infrastructure spotlight

Proofs, Not Models, Decide What Agents Can Execute

zkML, optimistic challenges, and TEE attestations all anchor AI outputs on-chain. The differentiator is the trust assumption your protocol can underwrite.

Crypto AI no longer has a proof-generation gap. It has a verification-selection gap. The stack converges on the same architecture—heavy inference off-chain, settlement on-chain—but the verification regimes carry distinct trust roots: zkML relies on cryptographic math and sometimes trusted setup, opML relies on 1-of-N honest validators and economic game theory, TEEs rely on a hardware root of trust and physical security.[^claim_959] What a dApp can safely automate depends on which failure it can price.

PoIm makes the split explicit. Model parameters and execution logic are stored fully on-chain, while actual inference runs off-chain as EVM view functions.[^claim_946] It also supports fully on-chain inference inside state-modifying transactions for DeFi interactions.[^claim_946] Contracts get two lanes: low-cost reads and atomic AI-informed writes.

OpenGradient’s HACA separates execution from verification across inference nodes, full nodes, and data nodes.[^claim_947] Developers choose TEE attestations for low-overhead LLM inference, zkML proofs for high-stakes ML like DeFi liquidation logic, or vanilla signatures for lightweight workloads.[^claim_947] That triage is already operating at scale: over 2,000,000 verifiable AI inferences and more than 500,000 zkML proofs plus TEE attestations executed on its network.[^claim_953]

The optimistic path shows what on-chain finalization looks like. ORA’s Onchain AI Oracle runs opML on Arbitrum: off-chain inference, then an on-chain challenge period before the result is dispatched to the user’s smart contract callback.[^claim_948] The challenge period is the primitive—a dispute window DeFi state machines can inherit. The reach is broad: Ethereum mainnet, Arbitrum One, Base, and Optimism all support on-chain AI inference through Ora Protocol and Ritual as of 2026.[^claim_955] Ritual’s Infernet runs agent model inference off-chain in sidecars and returns ZK proofs of outputs on-chain, so each action is auditable.[^claim_954]

The cryptographic path is moving from demo to settlement-grade. Polyhedra’s zkPyTorch compiler transforms PyTorch AI models into zero-knowledge circuits.[^claim_949] Extropy reports a production-ready system proving full inference of GPT-2, and zkLLM benchmarks verify a 13-billion-parameter model inference in under 15 minutes with a proof smaller than 200 kB.[^claim_952] That is not an AMM tick; it is a settlement-grade proof. zkML now works as a ‘digital passport’: an agent must present a ZK proof tied to a specific model weights hash, like a verified safety-aligned Llama-3 instance, so on-chain users can distinguish an approved model from a malicious fork.[^claim_958]

TEEs cover the latency and privacy side. Oasis ROFL executes complex computations privately in Trusted Execution Environments and anchors verifiable results on-chain.[^claim_950] Phala has integrated over 30,000 TEE devices into a verifier network for secure, privacy-preserving computation, including decentralized AI training and data processing, with cryptographic proof guarantees.[^claim_951] This moves trust to the hardware vendor and enclave operator, creating a new validator-class actor. Even with zk proofs advancing, zkML still faces hurdles scaling to LLMs, while opML and TEEs have already bridged the gap to production-grade on-chain AI applications.[^claim_959]

That infrastructure is now an execution surface for agents. WAIaaS exposes a unified REST API that executes actions across at least 14 DeFi protocols: Aave v3, LI.FI, Lido, 0x, Pendle, and Kamino on EVM; Jupiter, Drift, and Jito Staking on Solana; Across Protocol cross-chain; and Hyperliquid and Polymarket multi-chain.[^claim_956] Coupled with Ritual-style proofs, this is a control plane for verifiable multi-venue agent actions. Galaxy’s onchain inference capital markets describe the end state: users send prompts to GPU operators coordinated by crypto token incentives and onchain settlement, receiving cryptographic or economic guarantees about output correctness and privacy instead of centralized API access to frontier labs.[^claim_957]

The next product is not a better model. It is a risk market for proof regimes—slashing around optimistic challenges, model-hash allowlists, and TEE attestation pricing. Protocols that survive will underwrite a specific trust failure, not all of them.

