A conversation with Leo Fan, Founder at Cysic
“We are transforming computing resources from something scarce and controlled by a few into a shared economy owned by the community.”
Leo Fan is the co-founder of Cysic Network, a full-stack compute infrastructure project focused on zero-knowledge acceleration, decentralized GPU networks, and what the team calls ComputeFi.
Leo entered crypto through academia. He began his PhD in cryptography at Cornell University in 2015 and later worked at the National Institute of Standards and Technology, contributing to post-quantum cryptography standardization. He also helped develop a ZK-based cross-chain bridge at Algorand before becoming an Assistant Professor at Rutgers University, where he teaches cryptography and computer security.
Cysic was inspired by Leo’s background in Bitcoin mining and ZK research, combining specialized hardware with optimized software to power verifiable computation at scale.
Leo explained that Cysic originally started as a hardware acceleration company focused purely on speeding up zero-knowledge proof generation. While developing a cross-chain bridge at Algorand, he encountered major latency challenges in proof generation, even when using large cloud instances. Drawing from his early experience in Bitcoin mining, he recognized that specialized hardware could dramatically improve performance.
However, as customers began requesting end-to-end proof generation rather than just hardware optimization, Cysic expanded into a distributed prover marketplace. By tapping into idle GPU resources from crypto mining communities, the company transitioned into a broader ComputeFi model that combines hardware, software, and decentralized participation.
Leo emphasized that as AI systems become more powerful, verification becomes increasingly important. If a user pays for access to a premium AI model, there is currently no reliable way to verify that the provider actually used the promised model architecture during inference. Since humans cannot easily distinguish between outputs generated by different model versions, this creates an accountability gap.
Zero-knowledge proofs can serve as cryptographic certificates, proving that a specific model was used and that the output corresponds correctly to the input. Although current GPU-based proof generation for large models can take more than ten seconds, which is impractical for real-time use, ASIC-level acceleration could reduce that time to under one second. This would make verifiable AI both practical and scalable.
Leo described the mainnet as a transparency and settlement layer for computation. Every step in the proving lifecycle is recorded on-chain, including prover selection, proof generation, verifier selection, and final settlement. This ensures that computation cannot be altered or manipulated after the fact.
As Cysic expands into AI inference, AI training, and broader compute workloads, the mainnet becomes a foundational layer for accountability. It allows customers to trace and verify every computational action that takes place within the network.
Leo explained that Cysic built an internal ZK virtual machine that allows developers to write programs in Rust or Python without requiring deep expertise in zero-knowledge cryptography. These programs are compiled into opcode sequences that run on the ZKVM, after which the network handles proof generation and verification.
Provers are selected through a VRF-based mechanism to generate proofs, while verifiers, including mobile users via the MyCysic application, validate them efficiently. The entire process is abstracted to reduce friction, and typical onboarding can be completed within two to three weeks.
Leo outlined a flywheel dynamic where increased hardware participation attracts more customers, which in turn generates higher yields for hardware providers. As yields increase, more participants contribute idle hardware, strengthening the network’s capacity and reliability.
He compared this ambition to decentralized storage networks such as IPFS and Filecoin, but with an additional layer of software optimization. Rather than simply aggregating idle hardware, Cysic aims to maximize performance through optimized software stacks combined with decentralized infrastructure.
Cysic represents an intersection of zero-knowledge cryptography, AI verification, GPU infrastructure, and decentralized economics. Leo’s long-term vision is not simply about accelerating proof generation but about transforming compute into a shared economic layer.
In a world where AI infrastructure is increasingly centralized and expensive, ComputeFi proposes a different model. Instead of a few corporations controlling computational power, distributed participants can contribute hardware and earn value proportionally.
If decentralized storage defined one era of Web3 infrastructure, decentralized compute may define the next.