infrastructure spotlight

VGG-16 Proven in 2.2s, On-Chain for 506k Gas: The Verifiable Compute Asset Class Is Born

Proofs for non-trivial CNNs and GPT-2 now complete in seconds, while deterministic frameworks let AI agents act directly on DeFi protocols. The infrastructure is here—and it's measurable.

Cast your mind back to 1494, when Luca Pacioli first codified double-entry bookkeeping, transforming the Florentine market from a den of verbal promises into a mechanism where every transaction was both record and proof. That same architectural shift now arrives for artificial intelligence. Verifiable AI infrastructure ceases to be a pamphlet of visions; it ships with the hard, unforgiving numbers of a trusted ledger: a zero-knowledge proof of VGG-16 inference completes in 2.2 seconds [^claim_3369]; a complete GPT-2 proof exists, sprinting past earlier transformer approaches [^claim_3370]; and verifying these proofs on-chain costs a fraction of a cent, executing in constant time [^claim_3373]. The tools are no longer theoretical—they are turning AI decisions into cryptographic primitives a smart contract can bank on.

We dissect the stack layer by layer, much like tracing the power lines of an ancient grid. First, the proving engines: zkPyTorch shoves practical zero-knowledge into non-trivial CNNs [^claim_3369], while Lagrange DeepProve targets the towering complexity of LLMs [^claim_3370]. Once a proof is generated, Ethereum’s BLS12-381 precompiles allow verifier contracts to check it in milliseconds—think of a customs officer stamping a passport without caring how long the ship took to cross the ocean [^claim_3371]. The architecture is ruthlessly elegant: a model’s weights commit to a public hash on-chain, the heavy computation happens off in the digital hinterlands, and only the proof and output reach the chain [^claim_3373]. Verification cost does not swell with model complexity; it’s a few elliptic curve operations, as cheap as checking a signature on a cheque.

But proofs are only half the armory. Some models can run directly on-chain if you’re willing to pay the gas tax. A fully on-chain learning framework demonstrates inference for a small CNN on Layer-1 at 143,647 gas per call, while a hungrier variant consumes 506,397 gas [^claim_3372]. These aren’t just numbers; they are precise cost envelopes that allow a protocol designer to price a classifier’s trustlessness like a futures contract—deciding which models stay L1-native and which must slink behind zk or TEE boundaries.

Where zkML stumbles against the sheer mass of generative models, TEE-based schemes shoulder in. Optimistic TEE-Rollups harness NVIDIA H100 confidential computing to deliver sub-second provisional finality, backed by fraud proofs and stochastic ZK spot-checks that feel like random audits in a high-security vault [^claim_3374]. Oasis’s ROFL similarly runs AI inside enclaves and anchors the results to on-chain contracts with the grim certainty of a notary public [^claim_3375]. Both architectures shift the heavy inference off-chain, the digital equivalent of moving a factory to a low-tax jurisdiction while keeping the quality-control office onshore.

On the agent side, verifiability is the difference between an autonomous trader and a rogue algorithm. EigenAI provides a deterministic LLM API compatible with OpenAI, guaranteeing that prompts, models, and answers are tamper-proof across runs, each response sealed like evidence in a chain-of-custody bag [^claim_3376]. EigenCompute extends this guarantee to long-running agent logic off-chain, turning sprawling computations into auditable trails. Meanwhile, HeyAnon’s AUTOMATE framework wraps DeFi interactions in deterministic validation schemes, letting AI agents execute swaps and manage wallets via a TypeScript SDK as if they were clicking buttons on a Bloomberg terminal [^claim_3379]. Beneath it all, EIP-7702 account abstraction gives agents their own smart-contract wallets, enabling multi-chain orchestration—Blockeden’s uAgents already compose models and APIs into on-chain workflows that hum like automated trading desks [^claim_3380].