Provenance ledger

12 span-verified · 2 web-cited

12 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] PoIm is a decentralized L2 protocol that stores model parameters and execution logic fully on-chain while performing inference off-chain via EVM view functions, and also supports a mode where inference is executed entirely on-chain within state-modifying transactions for DeFi interactions. web-cited
Excerpt reported by researcher (not re-verified)
“We introduce PoIm, a decentralized L2 protocol that governs model training and propagates verified updates to L1 for inference… In this tier, the model parameters and execution logic are stored fully on-chain. However, the actual inference computation is performed off-chain by calling view functions, which are executed by any EVM client… Fully On-chain Inference… Here, the inference is executed entirely on-chain as part of a state-modifying transaction, typically one that interacts with a DeFi

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
[2] OpenGradient’s Hybrid AI Compute Architecture (HACA) separates execution from verification by assigning GPU-heavy inference workloads to specialized inference nodes, consensus and proof-settlement to full nodes, and attested data provision to data nodes, supporting three verification modes: TEE attestations, zkML proofs, and vanilla signatures. span-verified
Verbatim source span
“Hybrid AI Compute Architecture (HACA): The network separates execution from verification through specialized node types, so full nodes run consensus and settle proofs, inference nodes handle GPU workloads, and data nodes provide attested external data… Developers choose the proof method that matches their risk profile, with TEE attestations for low-overhead LLM inference, zero-knowledge proofs for high-stakes ML models like DeFi liquidation logic, and vanilla signatures for lightweight workload
SHA-256 of span
6fa33a7d2f06aff37680b72e250432924ebb1da806a16e9c36d42b449aa7a50f
↩ back to text
[3] ORA’s Onchain AI Oracle (OAO) runs AI inference via an opML optimistic protocol on Arbitrum, where each AI request triggers off-chain inference by opML nodes, followed by an on-chain challenge period in which opML validators can dispute incorrect results before the inference output is finalized and dispatched to the user’s smart contract callback. span-verified
Verbatim source span
“ORA is Ethereum's Trustless AI… Its main product, Onchain AI Oracle (OAO)… 1. The user contract sends the AI request to OAO on Arbitrum, by calling `requestCallback`… 2. Each AI request will initiate an opML inference… 4. opML node will run the AI inference, and then upload the result on Arbitrum, waiting for the challenge period. 1. During the challenge period, the opML validators will check the result and challenge it if the submitted result is incorrect… 3. After the challenge period, the su
SHA-256 of span
071662f9e5c455f9cae901f881a85b7bb489339efe65795fc23a9c9220e9a44d
↩ back to text
[4] Polyhedra’s zkPyTorch compiler, launched in March 2025, transforms PyTorch AI models into zero-knowledge circuits so that ML computations can produce cryptographic proofs of correct execution suitable for decentralized applications. span-verified
Verbatim source span
“Polyhedra's zkPyTorch Compiler: Launched in March 2025, zkPyTorch transforms PyTorch AI models into zero-knowledge circuits, enabling cryptographic proofs of correct execution for machine learning. This breakthrough brings verifiable AI to decentralized applications, making ML computations provable and trustless.”
SHA-256 of span
98d999731d5fd786582c74abf732c048556ee1ad28d83c472757fafa44c00bce
↩ back to text
[5] Oasis Protocol’s ROFL mainnet, launched July 2025, uses Trusted Execution Environments to execute complex computations privately off-chain and then anchor verifiable results on-chain as a framework for privacy-preserving, verifiable off-chain compute for AI and other dApps. span-verified
Verbatim source span
“Oasis Protocol's ROFL Mainnet: In July 2025, Oasis Protocol Foundation launched ROFL, a mainnet framework for verifiable off-chain compute. ROFL leverages Trusted Execution Environments (TEEs) to perform complex computations privately and securely, then anchors verifiable results on-chain—ideal for AI and privacy-first dApps.”
SHA-256 of span
cf9c774c349d1a7cbc8edc69402df4f9154eebed47c9184d7848a28c1f5cef7e
↩ back to text
[6] Phala Network has integrated over 30,000 TEE devices into a verifier network designed to provide secure, privacy-preserving computation for decentralized AI training and data processing with cryptographic proof guarantees. span-verified
Verbatim source span
“Phala Network's TEE Verifier Network: Phala Network has integrated over 30,000 TEE devices, creating a robust verifier network for secure, privacy-preserving computation. This infrastructure is powering decentralized AI training and secure data processing with cryptographic proof guarantees.”
SHA-256 of span
c44e0ed0efcddf30811fc5e54cd9a227df749fe570a5f0e334aa5365b88b58e8
↩ back to text
[7] Extropy’s analysis of zkML reports that a production-ready system based on tlookup and zkAttn can generate a cryptographic proof for full inference of OpenAI’s GPT-2, and that zkLLM benchmarks show verification of a 13-billion-parameter model inference in under 15 minutes with a proof size smaller than 200 kB. span-verified
Verbatim source span
“This system represents the first production-ready zkML environment to successfully generate a cryptographic proof for a full inference of OpenAI’s GPT-2 model… Benchmarks indicate that zkLLM can verify the inference of a 13-billion-parameter model in under 15 minutes, producing a proof smaller than 200 kB.”
SHA-256 of span
db91e48a090ea9e96d8acc513877e42e87be257f761d83a769c8e5aca6d1428e
↩ back to text
[8] OpenGradient reports over 2,000,000 verifiable AI inferences and more than 500,000 zkML proofs plus TEE attestations executed on its network, indicating large-scale deployment of cryptographically verified AI workloads. span-verified
Verbatim source span
“The OpenGradient Network powers high-performance verifiable computing for AI… 2 Million+ Verifiable AI Inferences 500K+ zkML Proofs + TEE Attestations.”
SHA-256 of span
b87786a8449ca6bfb26d06176a472dfbe1e3da7d51b6064d18a32c28e6a095a6
↩ back to text
[9] Ritual’s Infernet protocol runs AI agent model inferences off-chain in sidecar provers and returns zero-knowledge proofs of the outputs on-chain, so that each agent action can be audited by verifying an on-chain ZK proof of the claimed computation. web-cited
Excerpt reported by researcher (not re-verified)
“Ritual introduces a compute layer where AI agents can run models offchain but prove their outputs onchain using zero-knowledge proofs (ZKPs)… Its primary components include: Infernet: The core protocol that handles compute tasks and returns ZK proofs Sidecars: Offchain provers that run the actual model inferences.”

