infrastructure spotlight

Sell the Computation, Buy the Attestation: The Ledger’s New Asset

From TEE serverless sandboxes to pre-execution inference mempools, AI infrastructure is making outputs consumable by smart contracts without clogging consensus—or trusting the operator.

Treat the ledger as a border checkpoint. The expensive inference happens off-site, before validation; only the proofs—or attestations—cross into the settlement zone. Cysic Mainnet styles itself “the world’s first full-stack verifiable compute network,” a treaty between zero-knowledge proofs, AI computation, decentralized hardware, and blockchain infrastructure in a single trustless ecosystem [^1086]. It has shipped over 10 million ZK proofs and assembled 260,000+ nodes: provers, verifiers, mobile devices, and professional compute operators [^1087]. That is not an API catalog. That is an execution layer printing cryptographic evidence.

Marlin Oyster is running the heavy machinery inside Trusted Execution Environments—a decentralized cloud where DeFi strategies, automation scripts, and AI models go to sweat [^1088]. Oyster Serverless gives each function a 5-minute clock, a JavaScript or WASM sandbox sealed in a TEE, a shared vCPU, and 100 MB of memory, with request-response gates open to web2 endpoints or smart contracts [^1089]. The toolchain tames the output for on-chain consumption: reproducible builds via Nix, a Solidity library that verifies attestations with RISC Zero’s blessing, and remote-attestation-based TLS through Scallop [^1090]. A DeFi protocol can call a strategy, accept an output, verify the attestation—no need to trust the node operator, just the hardware’s sealed testimony.

OpenGradient treats the critical path like a toxic asset and moves it off the books. Its whitepaper is a cargo manifest: HACA, node architecture, a verification spectrum, consensus and settlement, x402 payment-gated LLM inference, PIPE on-chain ML execution, MemSync, token economics [^1091]. PIPE routes inference into an inference mempool before block inclusion, so smart contracts get AI models without EVM congestion, inferences running in parallel across the queue [^1092]. The pitch: hundreds, even thousands of pending transactions can run inference in parallel while block building stays light, because the expensive part is shifted into the mempool’s shadow economy [^1093]. The underlying stack is Cosmos SDK with EVM compatibility and CometBFT consensus, and verification can anchor TEE hardware attestation on-chain in a TEE Registry [^1094]. Model Hub now hosts 4,500+ AI models from 100+ developers, and BitQuant puts the user count at 1.8M+, with 69M+ AI agent queries logged [^1095].

SperaxOS stands at the agent layer, checking passports. TypeScript SDK, Python SDK, REST API, and an MCP server let existing agent frameworks plug straight into its DeFi tool catalog [^1096]. Every on-chain action passes through a dedicated security layer—spend caps, MEV protection, deterministic risk scoring, and a human in the loop when the move is radioactive [^1097]. Agents register as NFTs under the ERC-8004 standard on Arbitrum, which gives each one a persistent, verifiable identity [^1098]. The result: the agent becomes a named, permissioned actor, not a private key loose in the machine room.

Now the collateral question: which trust anchor do you accept as payment? Marlin and OpenGradient take TEE attestation; Cysic doubles down on ZK proofs. TEEs are the fast trade—speed but you believe the hardware and a registry, a counterparty risk that’s hard to hedge. ZK proofs are the cryptographic fortress—stronger, but the cost to generate them is a heavy toll. That spectrum—pure computation legend versus hardware faith—determines which workloads settle on-chain today.

Watch the protocols that treat inference as a mempool primitive, not a lumbering oracle call. The infrastructure that wins will make execution invisible to block production—off-book, off critical path—and make the attestation cheap enough to post. The market, as always, pays for evidence, not effort.

