Can You Tokenize an AI Model? Inside opML’s Real Test

The plumbing works. The question is what runs through it

Optimistic machine learning, opML for short, has spent about eighteen months proving a narrow but real point: you can run a large AI model off-chain, post the result to Ethereum with a cryptographic commitment, and let a challenge window plus a fraud proof stand in for re-running the whole thing on-chain. The idea, formalized by the team now called Ora in a January 2024 paper (arXiv 2401.17555), borrows its trust model wholesale from optimistic rollups: assume the answer is honest, and only do the expensive work if someone disputes it.

That part is settled. HOGE Wire has walked through the mechanics, the criticism, and the competitive squeeze from cheaper zero-knowledge proofs and faster trusted hardware in three earlier pieces. What none of them answered is the question every technology has to answer eventually: what is it actually for, and is anyone using it?

The market has an opinion. ORA, the token that meters Ora’s on-chain AI oracle, trades at $0.004764 with a market capitalization of roughly $792,000, ranked #3,075, on 24-hour volume of about $305 (CoinGecko, 22 August 2026). It is down 99.9% from a November 2024 high of $5.37. You do not arrive at $305 of daily turnover on a protocol whose core technology works unless the applications built on top of it have failed to find users. So this piece is about the applications: the oracle, the tokenized models, and the perpetual agents Ora has built on opML, and whether any of them is a business rather than a demo.

opML in one minute, for anyone who skipped the first three articles

opML publishes an AI inference result off-chain along with a Merkle commitment to the computation trace, then treats that result as final after a dispute window unless a watcher challenges it. If someone does challenge, the two sides play an interactive bisection game, narrowing a disagreement over millions of computation steps down to a single instruction that a Fraud Proof Virtual Machine can re-execute cheaply on-chain. The loser forfeits a stake. The pattern predates rollups: Jason Teutsch and Christian Reitwiessner described the same verification game in the Truebit paper years before Arbitrum and Optimism made it famous for general execution. Ora first floated the ML-specific version in an Optimism governance post in August 2023.

Ora’s implementation runs in two flavors. Single-phase replays every step inside the constrained VM, which is simple but slow. Multi-phase runs undisputed computation natively on GPUs and only drops into the VM for the one bisected step under dispute, which Ora says cuts the Merkle overhead from O(mn) to O(m+n) and speeds things up by “tens to hundreds of times” over the single-phase design (github.com/ora-io/opml). On the project’s own small-model benchmarks, inference lands in a couple of seconds and a full challenge cycle in a couple of minutes.

The security assumption is the part that matters. opML is safe only as long as at least one honest party is watching and willing to challenge. No cryptography guarantees the result; economics and vigilance do. Keep that sentence in mind, because it comes back to bite the most interesting thing Ora has built. Now the applications.

Three products, one settlement layer

Ora, founded in September 2022 as Hyper Oracle by former Google and TikTok infrastructure engineer Kartin Wong, with ex-Ethereum Foundation zkML researcher Cathie So as chief scientist, raised $20 million in June 2024 from Polychain, HF0 and HashKey Capital (CoinDesk). Wong framed the pitch at the time as wanting to “unlock the design space for AI dapps.” opML is the primitive; the design space is everything you can build once on-chain code is able to trust an AI output.

Ora has built three things on it, and they stack:

• OAO, the Onchain AI Oracle: a contract that lets any smart contract request an inference (an image, a chat completion, a score) and receive a verifiable result. This is the base layer, the thing every other product calls.
• opAgent: a framework for autonomous agents that live entirely on-chain, controlled by contract logic rather than private keys, with opML verifying each action they take.
• IMO, the Initial Model Offering: a scheme for tokenizing an AI model itself, so that the fees the model earns flow back to token holders.

The first is infrastructure. The second is a bet on the agent economy. The third is the most ambitious and the least covered, and it is where this piece spends most of its time, because it is the clearest test of whether opML enables something genuinely new or just re-plumbs things that already worked without it.

OAO: the oracle you can watch but can’t quite read

OAO is Ora’s flagship. It deploys at the same address, 0x0A0f4321214BB6C7811dD8a71cF587bdaF03f0A0, across more than a dozen chains including Ethereum, Optimism, Arbitrum, Base, Polygon, Linea, Mantle and BSC (github.com/ora-io/OAO). A developer calls the contract with a model ID and a prompt, opML nodes run the inference off-chain, and after the challenge window a callback delivers the verified output back to the calling contract. Fees run from 0.01 to 1.00 ORA per call. On Arbitrum One the live catalog is small and concrete: Stable Diffusion, Llama3 8B Instruct, and two OpenLM variants (Arbitrum docs).

Two honest caveats sit on top of the marketing. The first is the closed-model gap. opML’s fraud proof only means something if a challenger can independently re-run the model and reproduce the same trace. That works for open-weight models such as Llama and Stable Diffusion. It does not work for a proprietary API like GPT-4o, which OAO can relay but cannot verify in the same sense, because there is no weights file for a watcher to replay. When Ora lists GPT-4o alongside Llama, it is offering two different trust guarantees under one interface, and only one of them is really opML.

