infrastructure spotlight July 9, 2026

Verifiable AI Compute Splits Into Two Warring Factions: ZKML vs TEEs

ZKML proofs and TEE attestations each solve different halves of the verifiable inference problem, with concrete tooling now live for both approaches.

2 min read 10 claims web-cited

The AI×crypto stack is splitting into two distinct verifiable compute tracks, each with production-ready tooling and clear trade-offs. Zero-Knowledge Machine Learning (ZKML) and Trusted Execution Environments (TEEs) solve the same core problem—proving an AI inference ran correctly—but they serve fundamentally different latency, confidentiality, and composability profiles.

ZKML, via tooling like EZKL, converts deep learning models expressed as ONNX computational graphs into zk-SNARK circuits. This lets you verify inference on-chain in compressed form while keeping model details and intermediate activations hidden from the verifier [^claim_1602]. Modulus Labs has benchmarked these systems up to 18,000,000 parameters, proving on-chain inference verification is feasible for non-trivial models [^claim_1594]. Combined with Fully Homomorphic Encryption (FHE), ZKML lets you run inference on encrypted data with verifiable outputs in a single pipeline—neither the user nor the operator sees the other’s data [^claim_1597]. Worldcoin’s curated awesome-zkml repository highlights generic SNARK constructions that provide a reusable proving interface, independent of model architecture [^claim_1600].

TEEs take a different route. They run AI models natively and produce signed attestations that cryptographically link the inference output to specific model weights and prompts. You get onchain-verifiable inference receipts paid in stablecoins [^claim_1595]. Chainlink’s Confidential Reputable Execution (CRE) framework orchestrates AI workloads inside TEEs across decentralized oracle networks. Onchain agents can process sensitive financial data—credit histories, corporate ledgers—while ensuring the AI provider never touches the raw data [^claim_1596]. For workloads that need confidentiality, complex logic, or millisecond-level latency—block building, AI inference, private smart contracts—TEEs are described as more practical than purely on-chain execution or zero-knowledge approaches [^claim_1603].

On-chain inference remains a niche. Cortex is one of the few production chains that supports it directly, letting smart contracts call AI models deployed inside the blockchain execution environment. But this approach is largely absent from major ecosystems like Ethereum [^claim_1598].

Agent infrastructure is maturing in parallel. Ethereum’s proposed ERC-8004 standard defines on-chain registries for AI agent identity, reputation, and validation—agents can prove credentials and track performance in a decentralized way. Coinbase’s x402 protocol repurposes HTTP 402 to facilitate machine-to-machine cryptocurrency micropayments [^claim_1599]. AI oracles under development in the Chainlink ecosystem rely on multiple nodes independently executing an AI task and only reporting when consensus is reached on the inference output. That enables composable, trust-minimized AI services accessible directly from smart contracts [^claim_1601].

The convergence is clear: ZKML for integrity with privacy, TEEs for low-latency confidentiality, and oracle networks for Byzantine-resilient aggregation. Protocol teams can now plug into either track rather than building bespoke provers.

Provenance ledger

10 claims web-cited

Every claim below cites a source URL, and each URL was checked for validity before publish. The excerpt shown is the researcher's own summary of the page — it is not re-derived from the source, so it is not a verified verbatim quote. Follow the link to confirm any claim against the original. Citation markers in the text jump here.

[1] Modulus Labs’ ZKML proof systems can verify on-chain inference for machine learning models with up to 18,000,000 parameters, using SNARK circuits generated from ONNX models via the EZKL tooling. web-cited