Provenance ledger

9 span-verified · 4 web-cited

9 claims below are locked to a verbatim span re-verified against the source. The remaining 4 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] Cysic Mainnet is described as "the world’s first full-stack verifiable compute network" combining zero-knowledge proofs, AI computation, decentralized hardware, and blockchain infrastructure in a single trustless ecosystem. span-verified
Verbatim source span
“Today marks a defining milestone for Cysic as we officially launch the Cysic Mainnet, the world’s first full-stack verifiable compute network that unites zero-knowledge proofs, AI computation, decentralized hardware, and blockchain infrastructure into a single trustless ecosystem.”
SHA-256 of span
77833bec89156bed7bc0ee8e41fe982c57bb025e803ab682232696a16bad9ee9
↩ back to text
[2] Cysic says its network has delivered over 10 million ZK proofs and onboarded 260,000+ nodes, including provers, verifiers, mobile devices, and professional compute operators. span-verified
Verbatim source span
“Delivered over 10 million ZK proofs” and “Onboarded 260,000+ nodes, including provers, verifiers, mobile devices, and professional compute operators.”
SHA-256 of span
78dab15d9c1f24900b3d1149eced934b61d35052c5192ffad6df00281366d1e7
↩ back to text
[3] Marlin Oyster uses TEEs to run complex workloads such as DeFi strategies, automation tasks, and AI models over a decentralized cloud, with two execution modes: Oyster CVM and Oyster Serverless. span-verified
Verbatim source span
“Marlin is a verifiable computing protocol leveraging TEEs to allow complex workloads (like DeFi strategies, automation tasks or AI models) to be deployed over a decentralized cloud.”
SHA-256 of span
46bbce99bdd9284c4b4297140be5d50530341c522bca12977fe7fafe86720ac5
↩ back to text
[4] Marlin Oyster Serverless imposes a 5 minute execution limit, runs in a JS or WASM sandbox inside a TEE, uses shared vCPU and 100 MB memory, and supports request-response access through web2 endpoints or smart contracts. span-verified
Verbatim source span
Table entries list “Execution Time | 5 minute limit (may be modified)”, “Environment | JS or WASM (WIP) sandbox (in a TEE)”, “Resource Allocation | Shared vCPU (min 5%) and memory (100 MB)”, and “Access Modes | Request-Response (either through web2 endpoints or smart contracts).”
SHA-256 of span
a285b9c1f6d0e88db227893dd1f80f0a9658b4d65b1422b56cdac92c91373ac4
↩ back to text
[5] Marlin’s toolkit includes reproducible builds via Nix, attestation verification via a Solidity library with RISC Zero support, and remote-attestation-based TLS via Scallop. span-verified
Verbatim source span
“Reproducible Builds: Oyster uses Nix…”, “Attestation Verification: Solidity library with RISC Zero support…”, and “Remote Attestation-based TLS… through Remote Attestation-based TLS for encrypted communication using Scallop.”
SHA-256 of span
15b1b75502040fd2b3ce2778528edefd107c9d4579e0fb4d69f406595bd61301
↩ back to text
[6] OpenGradient’s whitepaper frames the architecture around HACA, node architecture, a verification spectrum, consensus and settlement, x402 payment-gated LLM inference, PIPE on-chain ML execution, MemSync, and token economics. span-verified
Verbatim source span
The table of contents lists “Hybrid AI Compute Architecture (HACA)”, “Verification Spectrum”, “Consensus and Settlement”, “x402: Payment-Gated LLM Inference”, and “PIPE: On-Chain ML Execution.”
SHA-256 of span
7e5d2730fb6d361b223b2869ab85b01517c3ef0a42ead487489a3428a74fdabf
↩ back to text
[7] OpenGradient’s PIPE executes AI inference in an inference mempool before block inclusion, allowing smart contracts to use AI models natively without EVM congestion and with inferences run in parallel. web-cited
Excerpt reported by researcher (not re-verified)
“PIPE (Parallelized Inference Pre-Execution Engine)” and “smart contracts and applications on our network can natively use and execute AI models without introducing any overhead or congestion inside the EVM… Inferences are executed in parallel.”

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
[8] OpenGradient says PIPE can run inferences for hundreds or thousands of pending transactions in parallel, with block building remaining fast because expensive AI work is shifted into the inference mempool. web-cited
Excerpt reported by researcher (not re-verified)
“With PIPE, we can run inferences for hundreds or thousands of pending transactions in parallel…” and “actual block building remains extremely fast.”

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
[9] OpenGradient’s research summary says the network is built on Cosmos SDK with EVM compatibility and CometBFT consensus, and that inference verification can use TEE hardware attestation anchored on-chain in a TEE Registry. web-cited
Excerpt reported by researcher (not re-verified)
“The OpenGradient network is built on the Cosmos SDK with EVM compatibility and CometBFT (BFT) consensus…” and “Inference verification supports… TEE hardware attestation… with attestation anchored on-chain in the TEE Registry.”

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] OpenGradient’s public materials state that its Model Hub hosts 4,500+ AI models from 100+ developers and that BitQuant has 1.8M+ registered users and 69M+ AI agent queries. web-cited
Excerpt reported by researcher (not re-verified)
“Model Hub (hub.opengradient.ai): decentralised repository hosting 4,500+ AI models from 100+ developers” and “BitQuant… 1.8M+ registered users, 69M+ AI agent queries.”

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
[11] SperaxOS exposes a TypeScript SDK, Python SDK, REST API, and MCP server so existing agent frameworks can plug directly into its DeFi tool catalog. span-verified
Verbatim source span
“A TypeScript SDK, Python SDK, REST API, and MCP server let any existing agent framework plug directly into the SperaxOS tool catalog.”
SHA-256 of span
862e988cdd1c6e17a1cef171d503b6a4be12556a7b154e1de914e48b75da3d54
↩ back to text
[12] SperaxOS claims guarded execution with spend limits, MEV protection, deterministic risk scoring, and human-in-the-loop approval for sensitive operations. span-verified
Verbatim source span
“Every on-chain action passes through a dedicated security layer with spend limits, MEV protection, deterministic risk scoring, and human-in-the-loop approval for sensitive operations.”
SHA-256 of span
d4dbb1716fe81b63caafdd08c7d138c63b0f9033698c84d4e3872065c1a8ea11
↩ back to text
[13] SperaxOS says its on-chain agent economy registers agents as NFTs under the ERC-8004 standard on Arbitrum, giving each agent a persistent, verifiable identity. span-verified
Verbatim source span
“Agents register as NFTs under the ERC-8004 standard on Arbitrum, giving every agent a persistent, verifiable identity.”
SHA-256 of span
9e733d15c5a91aee6d80b42232f52cce43cf41bff907e425b474f338a0e042f5
↩ back to text

Sources

  1. https://docs.cysicfoundation.org/
  2. https://docs.marlin.org/oyster/introduction-to-marlin/
  3. https://opengradient.foundation/whitepaper
  4. https://docs.opengradient.ai/learn/onchain_inference/execution
  5. https://www.binance.com/en/research/analysis/opengradient
  6. https://www.manilatimes.net/2026/07/15/tmt-newswire/globenewswire/sperax-and-ibm-enter-strategic-partnership-to-take-open-source-defi-agents-to-the-enterprise/2384510
verifiable-computeteezk-proofsonchain-aidefi-agentsinfrastructure
AUTOMATED

Get the synthesis

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