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
[10] ORA and Ritual-backed infrastructure enable on-chain AI inference on Ethereum mainnet, Arbitrum One, Base, and Optimism as of 2026, exposing AI inference as verifiable oracle-style calls from smart contracts. span-verified
Verbatim source span
“Ethereum mainnet, Arbitrum One, Base, and Optimism all support on-chain AI inference through Ora Protocol and Ritual as of 2026.”
SHA-256 of span
303b21920d14d47c0767beabde5f4f5451993669394eb20edd4a17bac8918d44
↩ back to text
[11] The WAIaaS AI DeFi agent framework exposes a unified REST API that can execute actions across at least 14 protocols including Aave v3, LI.FI, Lido, 0x, Pendle, Kamino (EVM), Jupiter, Drift, Jito Staking (Solana), Across Protocol (cross-chain), and Hyperliquid and Polymarket (multi-chain), allowing agents to control multiple DeFi venues via standard HTTP requests. span-verified
Verbatim source span
“WAIaaS solves this with a unified REST API that abstracts away protocol complexity while preserving full functionality… Here's the complete list of integrated protocols: EVM: Aave v3, LI.FI, Lido, 0x, Pendle, Kamino Solana: Jupiter, Drift, Jito Staking Cross-chain: Across Protocol Multi-chain: Hyperliquid, Polymarket.”
SHA-256 of span
03dddd9ca96fd5d4729a1983feaec3293b0cbee0df27795ea964af8cf36b70ea
↩ back to text
[12] Galaxy’s definition of onchain inference capital markets describes networks where users send prompts to GPU operators coordinated by crypto token incentives and onchain settlement, in some configurations receiving cryptographic or economic guarantees about output correctness and privacy instead of centralized API access to frontier labs. span-verified
Verbatim source span
“Rather than routing every API call through frontier model providers like OpenAI, Anthropic… users can send prompts to networks of GPU operators coordinated by crypto token incentives and onchain settlement, and in some configurations receive cryptographic or economic guarantees about output correctness and privacy.”
SHA-256 of span
18dffbd2174edc60fe20658bed1fbd70f201e658e6c12419282af8ef6c01b3ed
↩ back to text
[13] Extropy’s zkML Singularity report states that zkML is being used as a “digital passport” for safe AI interaction by requiring agents to present a ZK proof tied to a specific model weights hash (e.g., a verified Llama-3 instance) alongside outputs, so that on-chain users can cryptographically verify that responses came from an approved model rather than a malicious fork. span-verified
Verbatim source span
“Applications began using zkML to cryptographically prove that a specific response came from a specific, safe model (e.g., a verified, safety-aligned Llama-3 instance) and not a malicious DeepSeek fork. zkML became the 'digital passport' for safe AI interaction. In this paradigm, an AI agent must present a ZK proof alongside its output, certifying that the output was generated by a specific model weights hash.”
SHA-256 of span
189ce9b898a67fecdb23d51ecf3092da643fca8b80a89c8dadff28ceae8b5be2
↩ back to text
[14] The Uplatz overview characterizes zkML, opML, and TEE-based verifiable compute as three distinct verification regimes with different trust assumptions: zkML relies on cryptographic math and sometimes trusted setup, opML relies on 1-of-N honest validators and economic game theory, and TEEs rely on a hardware root of trust and physical security, with zkML still struggling to scale to LLMs while opML and TEEs already support production-grade on-chain AI applications. span-verified
Verbatim source span
“zkML represents the cryptographic gold standard… zkML (Zero-Knowledge)… opML (Optimistic)… TEE (Confidential Computing)… |Verification Basis|Cryptographic Math|Economic Game Theory|Hardware Root of Trust| |Trust Assumption|Math + Trusted Setup (sometimes)|1-of-N Honest Validators|Hardware Vendor + Physical Security|… While purely cryptographic solutions (zkML) still face hurdles in scaling to LLMs, optimistic (opML) and hardware-based (TEE) solutions have bridged the gap, enabling the first gen
SHA-256 of span
6f0e25cbb20f309517f4917238269d67aae5ab9f654034c33113bb8884604469
↩ back to text

