The AI×Crypto Stack Hardens: On-Chain Learning Meets TEE-Rollup Hybrids
A wave of specialized infrastructure layers—fully on-chain ML, zkML, TEE-rollup hybrids, and AI oracle L1s—is turning verifiable AI compute from theory into production primitives for DeFi, governance, and agent economies.
The AI×crypto stack has moved past the whitepaper stage. These are live, interoperable primitives now. The Dagstuhl on-chain learning framework proves you can run nontrivial models entirely on Ethereum within gas limits. Using quantization and loop-unrolling, linear models consume 57,603 gas, CNNs up to 506,397 gas—all bit-exact to off-chain counterparts, formally verified in Z3[^claim_2312]. The framework runs on a Proof-of-Improvement (PoIm) protocol that only accepts model updates improving at least one core metric without degrading others. That makes decentralized, verifiable ML training possible on L2, with verified propagation to L1[^claim_2311].
When full on-chain execution gets too expensive, zkML offers a cryptographic shortcut. Polyhedra’s zkML framework compiles PyTorch-native models through ONNX preprocessing, ZKP-friendly quantization, and hierarchical circuit optimization into zero-knowledge circuits. It produces a proof of correctness alongside the output[^claim_2313]. Verifiable AI compute pipelines freeze a model by committing its weights and architecture hash, quantize, compile to a circuit, and generate proofs on GPUs that can be succinctly verified on-chain[^claim_2314]. The ‘Lookup Singularity’—driven by Jolt and Lasso protocols—now lets you prove Ethereum blocks in real time on consumer hardware. Massive LLMs are still out of reach[^claim_2319].
If throughput and cost matter more than pure cryptographic verifiability, TEEs are the play. Optimistic TEE-Rollups (OTR) pair NVIDIA H100 Confidential Computing TEEs for sub-second provisional finality with optimistic fraud proofs and stochastic ZK spot-checks, formalizing Proof of Efficient Attribution (PoEA). Simulations show 99% of centralized throughput at $0.07 per query[^claim_2315]. Marlin’s analysis positions TEEs as the most viable current solution for secure off-chain AI training and inference in Web3—isolated execution that keeps confidentiality and integrity intact[^claim_2320].
Production systems are live. Allora, a Cosmos-SDK L1 with distinct worker, reputer, and consumer roles, has served over 690 million inferences since its mainnet launch in November 2025[^claim_2316]. On the agent layer, ERC-8004 defines a lean discovery framework: three onchain registries for autonomous AI agents to find and interact trustlessly[^claim_2317]. The x402 protocol embeds stablecoin payments directly into the HTTP request–response cycle using EIP-3009 and Permit2, giving agents frictionless onchain payment rails[^claim_2318].
These layers don’t compete—they compose. On-chain ML handles low-latency, fully verifiable scoring. zkML delivers trustless inference for governance and compliance. TEE-rollup hybrids provide high-throughput generative AI oracles. Agent standards enable autonomous discovery and payment. The infrastructure is here. The question is which DeFi protocols will integrate first.
Provenance ledger
10 claims web-citedEvery 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] A fully on-chain decentralized learning framework can perform gas-prohibitive ML training on a Layer-2, propagate verified model updates to Layer-1, and enable gas-bounded, low-latency inference inside smart contracts using a Proof-of-Improvement (PoIm) protocol that only accepts updates which improve at least one core metric on a public benchmark without degrading others. web-cited
“We present the first decentralized, fully on-chain learning framework that: (i) performs gas-prohibitive computation on Layer-2 to reduce cost, (ii) propagates verified model updates to Layer-1, and (iii) enables gas-bounded, low-latency inference inside smart contracts. A novel Proof-of-Improvement (PoIm) protocol governs the training process… Updates are accepted by PoIm only if they demonstrably improve at least one core metric (e.g., accuracy, F1-score, precision, or recall) on a public ben
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[2] With quantization and loop-unrolling, fully on-chain inference for ML models on Ethereum L1 can be executed within the block gas limit while remaining bit-exact to off-chain counterparts formally verified in Z3, consuming 57,603 gas for linear models, 143,647 gas for CNN(F2, K1), and 506,397 gas for CNN(F8, K4). web-cited
“We develop quantization and loop-unrolling techniques that enable inference for logistic regression, SVM, MLPs, CNNs, and gated RNNs (with support for formally verified decision tree inference) within the Ethereum block gas limit, while remaining bit-exact to their off-chain counterparts, formally proven in Z3… Fully on-chain inference consumes 57,603 gas (≈ $0.18) for linear models, 143,647 gas (≈ $0.49) for CNN(F2, K1), and 506,397 gas (≈ $1.77) for CNN(F8, K4) on L1 (e.g., Ethereum).”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[3] Polyhedra’s zkML framework compiles PyTorch-native models via ONNX-based graph preprocessing, ZKP-friendly quantization, and hierarchical circuit optimization into zero-knowledge circuits where operations such as convolutions, softmax, and ReLU can be cryptographically verified, producing a zero-knowledge proof of correctness alongside the model output at inference time. web-cited
“1. Build Your Model (PyTorch-native)… 2. Compile to Zero-Knowledge Circuits. The model is transformed into a ZK-friendly format using: - Graph-based preprocessing (via ONNX) - ZKP-friendly quantization - Hierarchical circuit optimization. This ensures all operations—like convolutions, softmax, or ReLU—can be cryptographically verified. 3. Generate a Proof of Correctness… produces: - The output… - A zero-knowledge proof that the output is correct for the specific model and inputs.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[4] Verifiable AI compute pipelines using zkML or zkVMs freeze a model by committing its weights and architecture hash, quantize to chosen numeric formats, compile the model graph to a circuit, define public/private input–output statements, and generate proofs on GPUs that can be succinctly verified on-chain without re-executing the computation. web-cited
