infrastructure spotlight

Verifiable AI goes settlement-grade: ERC-7992, 300M proofs, 86% cheaper rollups

zkML has crossed from toy demos to a standardized, high-throughput trust layer. ERC-7992 ties model commitments to on-chain verification, while proof networks and TEE rollups make auditable AI inference cheap enough for DeFi to settle on.

Verifiable AI inference has crossed from research into settlement infrastructure. ERC-7992 gives Ethereum a common ABI for model commitments and zk proof verification[^claim_748]; Mina’s zkML converts ONNX models into on-chain-verifiable proofs[^claim_749]; Inference Labs’ verified networks have passed 300 million proofs[^claim_757]; and Optimistic TEE-Rollups cut verification costs by ~86% at sub-second finality[^claim_753]. The take: zkML is no longer a boutique primitive — it’s a standardized, high-throughput trust layer that DeFi and agents can build on.

The standardization piece is ERC-7992. It defines a registry of immutable ModelCommitment objects — hashes of weights, architecture, proving circuit/AIR, and verifying key, plus a proofSystemId — and a single verifier interface verifyInference(modelId, inputCommitment, output, proof) that dispatches to Groth16, Plonk, or STARKs without ABI changes[^claim_748]. Inputs are bound by domain-separated, nonceable commitments for replay protection; outputs are ABI-encoded bytes; an optional extension persists verified records on-chain for auditability and deterministic settlement[^claim_754]. This gives any settlement contract a standard way to consume AI output as a provable fact.

Tooling sits on top. Mina’s zkML library converts ONNX-format models into zero-knowledge circuits, generates proofs on private inputs, and verifies them on-chain via o1js smart contracts, with a zkML Verifier tool exposing a REST API for arbitrary input validation[^claim_749][^claim_755]. The developer flow separates off-chain proving from lightweight on-chain verification — the same pattern ERC-7992 formalizes at the ABI level.

Latency is no longer the blocker. zkPyTorch, released in March 2025, proves VGG-16 inference in 2.2 seconds[^claim_752]; Modulus Labs has verified an 18-million-parameter model on-chain with SNARKs[^claim_751]. That is the shift from toy examples to near-interactive performance regimes[^claim_752].

Scale has followed. Inference Labs’ Proof of Inference protocol, live on testnet with mainnet eyed for late Q3, has processed over 281 million zkML proofs by August 2025[^claim_750]. Its Subnet-2 passed 300 million zk proofs by late 2025 and added 160 million more on testnet in 2026, running on restaked ETH[^claim_757]. Hundreds of millions of proofs settled in decentralized inference markets — that is the scale at which auditable AI autonomy becomes a real service.

For L2s, Optimistic TEE-Rollups combine TEE attestations with fraud-proof challenge games and stochastic ZK spot-checks, cutting verification costs by approximately 86% while delivering sub-second provisional finality through TEE attestations[^claim_753]. That is the cost profile for high-throughput verifiable AI inference.

On the execution side, HeyAnon’s AUTOMATE framework — released January 2025 — validates every on-chain call against predefined schemas before broadcasting, so AI-determined actions conform to protocol-safe transaction patterns[^claim_756]. It is the policy gate between a capable agent and an irresponsible transaction.

For DeFi, the pairing is direct: ERC-7992’s optional audit trail gives oracle networks and settlement contracts a deterministic record of which model produced which output[^claim_754]; inference markets like Inference Labs’ can feed yield routers, copy-trading agents, and automated oracles with provable AI signals secured by restaked ETH[^claim_757]; OTR cost curves make an AI coprocessor rollup serving DeFi message-passing viable[^claim_753].

What changes: AI outputs become settlement-grade assets. What to watch: whether ERC-7992 becomes the ABI that DeFi AI integration points reference, and whether mainstream L2s adopt OTR-style cost mechanics for inference. The stack is no longer speculative — it is a pipeline from ONNX weights to an on-chain, auditable, economically secured inference result.

Provenance ledger

8 span-verified · 2 web-cited

8 claims below are locked to a verbatim span re-verified against the source. The remaining 2 are web citations: the URL was checked, but the excerpt is the researcher's summary and was not re-derived from the page. Citation markers in the text jump here.

