AI is crowding out ZK proving — and Cysic says the solution is to stop treating GPUs as the only option. Why GPUs have become a bottleneck - Cysic founder and CEO Leo Fan warns that zero-knowledge (ZK) proving systems now compete directly with trillion-dollar AI data‑centers for the same Nvidia silicon. “AI models are converging. Compute isn’t,” Fan told crypto.news. Instead of cheaper chips driving down proving costs, proving teams are bidding against massive AI budgets for identical hardware. - The squeeze isn’t just about raw GPU supply. Cysic finds an architectural mismatch between zkVM software and the accelerators used to generate proofs. That mismatch leaves GPU capacity idle and raises the cost per proof. How Cysic exposed the problem - Cysic’s Venus proving engine — a hardware-focused extension of Polygon Hermez’s ZisK zkVM — models proof generation as a single connected computation graph. That lets the prover schedule work across the entire pipeline instead of treating each hardware call in isolation. - By integrating CUDA Graphs, doing kernel tuning, and changing shared-memory handling, Venus reduces repeated CPU–GPU transfers and synchronization. The company reported an end‑to‑end proof-time improvement of more than 9% versus ZisK 0.16.1 without swapping hardware — evidence that existing accelerators were not being fully utilized. Performance context and industry claims - Cysic says ZisK can produce an Ethereum block proof in 7.4 seconds using 24 GPUs and can submit real‑time proofs from a single RTX 4090. These vendor claims have not been benchmarked independently across energy use, proof size, security level and total cost. - Other teams have reported similar breakthroughs: Succinct in November 2025 said SP1 Hypercube proved 99.7% of a 954‑block Ethereum sample in under 12 seconds using 16 Nvidia RTX 5090 GPUs, with about 95.4% proven within 10 seconds. - The Ethereum Foundation’s real‑time proving definition: complete proofs for at least 99% of mainnet blocks within 10 seconds, plus requirements for open-source code, proof sizes below 300 KiB, ≥128‑bit security, hardware under $100,000 and power below 10 kW. The Foundation also warns that energy and cooling may be a greater limit than equipment cost for independent provers. Why GPUs aren’t ideal for ZK workloads - AI workloads emphasize massive matrix math; ZK proving relies heavily on multi-scalar multiplications and number‑theoretic transforms. GPUs can run both, but they weren’t designed specifically for the latter. Because GPU‑hours drive much of proving costs, competition with AI raises hardware and rental prices that pass directly into per‑proof cost. - Real‑time layer‑1 proving is especially vulnerable: a fresh proof is needed every block, so operators must keep spare capacity to avoid missing deadlines. That spare capacity multiplies the cost pressure across ZK rollups and proof marketplaces. One analysis cited by Cysic estimated proof generation accounted for 60–70% of fees on some ZK layer‑2s. Macro demand for accelerators - Nvidia’s data illustrates the scale of the AI pull: $75.2 billion in data‑center revenue for the quarter ending April 26 (up 92% year over year), with data‑center compute revenue of $60.4 billion (up 77%). - Amazon Web Services has reportedly committed to buying one million Nvidia GPUs, deliveries continuing through 2027; Nvidia executives and bond financing plans underscore massive ongoing investment in AI infrastructure. Even some crypto infrastructure owners are redirecting power and facilities toward AI workloads. Alternative hardware paths: FPGA and ASIC - Cysic is expanding beyond GPUs. Venus’ public repository includes GPU optimizations, a full FPGA backend, and an early ASIC‑oriented implementation. - The FPGA kernels cover Goldilocks field arithmetic, NTTs, Poseidon2, Merkle trees, FRI and expression evaluation, targeting AMD UltraScale+ and Versal devices with high‑bandwidth memory. The code is open under Apache 2.0 and MIT licenses. - FPGAs can be reprogrammed after production, enabling updates and experimentation. Custom ASICs are less flexible but can offer much better performance and energy efficiency for a stable set of ZK operations. Both approaches would free provers from the main AI hardware queue and avoid paying for GPU functionality ZK workloads don’t need — but design and production costs and limited volumes are real hurdles. A broader strategy and risks - Beyond hardware, Cysic proposes a global prover marketplace where devices from phones to clusters accept proving jobs, combining GPU and FPGA backends to reduce dependence on a single chip class. Cryptographic verification means verifiers reject invalid proofs regardless of who produced them, so opening participation doesn’t change soundness — but it raises exposure to implementation bugs in unaudited code. - Cysic notes Venus is under active development and recommends audits and redundant multi‑prover configurations to limit implementation risk. Operational economics: examples and exposed apps - An earlier analysis cited by Cysic estimated that proving a batch of 4,000 transactions could take 2–5 minutes on an Nvidia A100 and cost $0.04–$0.17 in cloud charges, depending on many variables. - Fan singled out zkML as particularly exposed because it combines AI workloads with proving overhead. For consumer use cases — private payments, on‑chain games — proof costs often must be in pennies to be viable. “Can cost limit adoption? Yes at the margin,” Fan said, noting that lower‑value applications would be deferred first if economics break down. - Fan also said more than 90% of ZK layer‑2 networks rely on a small group of prover services — his figure is Cysic’s estimate and would benefit from named datasets. Protocol-level developments - Draft EIP‑8025 would let Ethereum validators opt into generating or verifying execution proofs while keeping conventional block re‑execution as a fallback. It proposes a proof gossip channel and external proof nodes but currently lacks explicit incentives for operators that produce and broadcast proofs. Bottom line ZK proving is entering a new era where GPU scarcity and an architectural mismatch are driving costs up and forcing innovation. Cysic’s Venus shows that software and scheduling optimizations can deliver measurable gains today; broader adoption may hinge on diversified hardware (FPGAs, ASICs), open toolchains, careful auditing and economic designs that make proving affordable without surrendering decentralization. Read more AI-generated news on: undefined/news