AI Execution Becomes a Slashing-Backed Asset Class On-Chain
Verifiable inference, TEEs, and threshold cryptoeconomics are converging to make AI outputs executable by smart contracts, with slashing and staking for accountability.
In the year of our algorithm 2025, the blockchain archipelago is undergoing a cargo cult’s apotheosis—except the gods are now silicon-based. The transition from mere data feeds to fully-fledged, economically bonded on-chain actors mirrors the late 19th-century shift from signal flags to telegraphic trading, when latency became the first derivative of value. Just as the transatlantic cable turned arbitrage into a matter of minutes, modern protocols are embedding AI inference directly into execution and economic layers, turning models into verifiable, slashing-backed primitives. This is not just an upgrade; it’s a structural convergence across the stack, and the implications reek of the same systemic upheaval that followed the Buttonwood Agreement.
NEAR Protocol, with its claimed 1M TPS and 600ms block times[^claim_3469], is building the rail gauge for this new continent. Its sharded L1 hosts AI inference with the cold efficiency of a Swiss watchmaker’s wet dream. The IronClaw sandbox, nestled inside hardware-enforced TEEs like Intel TDX and NVIDIA GPU enclaves, isolates credentials with the paranoid precision of a Bond villain’s vault—user secrets never touch the model during private inference, gliding through a zero-trust labyrinth where even the janitor can’t peek[^claim_3470]. To marry access with token incentives, NEAR AI lets you stake NEAR for monthly compute credits, gaining entry to 43 models without a credit card’s sticky residue, all while retaining token ownership until you unstake[^claim_3472]. And NEAR Intents executes cross-chain actions with sub-cent fees and sub-second settlement across 35+ chains, integrating AI-driven trades into the very wallets you carry—Ledger included[^claim_3471]. The interface hums with the quiet menace of a bespoke trading floor, every transaction a micro-contract with fate.
Supra’s Threshold AI Oracles are the realpolitik answer to single-party cowboys. Instead of one node swearing on a Bible, a randomized committee seals each answer with a compact BLS threshold signature, slashing gas costs while preserving verifiability[^claim_3473]. But the roadmap is where the blood money flows: Phase 1 spits structured numerical outputs via statistical summarization; Phase 2 upgrades to JSON-formatted commands with semantic checks through clustering and pairwise logic validation; Phase 3 aims to generate executable smart contract code on demand, once a quorum nods and simulation-based testing greenlights the deployment[^claim_3474]. This progression transforms oracles from passive data pipes into code generators that smart contracts can trust—a move that smells of a dark pool clearing house, where information trades on its own credibility spread.
Meanwhile, systems like Keryx embed AI inference into consensus itself, turning the BlockDAG into a high-frequency truth machine. Its OPoI (Optimistic Proof of Inference) churns at 10 blocks per second, with miners committing to an AI execution rather than a puzzle hash. Results are accepted instantly but can be challenged within a window; fraud is sniffed out by re-running a deterministic model on-chain, and dishonest miners lose 20% of escrowed rewards—a swift, cold blade to the jugular[^claim_3477][^claim_3478]. Though on-chain slashing is, as of June 2026, temporarily disabled[^claim_3478], the design whispers of a bazaar where inference is a credit default swap, priced by the cost of deception. The BlockDAG pulses like a server room in a cyber-noir flick, each block a heartbeat of economic coercion.
Ritual’s Infernet pushes further, treating inference as a primitive commodity. It exposes precompiles for AI, ZK, and TEEs, while its Cascade protocol splits prompts into tokens among nodes so no single node sees the whole picture—reassembly and proofs land on-chain with the delicacy of a cocktail napkin’s coded message[^claim_3476]. Complementing TEE-based privacy, zero-knowledge machine learning (ZKML) proves correctness off-chain and returns a succinct proof for on-chain verification, with Groth16, Plonk, STARKs, and Halo2 serving as the verifiable compute layers[^claim_3481]. This is the cryptography of the paranoid elite, a world where even the light you read by is audited.