[1] ERC‑7992 standardizes verifiable ML inference on Ethereum by defining a registry of immutable `ModelCommitment` objects (hashes of model weights, architecture, proving circuit/AIR, and verifying key plus a `proofSystemId`) and a verifier interface `verifyInference(modelId, inputCommitment, output, proof)` that dispatches to different proof systems (Groth16, Plonk, STARKs) without changing the ABI. span-verified
Verbatim source span
This ERC standardizes how smart contracts reference machine-learning (ML) models and accept zero-knowledge attestations of their inferences. It defines a registry that issues a `modelId` for a `ModelCommitment`, hashes of the model’s weights/architecture, proving circuit/AIR, and verifying key, along with a `proofSystemId` for the proving system... A verifier interface provides `verifyInference(modelId, inputCommitment, output, proof)`: it retrieves the model commitment, dispatches verification
SHA-256 of span
e5342dc9947778c7abce09d4a078d82fcc27fcba0186fab8a847f1c590889697
↩ back to text
[2] Mina’s zkML library converts AI models in ONNX format into zero‑knowledge proof circuits, generates proofs of AI inference on private inputs, and verifies these proofs on‑chain via o1js smart contracts and a zkML Verifier tool that exposes a REST API and supports arbitrary input submission for validation. span-verified
Verbatim source span
Mina’s zkML library enables anyone to generate a zero knowledge proof from an AI inference job using private inputs... This allows you to: Convert AI models (in the widely used ONNX format) into zero knowledge proof circuits; Generate a Mina proof of the AI inference job of the ONNX model on private inputs... Submit these proofs to the Mina blockchain, where they can be verified in a trustless environment... To make it easy for anyone to verify a proof and record it on the Mina chain we are also
SHA-256 of span
de5702d28aaedcd642cb5b29d296332b7a2d23b9ce98763ce5409061d555da49
↩ back to text
[3] Inference Labs’ Proof of Inference protocol, implemented in their Verified Inference Network, has processed over 281 million zkML proofs by August 2025 on testnet, enabling AI models to be converted to zero‑knowledge circuits where execution traces are proven off‑chain and outputs attested on‑chain as on‑chain AI proofs for settlement in decentralized inference markets. web-cited
Excerpt reported by researcher (not re-verified)
The Proof of Inference protocol stands as Inference Labs' flagship innovation... Live on testnet, with mainnet eyed for late Q3, it has already processed over 281 million zkML proofs by August 2025... At its core, zkML transforms AI inference into a provable computation... Inference Labs converts them into zero-knowledge circuits, proving execution traces off-chain while attesting results on-chain. By generating on-chain AI proofs, Inference Labs enables provers, verifiers, and users to settle t

This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.

