Leo Fan holds a PhD in cryptography from Cornell University and has a diverse background that includes academic research, FPGA/ASIC design, and protocol engineering. He has played a significant role in advancing zero-knowledge proofs from theoretical ideas to industrial-scale systems.
At Cysic, he leads a multidisciplinary team developing ZK and AI-optimised ASICs, high-throughput GPU clusters, and a decentralised compute network that turns compute into a programmable on-chain asset.
We recently spoke with Leo Fan, Founder and CEO of Cysic, about the crucial hardware limitations that hinder the adoption of zero-knowledge (ZK) technology and how his company’s new custom chips are poised to overcome them.
Cysic is a vertically integrated compute network that combines custom silicon, high-performance infrastructure, and a programmable chain to turn GPUs and ASICs into liquid, yield-generating assets. With the recent launch of their mainnet, Cysic aims to make real-time ZK proving a universal utility for the decentralised web.
Read more about Cysic’s “full-stack” approach to verifiable compute and the convergence of AI and ZK in the interview below.
You founded Cysic with the belief that zero-knowledge proofs are a vital breakthrough for the internet, yet their deployment has faced delays. What specific bottlenecks did you identify in the current landscape, and how did you address them?
We believed zero-knowledge proofs were the most significant breakthrough for the Internet in decades. They could make trust affordable and scalable. However, upon examining their implementation, we discovered it was far from ideal. Generating proofs was slow and costly, leading to stagnation in the adoption of ZK proofs on the Internet, particularly on Ethereum.
We founded Cysic to address this issue, starting with the hardware involved. We realised that to make ZK a genuine utility, it was not enough to just develop better software. A new type of machine had to be built.
The industry relies heavily on GPUs, but you have argued that they are inefficient for ZK workloads. Why is the architecture of a GPU ill-suited for the specific finite field arithmetic required for zero-knowledge proofs?
GPUs excel at what they were designed for: parallel processing in graphics and AI. However, ZK proof generation demands a very different kind of mathematics. It is characterised by complex, sequential operations in finite field arithmetic.
It is not about raw power but efficiency. Using a GPU for ZK is like using a racecar to plough a field. It consumes vast energy for minimal results. This inefficiency acts as a hidden tax, driving up L2 fees and hindering real-time applications. To overcome this bottleneck, we need hardware that understands the language of ZK maths, not graphics processing.
Cysic is introducing custom silicon with ZK Air and ZK Pro. How do these two ASICs differ in their target use cases, and what does the existence of a portable device like ZK Air signal about the future of decentralised proving?
ZK Pro is our solution to the challenge of industrial-scale proving. It is a custom ASIC explicitly built for ZK computations. It enables us to generate proofs rapidly enough to keep up with block production, significantly reducing cost and latency compared to alternatives.
ZK Air represents the other side of the coin. It is a custom ASIC, about the size of a MacBook charger, that outperforms a server GPU. Its existence proves that high-stakes proving and verification do not need to be centralised in an expensive data centre. Together, these two products redefine how proofs will be generated in the future, whether by large-scale enterprises, households, or retail users on the go.
For developers, integrating ZK today often involves assembling disconnected tools and hardware. When you describe Cysic as a “full-stack compute network,” what complexity are you simplifying?
For developers today, using ZK is a fragmentation nightmare. They must piece together algorithms, cloud hardware, and coordination layers, each with its own failures and costs. “Full-stack” means we deliver a unified pipeline, from our custom hardware to our decentralised prover network, that hides all that complexity.
As a result, ZK ceases to be an R&D challenge and becomes a dependable utility. Developers can finally build applications that rely on verifiable compute without becoming experts in hardware logistics or distributed systems engineering.
Current ZK implementations often rely on batch processing, which introduces latency. If we shift to a “real-time” proving model, what new categories of applications become feasible that simply cannot exist today?
Today, ZK is often a background batch process, causing latency between an action and its final verification. Real-time proving nearly eliminates that delay.
This introduces a new category of application: responsive dApps, where every user action is privately and instantly verified. For instance, on-chain games with complex logic become feasible, as do DEX trades that settle with proof. The user experience becomes seamless.
You are working with major teams, such as Scroll. Why are these L2s choosing to offload their provisioning infrastructure to Cysic instead of building it in-house?
When teams are building the networks of the future, they cannot afford infrastructure bottlenecks. They choose Cysic because we provide the reliability, ultra-low latency, and predictable economics they need to deliver a seamless experience to their own users.
It all comes down to focus. Their expertise lies in building great L2s or privacy protocols. Our strength lies in making ZK proofs cheap and fast at a global scale. By relying on our backbone, they can focus on innovation at their layer, knowing that the foundational compute layer is handled.
Your testnet attracted over 500,000 users before the mainnet launch. Beyond stress testing the network, what did that level of participation teach you about the latent demand for verifiable compute?
The main conclusion was that the demand for verifiable compute is even greater than we anticipated. The tests certainly succeeded in stress testing the network, which it passed with flying colours.
With the mainnet now active, what immediate features are accessible to developers and partners? Does this signify the beginning of the network’s actual economic cycle?
On Day One, developers and partners gain access to a production-ready, live, and cost-effective prover-verifier network. They can build commercial applications on top of it, with real staking, real fees, and tangible rewards, right now.
The launch activates the whole Cysic economy, transforming our testnet’s potential into a live utility that can immediately reduce costs and accelerate innovation across the entire ecosystem.
We hear a lot about the intersection of AI and crypto. You see the “black box” nature of AI as an existential crisis for decentralisation. How does Cysic’s verifiable compute layer address this?
ZK and AI face the same fundamental problem: the black box. How can one trust the output of a complex, opaque computation? As AI becomes central to dApps, this issue transcends technical concerns. It turns into an existential crisis for decentralisation. A smart contract cannot depend on a centralised API’s promise and still claim to provide a non-custodial, decentralised solution.
This is why we have decided to build a verifiable compute layer that solves this problem for both fields.
Looking ahead, you have discussed the “commoditisation of trust”. What does the internet look like when high-performance, verifiable compute becomes a widespread public good?
The future involves the complete commodification of trust. High-performance, verifiable computing will become a widespread, decentralised public resource. It will be as vital to the next internet as broadband is today. The infrastructure will recede into the background, powering applications we can barely conceive of now.
Cysic’s goal is to help build that foundational layer.