At the application layer, the Crypto.com AI Agent SDK and HeyAnon AUTOMATE framework let devs conjure deterministic DeFi logic from natural language, with on-chain calls validated against predefined schemes[^claim_3479]. It’s the whispering gallery of high finance, where a trader’s mutterings become smart contracts. Lightchain AI’s AIVM processes smart contract–verified inference jobs and autonomously pays GPU operators in LCAI tokens, a self-regulating bazaar of compute[^claim_3480]. Even model training slithers on-chain: the PoIm protocol on Ethereum L2 verifies each micro-step update as a self-verifying transaction and runs full inference of a 10-layer CNN within gas limits[^claim_3475]. The yield on this labor is priced in trust, not just tokens.
The playbook is obvious: AI compute is becoming a first-class resource on public blockchains, collateralized by economic stakes and cryptographic proofs. Smart contracts can now call inference directly, condition on its outputs, and slash for fraud—creating a trust-minimized substrate for agentic DeFi. The missing piece, full on-chain slashing and robust fraud proofs for general models, remains under development like a derivatives contract awaiting regulatory green light. But the architectural shift is priced in already: AI compute is no longer just an oracle feed; it’s a programmable, slashing-backed execution layer, and the market for verifiable truth is going ex-dividend.
Provenance ledger
9 span-verified · 4 web-cited9 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] NEAR Protocol’s sharded L1 infrastructure is specified to support 1,000,000 transactions per second with 600 ms block times and 1.2 s finality, while maintaining 100% mainnet uptime for over 5 years. span-verified
“NEAR Protocol is fully sharded, quantum-adaptive blockchain infrastructure that has maintained 100% uptime on mainnet for over 5 years. NEAR protocol supports 1 million TPS, with 600ms blocks and 1.2s finality.”
222c492591fb2c17c49fa9cb1c65c14f8ce50a5da0f56b33dd873f2d7ec7dadc [2] NEAR AI’s IronClaw system sandboxes tools and isolates credentials inside hardware-enforced TEEs (including Intel TDX and NVIDIA GPU TEEs) so that user secrets never reach the model during private inference and agent execution. span-verified
“Deploy agents using the private inference infrastructure serving Abound, Brave, and Venice AI. IronClaw sandboxes tools and isolates credentials inside hardware-enforced TEEs to ensure your secrets never reach the model.” and “NEAR AI uses Intel TDX gateways for CPU-level confidential compute, securing inference from edge to cloud alongside NVIDIA GPU TEEs.”
328b4ca307a7fbbac9aa2ac116d7762ce11d9e4a3b5e15688a063b21614dd0bd [3] NEAR Intents provides cross-chain execution with sub-cent fees and sub-second settlement across 35+ chains and 150+ assets, with $19B in cumulative volume, integrated into wallets like Ledger and Brave. span-verified
“NEAR Intents: universal cross-chain execution… Integrated by Ledger, Brave Wallet, Infinex, ThorSwap, HOT Wallet, and leading DeFi protocols for cross-chain execution with sub-cent fees and sub-second settlement. $19B in volume to date.” and “Trade across 35+ chains and 150+ assets from one account.”
e23ab889bba5be7398a252b98af931999f5bb1b108363d27994a510da6def286 [4] NEAR AI’s staking-based payment system converts staked NEAR into monthly compute credits for confidential AI inference and autonomous agents, allowing access to 43 AI models without a credit card while keeping stakers’ token ownership until unstaking. span-verified
“NEAR Protocol has launched a staking-based payment system for its AI platform, allowing users to access confidential inference and autonomous AI agents by locking NEAR tokens instead of paying with a credit card… Users can access all 43 AI models on NEAR AI without a credit card, while retaining ownership of their staked tokens until they choose to unstake… ‘staking NEAR equates to AI usage, prepaid in a form you can recover.’”
