Fully Homomorphic Encryption Is Overhyped for On-Chain AI

FHE operations introduce ~10-1000x slowdown versus plaintext computation, making interactive AI agents impossible. Zero-Knowledge Machine Learning (ZKML) frameworks like EZKL and Giza prove privacy with verification, not execution, offering a pragmatic path forward.

• Key Benefit 1: ZK proofs verify a private inference in seconds, FHE computes it in minutes.
• Key Benefit 2: Enables verifiable AI without the prohibitive runtime cost of homomorphic ops.

FHE's massive computational burden translates directly to gas costs. A single private inference could cost >$100 on Ethereum L1, destroying any viable economic model. Projects like Infernet and Ritual are exploring hybrid models that use trusted hardware or off-chain compute with on-chain settlement to bypass this.

• Key Benefit 1: Offloads FHE's heaviest work to specialized co-processors or networks.
• Key Benefit 2: Creates a sustainable fee market for private AI, unlike pure on-chain FHE.

FHE protects computation, but the fundamental input/output data must still be posted on-chain. For most DeFi or gaming AI agents, the request and result are the valuable secrets—FHE doesn't hide them. Solutions like TCP and Automata use TEEs (Trusted Execution Environments) to guarantee private execution with integrity, a more complete solution.

• Key Benefit 1: TEEs (e.g., Intel SGX) provide full data confidentiality and code integrity.
• Key Benefit 2: Avoids the cryptographic overhead of FHE for equivalent privacy guarantees.

Despite hype, no production-grade FHE-optimized Layer 2 or co-processor network has meaningful adoption. Contrast with the ~$20B TVL in ZK-Rollups like zkSync and Starknet, which have spent years optimizing provers. FHE requires entirely new hardware (e.g., FHE ASICs) not yet deployed at scale.

• Key Benefit 1: ZK-Rollups have proven scaling paths (recursive proofs, custom VMs).
• Key Benefit 2: Hardware acceleration for FHE is still a research project, not a product.

FHE operations are orders of magnitude slower than plaintext computation. On-chain AI inference is already a latency and gas-cost nightmare; FHE makes it economically impossible for real-time applications.

• Latency: FHE ops are ~10,000x to 1,000,000x slower than plaintext.
• Cost: A single encrypted inference could cost >$100 in gas, versus cents for a verifiable ZK proof.

For on-chain AI, you don't need to hide the model/data during execution; you need to cryptographically verify the result. Zero-Knowledge Machine Learning (ZKML) like zkML, EZKL, or Giza proves a computation was performed correctly off-chain with ~100-1000x less overhead than FHE.

• Key Use: Model integrity, verifiable inference, and fair sequencing.
• Entities: Modulus Labs, Gensyn, Risc Zero.

If input/output privacy is the goal, use specialized hardware (TEEs) or multi-party computation. Trusted Execution Environments like Intel SGX provide a secure, isolated enclave with near-native speed. MPC distributes trust among parties.

• TEEs: ~native speed, trusted hardware requirement.
• MPC: Trustless but slower, ideal for high-value, low-frequency ops.

FHE's killer app is not on-chain AI inference, but encrypted state and conditional logic. Projects like Fhenix and Inco use FHE to enable private smart contracts where state is always encrypted, enabling novel DeFi and gaming mechanics impossible with ZK or TEEs alone.

• Use Case: Private auctions, sealed-bid governance, hidden-game-state RPGs.
• Trade-off: Accept high cost for specific, high-value privacy guarantees.