And here comes the financialization, the part where the capital breathes. Cysic’s ComputeFi turns verifiable compute into an on-chain resource, a commodity that smart contracts can demand with model-specific authorization, much like demanding only organic cotton in a supply contract [^claim_3377]. Idle proof-of-work hardware, like Dogebox1, can toggle between mining and ZK proving as if switching between drilling for oil and refining it, depending on market spreads. Unique Network’s proof-of-compute has already minted over 3 million NFTs—one per AI task—each an immutable record with encrypted metadata and miner signatures, fully on-chain and as legally binding as a public deed [^claim_3378]. These are not sandbox demos; they are live systems trading on the street.

The implications for DeFi hit with the force of a margin call. zkML proofs let lending protocols accept privacy-preserving credit scores without ever trusting an oracle, turning what was once a whisper network into a market signal. On-chain inference at known gas costs lets AMMs embed ML-based rebalancing directly in their smart contracts, like automating a portfolio’s delta-hedge. TEE rollups can serve verifiable price feeds or MEV signals with sub-second latency, the speed of insider information, but now with a paper trail. Agent frameworks translate LLM-driven intents into safe contract calls, reducing the risk of rogue trades to a known, insurable quantity. Tokenized compute markets mint new roles for validators and miners as AI service providers, aligning economic security with the insatiable demand for gigaflops—labor and capital, unified under a single proof.

Proof times are seconds. Verification is constant-time and cheap. Agent frameworks are live. Protocol designers now hold a menu of verifiable compute primitives with concrete performance and cost profiles. The remaining question isn’t whether AI can be brought on-chain trustlessly—that trade is already priced in. The question is which applications will seize the high ground first, and which will be left holding the bag.