c9007343d97b9ff32a6bfdfb1b71ff6e4212e200f565270f1421d7540cda7c08 [5] Supra’s Threshold AI Oracles validate every AI answer via a randomized sub-committee of nodes and seal it with a compact BLS threshold signature, providing cryptographically verifiable AI outputs with reduced gas footprint. span-verified
“Threshold AI Oracles eliminate that friction by fusing inference and execution into one seamless, decentralized workflow that delivers AI insights on-chain securely on Supra… Every answer is validated by a randomized sub-committee of Supra nodes and sealed with a compact BLS threshold signature… Gas Efficiency: This protocol uses threshold signatures, which are extremely compact… Both of these design choices reduce gas bloat and gas costs by a large margin.”
ac4c861e1849ed764ed4077b7c496dbc655cc0ab086eddbef5b09749aa6d4da8 [6] Threshold AI Oracles on Supra are designed to roll out in three phases: Phase 1 returns structured numerical outputs via statistical summarization; Phase 2 returns JSON-formatted commands plus reasoning with clustering and pairwise logic checks; Phase 3 enables on-demand generation of executable smart contract code once a quorum of agents agrees and logic passes verification or simulation-based testing. span-verified
“Phase 1… will be able to return simple, verifiable numerical values… These outputs are generated via statistical summarization (e.g., median of agent responses)… Phase 2… produce JSON-formatted outputs containing a predefined command, parameters, and an explanation… To ensure semantic consistency, agent responses are clustered using AI embeddings and validated with pairwise logic checks… Phase 3, Threshold AI Oracles can generate executable smart contract code on demand… Once a quorum of agents
a300bdd76212cd423d3516d814c8e317552d2d8393e5ae7aa1342944a134e5d5 [7] The Proof-of-Improvement (PoIm) protocol for decentralized learning on Ethereum L2 verifies each micro-step model update as a self-verifying training transaction and supports fully on-chain inference of moderately complex models such as 10-layer CNNs with quantized integer arithmetic within Ethereum gas limits. web-cited
“A novel Proof-of-Improvement (PoIm) protocol governs the training process and verifies each decentralized micro update as a self-verifying training transaction… PoIm is a decentralized protocol on L2 that governs and verifies micro-step model training and its L1-propagated updates… While this incurs gas costs, our optimized design allows even moderately complex models, such as 10-layer CNNs with quantized integer arithmetic, to execute efficiently within Ethereum’s gas limits.”
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[8] Ritual’s Infernet is a decentralized oracle layer for heterogeneous compute that treats inference as a primitive and exposes precompiles for AI, zero-knowledge, and TEEs, while the Cascade protocol for private LLM inference splits prompts into tokens dispatched to multiple nodes so no single node sees full context, with outputs reassembled on-chain with proofs. web-cited
“Ritual introduces: Inference as a primitive; Precompiles for AI, ZK, TEE; Cascade for privacy-preserving LLM execution… Infernet… a decentralized oracle layer built for heterogeneous compute… Use cases: AI inference: Feed an on-chain prompt to a remote LLM; TEE workflows… ZK proof generation and verification… Cascade is Ritual’s peer-reviewed protocol for private AI inference, accepted to ICML 2025… Prompts are split into tokens; Tokens are dispatched to multiple nodes; No single node sees the
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[9] Keryx’s OPoI (Optimistic Proof of Inference) embeds AI inference directly into a BlockDAG proof-of-work consensus, committing each block to an AI execution whose result is accepted immediately but can be challenged within a time-bounded window, with fraud detected by re-running a deterministic fixed-point model on-chain and dishonest miners losing 20% of escrowed block rewards. span-verified
“We propose Keryx, a high-throughput BlockDAG protocol that embeds Artificial Intelligence inference directly into the proof-of-work consensus. Miners do not merely hash — they compute. Each block carries a cryptographic commitment to an AI execution… The core mechanism, OPoI (Optimistic Proof of Inference), follows the established optimistic-rollup pattern: inference results are accepted immediately and can be challenged within a time-bounded window. Fraud is detected by re-running a determinis
