and cooling can be a bigger barrier than hardware cost for independent/home provers: a proof can meet the time deadline but still require too much power or capital to be practical for many operators. AI demand is enormous and growing The reason ZK systems are competing so fiercely for GPUs is plain: AI demand has exploded. Nvidia reported $75.2 billion in data‑center revenue for the quarter ending April 26 (up 92% year‑over‑year), including $60.4 billion in data‑center compute revenue (up 77%). AWS has agreed to buy one million Nvidia GPUs with deliveries through 2027. Nvidia CEO Jensen Huang has estimated a $1 trillion sales opportunity for its Blackwell and Rubin product families through that period. Meanwhile, Bitcoin miners and other crypto infrastructure owners have also announced large AI and HPC contracts, illustrating how traditionally crypto‑oriented power and facilities are being redirected toward AI. Why layer‑1 real‑time proving is most exposed Real‑time layer‑1 proving is particularly vulnerable because it requires a fresh proof for every block; any delay risks missing Ethereum’s time limit, so operators must maintain spare capacity beyond normal demand. That dynamic pushes provers into higher‑cost hardware or rented GPU hours. An analysis cited by L2Beat estimated proof generation accounts for 60–70% of fees on ZK layer‑2 networks. Example cloud figures: a proof for a batch of 4,000 transactions could take two to five minutes on an Nvidia A100 and cost roughly $0.04–$0.17, depending on configuration. Fan singled out zkML as especially exposed because it combines AI compute with the added expense of cryptographic proofing. Concentration risk and a hardware diversification strategy Fan warned that operator concentration is already high — Cysic estimates more than 90% of ZK layer‑2 networks rely on a small set of prover services — and that GPU competition could further concentrate proving infrastructure. Cysic’s response is to diversify hardware rather than rely on GPUs alone. The public Venus repository includes: - GPU optimizations and the connected computation‑graph design, - a full FPGA acceleration backend, and - an early ASIC‑oriented implementation. The FPGA backend provides kernels for Goldilocks field arithmetic, NTTs, Poseidon2, Merkle trees, FRI and expression evaluation, targeting AMD UltraScale+ and Versal devices with high‑bandwidth memory. That code is available under Apache 2.0 and MIT licences. FPGAs vs ASICs: tradeoffs FPGAs are reprogrammable, letting developers iterate on circuits and new proving systems — useful while ZK algorithms keep evolving. ASICs are less flexible but can deliver much better performance and energy efficiency for a stable set of ZK operations. Both approaches would take ZK provers out of the main AI GPU queue and avoid paying for GPU features ZK workloads don’t need, though development costs and limited manufacturing volumes remain meaningful hurdles. Marketplace and openness — with caveats Cysic also proposes a global prover marketplace in which devices from mobile hardware to professional clusters accept jobs, with multiple backends (GPU, FPGA, ASIC) reducing dependence on any single chip class. Cryptographic verification ensures that verifiers will reject incorrect proofs regardless of which operator produced them, so opening participation does not change proof soundness — but it does increase exposure to implementation bugs in unaudited or unfinished code. Cysic notes Venus is under active development and calls for audits and redundant multi‑prover setups to limit implementation risk. Protocol developments Draft EIP‑8025 would let Ethereum validators opt into generating or verifying execution proofs while conventional block re‑execution remains available. The draft adds a proof gossip channel and external proof nodes, but in its current form it does not include incentives for operators that generate and broadcast proofs. Bottom line As AI soaks up GPU supply, ZK proving is hitting a new cost frontier. Cysic’s experience suggests much of the problem is inefficient use of existing accelerators, not just raw scarcity — and that a hardware‑diversified strategy (FPGAs, ASICs, plus smarter scheduling on GPUs) plus open marketplaces could be the path to cheaper, more resilient proving infrastructure. The tradeoffs include development expense, manufacturing limits, and increased need for audits and redundancy to keep systems secure. Read more AI-generated news on: undefined/news