Sources

  1. https://drops.dagstuhl.de/storage/00lipics/lipics-vol354-aft2025/LIPIcs.AFT.2025.35/LIPIcs.AFT.2025.35.pdf
  2. https://new.deal/cryptocurrencies/opengradient/
  3. https://docs.arbitrum.io/for-devs/oracles/ora
  4. https://restakinglayer.org/2025/10/16/the-rise-of-verifiable-compute-offchain-proofs-and-flexible-trust-models-in-crypto-infra
  5. https://academy.extropy.io/pages/articles/zkml-singularity.html
  6. https://www.opengradient.ai/
  7. https://medium.com/@friscofr/ritual-the-missing-infrastructure-for-ai-in-web3-98a0f292711c
  8. https://www.ancilar.com/knowledge-hub/blogs/quick-guide-running-your-first-gen-ai-inference-on-ethereum
  9. https://dev.to/walletguy/14-defi-protocols-one-api-building-an-ai-defi-agent-49cj
  10. https://www.galaxy.com/insights/research/inference-capital-markets-ai-compute-gpu-futures-onchain-crypto
  11. https://uplatz.com/blog/verifiable-compute-for-ai-models-on-blockchain-the-convergence-of-cryptography-intelligence-and-consensus/
verifiable-computezkmlopmlteeai-oraclesdefi-agentsinference-capital-marketscrypto-infrastructure
AUTOMATED

Get the synthesis

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