Provenance ledger

11 span-verified · 1 web-cited

11 claims below are locked to a verbatim span re-verified against the source. The remaining 1 is a web citation: 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] zkPyTorch (Polyhedra Network, March 2025) can generate a zero-knowledge proof of VGG-16 inference in 2.2 seconds, demonstrating practical zkML proving performance for non-trivial CNN models. span-verified
Verbatim source span
“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
d00070b979d520ca982c1187e54e2d80fab0f468a53848c5c05523ab57474580
↩ back to text
[2] Lagrange DeepProve generated the first complete GPT-2 proof and achieves major speedups over EZKL for transformer architectures, targeting large LLM inference with zkML. span-verified
Verbatim source span
“Lagrange DeepProve achieves major speedups over EZKL for transformer architectures and generated the first complete GPT-2 proof (ICME, Aug 2025).”
SHA-256 of span
198c4258da76ded830a3c86e05d6696287bc539d82b818c6f4073264bc348bff
↩ back to text
[3] On-chain zero-knowledge ML proof verification on Ethereum can be performed efficiently using precompiled BLS12-381 pairing operations, allowing smart contracts to verify zk-SNARK-based ML proofs in milliseconds to seconds without executing the model. span-verified
Verbatim source span
“Smart contracts can include verification logic to check the proof, often using precompiled cryptographic primitives. For example, Ethereum has precompiles for BLS12-381 pairing operations used in many zk-SNARK verifiers, making on-chain verification of SNARK proofs efficient… zkML compresses expensive AI inference into a small proof that blockchains (or any verifier) can check in milliseconds to seconds.”
SHA-256 of span
911384d5e9712c0152fa9733d3dcb37879b0446cbb567dbcb9c70b4af1b9b2f0
↩ back to text
[4] A fully on-chain decentralized learning framework can perform training on Layer-2, propagate verified model updates to Layer-1, and support gas-bounded low-latency inference inside Ethereum smart contracts, with CNN(F2, K1) inference consuming 143,647 gas and CNN(F8, K4) consuming 506,397 gas on L1 while remaining bit-exact to off-chain models formally proven in Z3. span-verified
Verbatim source span
“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… We develop quantization and loop-unrolling techniques that enable inference for logistic regression, SVM, MLPs, CNNs, and gated RNNs… while remaining bit-exact to their off-chain counterparts, formally proven in Z3… Fully on-ch
SHA-256 of span
2116007621ccea8c7ccb75d8a7cb5f78fca433700cc663f7f22f9ed81d6c0110
↩ back to text
[5] The ZKML-for-on-chain-agents architecture commits model weights via a public hash, keeps model weights and input data off-chain, and verifies proofs on-chain in constant time via a few elliptic-curve pairing operations, with verification costing a fraction of a cent in gas regardless of model complexity. span-verified
Verbatim source span
“Model weights and input data never leave the prover's machine. Only the proof and output go on-chain… The proof output becomes a first-class on-chain primitive… It is about proving off-chain inference so that on-chain systems can trust the result without trusting the agent… The proof, the model's public commitment (a hash of the weights), and the output are submitted on-chain. A smart contract (or Compact circuit on Midnight) verifies the proof in constant time — typically a few elliptic curve
SHA-256 of span
b3e3c973e8b203e8c02fd8a1c1744e12c70dfea7e28fb055ed022ea0d487bd08
↩ back to text
[6] Optimistic TEE-Rollups (OTR) use NVIDIA H100 Confidential Computing TEEs to provide sub-second provisional finality for generative AI inference on blockchain, combined with an optimistic fraud-proof mechanism and stochastic zero-knowledge spot-checks; the protocol defines Proof of Efficient Attribution (PoEA) to cryptographically bind execution traces to hardware attestations and guarantee model authenticity. span-verified
Verbatim source span
“We introduce Optimistic TEE-Rollups (OTR), a hybrid verification protocol… 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, thereby
SHA-256 of span
9a9292c889347af47f095aeb258352ecfe4556d404c141555dac69cca8e0e2cc
↩ back to text
[7] ROFL (Runtime Offchain Logic) on Oasis executes AI training, inference, and data analysis inside TEEs off-chain, then cryptographically verifies and connects results back to on-chain smart contracts, providing a verifiable off-chain compute framework targeted at AI workloads while preserving blockchain-level trust and privacy. span-verified
Verbatim source span
“ROFL enables developers to build privacy-preserving consumer and finance applications by leveraging Trusted Execution Environments (TEEs)… designed to enable developers to perform complex computations offchain while retaining blockchain-level trust, verification, and privacy… developers can execute resource-intensive operations off-chain within secure enclaves, then cryptographically verify and connect the results back to smart contracts on-chain.”
SHA-256 of span
6b501e0ec554d06ef983f28f740a604a35e1e2f71e6dd836f77e1800ddf0014c
↩ back to text
[8] EigenAI provides a deterministic, verifiable LLM inference API compatible with the OpenAI API, ensuring consistent results across runs and allowing developers to verify that prompts, models, and answers were not tampered with; EigenCompute extends this with verifiable compute for complex, long-running agent logic run off-chain while preserving smart-contract-like integrity. span-verified
Verbatim source span
“The EigenAI platform enables developers to create verifiable applications using LLM inference, ensuring consistent results across different runs of the same LLM call. The solution offers a deterministic, verifiable application programming interface compatible with the OpenAI API that supports open-source LLMs and tool-calling… Eigen Labs said its method of verifying inference for LLMs relies on a technical breakthrough… for making inference deterministic… EigenCompute provides a verifiable comp
SHA-256 of span
4e59352b5c68c0bcc70a89160253eadc28f1a8aaa8bc3b83adc037509063de79
↩ back to text
[9] Cysic’s mainnet exposes verifiable compute as an on-chain resource (ComputeFi), enabling smart contracts and agents to verify that outputs are produced by specific, authorized AI models and supporting decentralized scientific computing workloads, including the ability for idle PoW infrastructure (e.g., Dogebox1 devices) to toggle between mining and ZK proving based on market demand. web-cited
Excerpt reported by researcher (not re-verified)
“Cysic describes this model as the foundation of ComputeFi – a decentralized economy where verifiable compute becomes a programmable, onchain resource… AI model verification: enabling smart contracts and agents to verify that outputs came from specific, authorized models… Decentralized scientific computing for verifiable workloads in genomics, climate research, and cryptography… Idle PoW infrastructure re-use, with devices like the Dogebox1 able to toggle between mining and ZK proving, based on