↩ back to text
[4] Modulus Labs demonstrated on‑chain verification of an ML model with 18 million parameters using SNARKs, and their work cited in “The Cost of Intelligence” shows proof systems capable of verifying machine learning models with up to 18 million parameters directly on blockchain networks, representing production‑grade zkML scale. span-verified
Verbatim source span
Modulus Labs shows the benchmarking proof systems for on-chain inference up to 18 million parameters... Their paper titled “The Cost of Intelligence” presents proof systems capable of verifying machine learning models with up to 18 million parameters directly on blockchain networks.
SHA-256 of span
09a67162f5a3eba73536979ac9d5783ca1e562d1cc61b6695b6c8a521b4951fe
↩ back to text
[5] The ICME Labs 2025 zkML guide reports that zkPyTorch, released in March 2025, can generate a proof of VGG‑16 inference in 2.2 seconds, and that by 2025 there are models that can be proven in seconds with ZKP, moving zkML beyond toy examples into near‑interactive performance regimes. span-verified
Verbatim source span
In 2025, ZKML is still expensive... zkPyTorch dropped in March 2025 and suddenly you can prove VGG-16 inference in 2.2 seconds... In 2025, we're way past the toy phase. There are models that can be proven in seconds with ZKP right now.
SHA-256 of span
aae289e3f7668411f3f6a3a864ad8061f8f2254e94f3f82676a511b0f203ec2a
↩ back to text
[6] Optimistic TEE‑Rollups (OTR) are hybrid L2 protocols that use TEEs for hardware‑rooted attestations, optimistic rollup fraud‑proof style challenge games, and stochastic zero‑knowledge spot‑checks; empirical evaluations show OTR can reduce verification costs by approximately 86% while providing sub‑second provisional finality (≤1s) through TEE attestations for high‑throughput verifiable AI inference workloads. span-verified
Verbatim source span
Optimistic TEE-Rollups (OTRs) are hybrid layer-2 blockchain protocols combining TEEs with fraud-proof mechanisms to reduce gas costs and latency... Empirical evaluations show OTRs can lower verification costs by ~86% and support high-throughput applications like scalable verifiable AI inference... TEEs generate attestations binding inputs, outputs, and code hashes... Smart contracts verify these attestations rapidly (≤1s), enabling sub-second “provisional finality” for users... Advanced Use Case
SHA-256 of span
ecb3df640ce3f290a1f68936ed96f2a73c1ed2096d56a3cf26567e3c28d236e0
↩ back to text
[7] The ERC‑7992 verifier interface binds inferences to domain‑separated `inputCommitment`s (which can be nonce‑based for replay protection) and allows arbitrary ABI‑encoded `bytes` outputs, while an optional extension persists verified inference records on‑chain for auditability and deterministic settlement of AI‑driven contracts. span-verified
Verbatim source span
Inputs are bound by domain-separated commitments (nonceable for replay protection), outputs are ABI-encoded bytes whose schema can be application-defined or additionally committed on-chain, and proof systems (e.g., Groth16/Plonk/STARK) are pluggable without ABI changes. An optional extension persists verified inference records to enable auditability and deterministic settlement.
SHA-256 of span
a3c507e4327d81375ff328948d5fbdcb77368a47f690a4b55247b808674cb7ef
↩ back to text
[8] The Mina zkML Verifier tool allows developers to deploy o1js smart contracts that verify zkML proofs and push them to the Mina chain via a REST API, separating proof generation off‑chain from lightweight on‑chain verification and enabling arbitrary client applications to submit inputs and retrieve verified results without handling low‑level proving details. span-verified
Verbatim source span
To make it easy for anyone to verify a proof and record it on the Mina chain we are also releasing the zkML Verifier— a tool for deploying smart contracts to verify proofs on chain. zkML verifier allows you to verify proofs using o1js smart contracts and push proofs to chain using a REST API for proof verification, it also allows you to upload arbitrary input for validation.
SHA-256 of span
b616ed2601f99008afd408cf930661599bfe30c1faeb5aeb05c9d7c794f7df07
↩ back to text
[9] HeyAnon’s AUTOMATE framework (released January 2025) is a TypeScript framework that integrates new DeFi protocols via deterministic logic where every on‑chain call is validated against predefined schemas, effectively acting as a structured agent execution layer that checks that AI or off‑chain agent actions conform to protocol‑safe transaction patterns before broadcasting. span-verified
Verbatim source span
In January, HeyAnon introduced AUTOMATE, a TypeScript framework that enables developers to easily integrate new DeFi protocols into the HeyAnon ecosystem. The framework leverages deterministic logic, where on-chain calls are validated against defined schemes, to ensure the accuracy of on-chain actions.
SHA-256 of span
64ebee9168b31a57b8a50e9386fee3fff5450a2b6f86f5053f27eaa9555f2811
↩ back to text
[10] Inference Labs’ Subnet‑2, part of their Verified Inference Network, processed over 300 million zero‑knowledge proofs by late 2025 and an additional 160 million on testnet in 2026, demonstrating that their restaked‑ETH‑secured network can operate at hundreds‑of‑millions‑of‑proofs scale for auditable AI autonomy. web-cited
Excerpt reported by researcher (not re-verified)
Their Subnet-2 already smashed 300 million zk proofs by late 2025, and now with 160 million more on testnet in 2026, they're scaling like wildfire... Inference Labs' Subnet-2 achieved over 300 million zero-knowledge proofs, marking a major step in auditable and verifiable AI autonomy.

This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.

↩ back to text

Sources

  1. https://eips.ethereum.org/EIPS/eip-7992
  2. https://minaprotocol.com/blog/minas-zkml-library-developer-guide
  3. https://decentralizedinference.org/2026/02/11/inference-labs-zkml-proofs-for-verifiable-decentralized-ai-inference-markets
  4. https://kudelskisecurity.com/modern-ciso-blog/zkml-verifiable-machine-learning-using-zero-knowledge-proof
  5. https://blog.icme.io/the-definitive-guide-to-zkml-2025/
  6. https://www.emergentmind.com/topics/optimistic-tee-rollups-otr
  7. https://crypto.com/en/research/defai-jan-2025
  8. https://decentralizedinference.org/2026/02/22/inference-labs-zkml-proofs-for-verifiable-decentralized-ai-inference-in-defi
zkmlerc-7992verifiable-inferenceteesoptimistic-rollupsdefi-agentsminainference-labs
AUTOMATED

Get the synthesis

AI×crypto research, repackaged with every claim hash-locked to its source. New arXiv → analysis in ~3 hours.