“Verifiable AI compute means generating cryptographic proofs that a model inference or training step was executed correctly… zkML frameworks compile neural network graphs into proof-friendly circuits… 1. Freeze the model: commit weights and architecture hash. 2. Quantize deliberately: pick numeric formats that preserve enough accuracy. 3. Compile to a circuit: generate proof artifacts from the model graph. 4. Define the statement: what inputs and outputs are public vs private? 5. Prove and verif
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[5] Optimistic TEE-Rollups (OTR) use NVIDIA H100 Confidential Computing TEEs for sub-second provisional finality combined with an optimistic fraud-proof mechanism and stochastic zero-knowledge spot-checks, formalizing a Proof of Efficient Attribution (PoEA) consensus that cryptographically binds execution traces to hardware attestations and achieves 99% of the throughput of centralized baselines with a marginal cost overhead of $0.07 per query. web-cited
“OTR leverages NVIDIA H100 Confidential Computing Trusted Execution Environments (TEEs) to provide sub-second Provisional Finality, underpinned by an optimistic fraud-proof mechanism and stochastic Zero-Knowledge spot-checks to mitigate hardware side-channel risks. We formally define Proof of Efficient Attribution (PoEA), a consensus mechanism that cryptographically binds execution traces to hardware attestations… simulations demonstrate that OTR achieves 99% of the throughput of centralized bas
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[6] Allora is a Cosmos-SDK L1 chain that functions as a decentralized AI inference network with distinct worker, reputer, and consumer roles, delivering context-aware, machine-learning-based predictions as on-chain signals and AI oracle outputs, and has served over 690,000,000 inferences since its mainnet launch in November 2025. web-cited
“Allora… is a decentralized AI inference network that produces context-aware, machine-learning-based predictions usable as on-chain signals and an AI oracle. Built as a Cosmos SDK L1 with worker, reputer and consumer roles… Mainnet launched Nov 2025; over 690M inferences served.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[7] ERC-8004 defines a lean discovery framework for autonomous AI agents on Ethereum composed of three onchain registries that provide primitives for agents to find, evaluate, and interact with each other trustlessly, built on top of the A2A protocol. web-cited
“ERC-8004 is a proposed standard that defines a discovery framework for autonomous AI agents on Ethereum. Built on top of A2A, its design is intentionally lean, consisting of three onchain registries that provide the basic primitives for agents to find, evaluate, and interact with each other trustlessly.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[8] The x402 protocol for AI agents and automated services embeds stablecoin payments directly into the HTTP request–response cycle, using EIP-3009 (transferWithAuthorization) for gasless USDC transfers and Permit2 for other ERC-20 tokens to handle onchain payment execution as part of internet-native resource access. web-cited
“x402 is a payment execution protocol. It handles the actual movement of money onchain… The protocol uses EIP-3009 (transferWithAuthorization) for gasless USDC transfers and Permit2 for other ERC-20 tokens… By embedding stablecoin payments into the request-response cycle, x402 gives AI agents, automated services, and MCP tool providers a frictionless way to transact.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[9] Recent zkML systems leveraging lookup arguments such as Jolt and Lasso have achieved real-time proving of Ethereum blocks on consumer hardware, marking a "Lookup Singularity" where zkML becomes a specialized production layer; however, these protocols still do not efficiently support massive large language models despite performance gains. web-cited
“The report details the technological breakthroughs that have occurred in late 2025, specifically the ‘Lookup Singularity’ in zkML driven by Jolt and Lasso protocols, which has enabled real-time proving of Ethereum blocks on consumer hardware… By late 2025, zkML has transitioned from a theoretical curiosity to a specialized production layer. While it has not yet scaled to support massive Large Language Models (LLMs) efficiently, significant breakthroughs in ‘Lookup Arguments’ have radically alte
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[10] Trusted Execution Environments (TEEs) are identified as the most viable current solution for secure, off-chain AI training and inference for Web3, providing isolated execution that safeguards confidentiality and integrity of processed data and serving as an optimal choice for off-chain computations necessary for AI applications on the blockchain. web-cited
“Trusted Execution Environments (TEEs) come with significant advantages when it comes to training and performing AI inference, in particular in terms of security assurances, isolation and data privacy and protection… Currently, TEEs emerge as the most viable solution for this requirement. TEEs provide a secure and isolated execution space for code, safeguarding the confidentiality and integrity of the data being processed. This makes them an optimal choice for offchain computations necessary for
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
Sources
- https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.AFT.2025.35
- https://www.polyhedra.network/zkML
- https://tokentoolhub.com/verifiable-ai-compute/
- https://www.arxiv.org/abs/2512.20176
- https://tempo-track-nu.vercel.app/companies/allora
- https://oasis.net/blog/erc-8004-trustless-agents
- https://eco.com/support/en/articles/12328618-x402-protocol-explained-how-ai-agents-pay-onchain
- https://uplatz.com/blog/verifiable-compute-for-ai-models-on-blockchain-the-convergence-of-cryptography-intelligence-and-consensus/
- https://blog.marlin.org/benchmarking-oyster-the-tee-based-cryptoxai-coprocessor-against-zkml