d470dba05be4b2d55f42c31d3a039c90afbd72acea43f700f9e12d2dddf4399b [10] The Keryx network operates at 10 blocks per second via GHOSTDAG, stores full inference results on IPFS with only a 34-byte CID (SHA-256 of content with prefix) committed on-chain, and since the Phase 3 hardfork (DAA 15,550,000) supports multiple open-weight models with per-model economic minimums enforced at the consensus level, although on-chain slashing is temporarily disabled as of June 2026. span-verified
“The network operates at 10 blocks per second via the GHOSTDAG protocol, enabling near-instant finality for AI tasks… As of the Phase 3 hardfork (DAA 15,550,000 on mainnet), Keryx supports multiple open-weight models with per-model economic minimums enforced at the consensus level. Inference results are stored on IPFS; only a 34-byte CID is committed on-chain… The on-chain CID is the SHA-256 of the content prefixed with… Current status (June 2026): on-chain slashing is temporarily disabled while
537c8a69e2d041216651738cb883e9392eada9a60bb494cec62beda0dad88896 [11] Crypto.com’s AI Agent SDK exposes blockchain interaction capabilities such as chain data queries, wallet creation and transfer, and smart contract interactions (including token swaps and wrapping zkCRO) via a natural-language interface, while HeyAnon’s AUTOMATE TypeScript framework integrates new DeFi protocols using deterministic logic where on-chain calls are validated against predefined schemes. span-verified
“Crypto.com’s AI Agent SDK… processing natural language and executing instructions directly, such as calling chain data (e.g., balance enquiry), wallet management (transfer functions), and basic smart contract interactions (swap)… The SDK is able to handle various on-chain functions like calling chain data… wallet management (create and transfer funds, get latest block, get transactions by address), and smart contract interactions (swap token, wrapping zkCRO)… In January, HeyAnon introduced AUTO
b4b8970a39d1f8081288fd2c8deb9769bc3d186bf7640e588b8d0a14ab6b4ba4 [12] Lightchain AI’s on-chain inference module uses an AI Virtual Machine (AIVM) to process smart contract–verified inference jobs in a decentralized network, coupled with an autonomous payment system that rewards GPU operators in LCAI tokens for successfully completing inference tasks. web-cited
“The team is finalizing a robust on-chain inference module that will allow users to submit and receive AI queries directly from the Lightchain dashboard. Powered by Lightchain’s AI Virtual Machine (AIVM), this module enables smart contract-verified inference jobs that are processed in a decentralized network… To support this decentralized AI backbone, Lightchain AI is implementing an autonomous payment system that rewards GPU operators for successfully completing inference tasks. Rewards will be
This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.
[13] Zero-knowledge machine learning (ZKML) proves that a specific neural network inference was executed correctly off-chain and returns a succinct proof for on-chain verification, with protocols such as Groth16, Plonk, zkSTARKs, and Halo2 forming the core verifiable compute layer for AI outputs in crypto systems. web-cited
“Zero-knowledge proofs, specifically zk-SNARKs, provide the cryptographic bridge, enabling verifiable computation… ZKPs enable verifiable inference. A model’s computation is proven correct off-chain, with a succinct proof verified on-chain by a smart contract… Zero-Knowledge Machine Learning proves a specific neural network inference was executed correctly, off-chain, and outputs a succinct proof for on-chain verification… Feature / Metric: zkSNARKs (Groth16, Plonk); zkSTARKs; Halo2 (PSE, Scroll
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://near.org/
- https://crypto.news/near-lets-users-pay-for-ai-services-by-staking-tokens/
- https://supra.com/academy/threshold-ai-oracles-brings-intelligence-on-chain/
- https://drops.dagstuhl.de/storage/00lipics/lipics-vol354-aft2025/LIPIcs.AFT.2025.35/LIPIcs.AFT.2025.35.pdf
- https://medium.com/@rajpal.pagare44/title-ritual-the-execution-layer-for-the-ai-crypto-synthesis-23cc113c7426
- https://keryx-labs.com/whitepaper
- https://crypto.com/en/research/defai-jan-2025
- https://www.binance.com/bg/square/post/26059299298930
- https://www.chainscorelabs.com/en/blog/ai-x-crypto-agents-compute-and-provenance/verifiable-computation-and-zkml/why-zk-proofs-are-the-missing-link-for-ai-x-crypto