AI's competition for GPU is pushing the zero-knowledge proof hardware problem to the stage. Cysic founder and CEO Leo Fan states that ZK has proved that today it is competing with the same chips as the large AI data centre, and that costs are no longer dependent solely on algorithm optimization, but are increasingly subject to calculator supply and hardware prices.

Prove that costs are starting to rise.

According to Fan, the current pressure is not only not enough for GPU, but also because the zkVM software does not fully match the existing accelerator, resulting in part of the GPU resources not being fully utilized and a single demonstration of cost increases.

Cysic states that by reducing the cost of coordination between CPU and GPU, its Venus certification engine increased the end-to-end certification time by more than 9 per cent than ZisK 0.16.1 without replacing the bottom hardware. The system is based on the Zisk zkVM extension of Polygon Hermez, which attempts to move the entire certification process as an integrated computational map rather than breaking into multiple independent hardware calls.

The rate of certification within the industry has increased rapidly over the past two years. Cysic had previously stated that ZisK could use 24 GPUs to produce a certificate in 7.4 seconds with a RTX 4090 block and could also submit a real-time certificate. However, these data are mainly from corporate disclosures and have not yet been independently validated under the harmonized benchmarks.

It's a higher threshold.

The term “real-time certification” is defined by the ETA Foundation as at least 99 per cent of the main network blocks are to be certified within 10 seconds. In addition to speed, the code is required to be fully open, certified to be less than 300 KiB, a security level of at least 128 bits and a hardware cost of not more than $100,000, and a power consumption of not more than 10 kW.

The Foundation also noted that energy consumption may be more difficult for independent certifying officers than the cost of equipment procurement. Even if it is proved that it can be done within the time frame, it is difficult to support a more decentralized network of proof if it is power-consuming, heat-dissemination or when capital investment is too high.

Fan states that real time Layer1 proves that cost pressure is felt first, because each block needs to generate new proofs in a timely manner, and the operator needs not only sufficient hardware, but also redundancy. ZK Rollup and the certification market will then be under pressure, as GPU use time limits will enter operating costs directly and, in some cases, transfer to user costs.

FPGA and ASIC as an alternative

According to the article, AI training and reasoning rely mainly on matrix calculations, while ZK has demonstrated a greater reliance on multi-scale multipliers and numerical variations. While GPU can handle such tasks, it is not specifically designed for this purpose. This means that while ZK projects compete with AI chips, they do not fully take advantage of the universal GPU structure.

According to the British financial report, AI infrastructure needs are still growing rapidly. The company ' s data centre revenue as of the quarter of 26 April was $75.2 billion, an increase of 92 per cent over the same period; of that amount, the data centre calculated income was $60.4 billion, an increase of 77 per cent over the same period. Reuters had also previously reported that AWS had agreed to purchase 1 million British Weida GPUs and that delivery would continue until 2027.

Against this background, Cysic is advancing a multi-back-end solution rather than relying on GPU alone. Its open Venus repository already contains GPU optimization, complete FPGA acceleration backend and early realization for ASIC. Fan believes that a shift to FPGA and ZK ASIC would allow the operator to prove that it is less dependent on the AI hardware supply chain and avoid paying for GPU functions that are not required for the ZK load.

However, there are still realistic constraints on this route. While FPGA can be reprogrammed after being out of the plant and suitable for rapid iterativeity, ASIC is less flexible, it may result in higher performance and lower energy consumption in fixed ZK operations. Both face high development costs and limited manufacturing.

Additional information:Cysic also proposed the establishment of a global certification market that would allow mobile devices to be placed on a professional cluster. Fan indicated that the password certification itself would not change validity by proving who generated it, but that opening up more participants would increase the risk of errors without an audit or mature implementation code.