The decentralized compute network is pointing its GPU fleet at large language models after layer 2 fees rose above Ethereum's own
Cysic built a business around generating zero-knowledge proofs for Ethereum layer 2 networks. Now it is pointing that same hardware at AI.
The decentralized compute network has moved its focus from layer 2 proving to AI inference. The reason is a familiar one: demand for its original product fell, and layer 2 fees climbed above the fees on Ethereum itself.
From proofs to prompts
A zero-knowledge proof is a cryptographic receipt. It lets one party show that a computation was done correctly without anyone having to redo the work.
Layer 2 networks lean on these proofs to bundle transactions and settle them on Ethereum. Generating the proofs takes serious computing power, and that is the job Cysic signed up for.
AI inference is a different workload. It is the moment a trained model actually answers a question, the part of the process users see when a chatbot replies.
Both jobs eat GPUs. That overlap is what makes Cysic’s pivot workable rather than a total reinvention.
According to CEO Leo Fan, the plan is to put the company’s existing hardware to work on large language model (LLM) inference. The company is not abandoning ZK work entirely, though.
Fan has said Cysic still holds aspirations for proving Ethereum blocks in under eight seconds.
What Cysic is bringing to the table
The company’s mainnet launched on September 24, 2025, and has since handled more than 13 million transactions.
It has also produced more than 10 million ZK proofs. Those numbers were built in part through partnerships with Scroll and Aleo.
The network is large on paper too. Cysic counts more than 118,000 provers and 200,000 verifiers across its decentralized system.
Under the hood, Cysic runs a Proof-of-Compute layer 1 network that is compatible with the Ethereum Virtual Machine. That chain coordinates tasks for both ZK and AI workloads.
Cysic says its Venus zkVM and GPU optimizations improved proof-time efficiency by more than 9%.
Its custom C1 chip benchmarks at 1.31 million Keccak functions per second. Keccak is the hashing function at the heart of Ethereum, so speed there translates directly to faster proving.
The AI marketplace runs under the name CyAI, which sits alongside the company’s broader ComputeFi marketplace.
CyAI takes a notably pragmatic approach to trust. Instead of generating full zero-knowledge proofs for every AI task, it uses random statistical audits to check the work.
Why the pivot makes sense now
Fan has pointed to a broader force shaping the hardware market. AI data centers are soaking up GPU supply, and he cites those constraints as a driver behind the push for specialized hardware.