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] Unique Network’s proof-of-compute system mints an NFT for every completed AI task, embedding encrypted task metadata and results, digital signatures from miners’ cryptographic keys, bindings to miner hotkeys, and structured on-chain metadata, yielding more than 3,000,000 NFTs since May as tamper-proof compute records providing 100% on-chain verifiability with no off-chain dependencies. span-verified
Verbatim source span
“Since May, the system has minted more than 3 million NFTs, each serving as a verifiable, tamper-proof record of completed AI tasks… Every completed compute task is minted as a signed, immutable NFT on Unique Network, transforming each token into a ‘proof object’ containing encrypted task metadata and results, digital signatures from miners’ cryptographic keys, ownership bindings to miner hotkeys, and structured, on-chain metadata. This ensures 100% on-chain verifiability with no off-chain depen
SHA-256 of span
0b884301dc5cea26446c592301f290bb0f0661e7086651c0fa6f8358adb9e104
↩ back to text
[11] HeyAnon’s AUTOMATE framework is a TypeScript SDK for DeFi × AI agents that validates on-chain calls against predefined deterministic schemes and exposes functions for chain data access, wallet management, and smart contract interactions (including token swaps and zkCRO wrapping), providing a structured agent integration layer over multiple DeFi protocols. span-verified
Verbatim source span
“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… The SDK is able to handle various on-chain functions like calling chain data (e.g., balance enquiry), wallet management (create and transfer funds, get latest block, get transactions by address), and smart c
SHA-256 of span
209b2f6f9a00a76f815a1cca9a0848ad9d0c94807574fdf87ae31167bb688334
↩ back to text
[12] Account abstraction via Ethereum’s EIP-7702 (live in Pectra since May 2025) underpins multi-chain AI-DeFi agents, enabling agent-controlled smart-contract wallets; frameworks like Blockeden’s uAgents turn data streams, ML models, and APIs into on-chain agents that can be composed via an Agentverse marketplace. span-verified
Verbatim source span
“The underlying technical foundation that makes this possible is account abstraction — specifically Ethereum's EIP-7702, which went live with the Pectra upgrade in May 2025… Its uAgents turn data streams, ML models, and APIs into on-chain agents that can be deployed and orchestrated into complex workflows via the Agentverse marketplace.”
SHA-256 of span
71b87fa0c44ce3a30aff2b21040a0e5ea7c45b91279c2c3b766d1caa273cc92b
↩ back to text

Sources

  1. https://blog.icme.io/the-definitive-guide-to-zkml-2025/
  2. https://www.ancilar.com/knowledge-hub/blogs/verifiable-ai-inference-how-zkml-enables-trust-minimised-on-chain-model-execution
  3. https://blockeden.xyz/blog/2025/04/22/verifiable-on-chain-ai-with-zkml-and-cryptographic-proofs
  4. https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.AFT.2025.35
  5. https://cercuit-ola.github.io/zklm-docs/
  6. https://arxiv.org/abs/2512.20176
  7. https://unchainedcrypto.com/press-release/oasis-protocol-foundation-launches-rofl-mainnet-verifiable-offchain-compute-framework-powering-ai-applications/
  8. https://siliconangle.com/2025/10/01/eigencloud-launches-platform-verifiable-ai-infrastructure/
  9. https://natlawreview.com/press-releases/cysic-launches-mainnet-break-zk-and-ai-bottlenecks-turning-compute
  10. https://www.coindesk.com/press-release/2025/11/11/unique-network-redefinestrust-in-decentralized-ai-with-nft-based-proof-of-compute
  11. https://crypto.com/en/research/defai-jan-2025
  12. https://blockeden.xyz/blog/2025/09/21/ai-agents-orchestrating-multi-chain-services/
verifiable-aizkmltee-rollupsdefi-agentson-chain-inferencecomputefiaccount-abstraction
AUTOMATED

Get the synthesis

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