The second is the usage question. Ora points to a public Dune dashboard that tracks OAO request counts by chain and by model, plus revenue and callback gas. The requests are on-chain, so anyone can count them. What is missing is the kind of number a healthy protocol puts in a press release: millions of paid inferences a month, a growth curve, marquee integrations moving real volume. Set that against the trusted-hardware camp, where Phala Network reports serving billions of confidential model tokens a day through OpenRouter, and the contrast is hard to ignore. opML has a working oracle. It has not yet shown a busy one.

The big idea: turn a model into an asset

Here is where Ora tries to do something no verification layer has done before: make the AI model itself a tradable, income-producing asset. The mechanism is the Initial Model Offering, and Cointelegraph’s coverage of the 2024 raise described the ambition plainly, that Ora is tokenizing AI models so buyers can “own and share revenue generated by an AI model.”

Think about what that requires. If you want to sell shares in a model’s future earnings, three things have to be true. Someone has to actually run the model and pay for it. The fees have to be counted honestly and split among holders without a trusted middleman. And the outputs have to be attributable to that specific model, so revenue can be assigned rather than faked. The first is a demand problem. The second and third are exactly what opML is for: a fraud-proofable record that this model, not some cheaper substitute, produced this output, and therefore earned this fee.

That is the elegant part of the design. A revenue-share token on an AI model is only as trustworthy as the accounting behind it, and on a normal server the accounting is a private database the operator controls. opML replaces that database with a challengeable on-chain record. In principle it converts “trust the model provider’s revenue report” into “verify the model’s revenue on-chain.” Whether the principle survives contact with real usage is the open question, but the architecture is genuinely novel, and it is the strongest single argument anyone has made for why opML needs to exist at all.

ERC-7641 and ERC-7007: the two halves of a tokenized model

Ora splits a tokenized model into two token standards, and the split is worth understanding because it maps onto two separate revenue streams.

ERC-7641 is the ownership layer, an ERC-20 extension that builds revenue sharing into the token itself: hold it, and you can claim a proportional share of a communal pool that fills up every time the model is used. ERC-7007 is the output layer, and unlike ERC-7641 it is a finalized Ethereum standard. Authored by Ora’s founding team, including Cathie So, KD Conway and Kartin Wong, and created in May 2023, ERC-7007 reached Final status as a Standards Track ERC (eips.ethereum.org). It extends the ERC-721 NFT standard to bind a generated image or video to the exact prompt that produced it, and it supports both proof systems: zkML validity proofs and opML fraud proofs. It even indexes tokens by prompt hash rather than sequential ID, so the same prompt cannot be minted twice under identical parameters.

The provenance angle here is not trivial in a year when the hardest problem in online identity is telling real from synthetic. The same verification that lets a marketplace prove an image came from a licensed model is a defense against the flood of unattributed synthetic media reshaping everything from fraud to identity checks, a fight HOGE Wire traced in its look at crypto KYC and the deepfake arms race. A cryptographically attributable output is the opposite of a deepfake: it arrives with a receipt.

Ora’s first live example of the whole stack is $OLM, a token tied to OpenLM, an open language model, launched as the protocol’s inaugural Initial Model Offering. It is the proof of concept for the entire thesis: an open model, its inference metered through OAO, its revenue routed to ERC-7641 holders, its outputs mintable as ERC-7007 assets.

Why opML is the load-bearing wall under all of this

Strip away the token standards and the question is simple: why can’t you just do this on a normal cloud? You can, badly. A company could sell revenue-share tokens in a model it hosts on ordinary servers and promise to pay out honestly. Investors would be trusting a private revenue report, which is to say trusting the company, which is to say holding an IOU dressed up as a token. The entire pitch of putting it on-chain is to remove that trust, and opML is the only part of the stack that does.

The fraud-proof record is what makes the fee stream enforceable. When OAO logs an inference, the commitment to that computation can be challenged; if an operator tries to under-report usage, or quietly swaps in a cheaper model while billing for the expensive one, an honest watcher can dispute the record and the operator loses a stake. That is the difference between a revenue-share token backed by cryptographic accounting and one backed by a pinky promise.

But the load-bearing wall has two known cracks, and an IMO leans directly on both. The closed-model gap means the design only fully works for open-weight models; a tokenized GPT-4o would inherit the weaker attested-not-verified guarantee. And opML’s honest-watcher assumption carries a classic flaw known as the verifier’s dilemma: if fraud is rare, there is little economic reason to spend money running a watcher, so watchers may not exist precisely when they are needed. Truebit tried to patch this with deliberately introduced errors and jackpot rewards to keep verifiers paid and alert; Ora’s public design has no deployed equivalent. A revenue-share asset whose integrity depends on someone volunteering to watch is only as safe as the assumption that someone actually does. This is the same accountability gap the audit business keeps running into, where a clean report is only as good as the incentive to look hard; HOGE Wire’s profile of Halborn’s move to Wall Street is a study in how expensive genuine verification turns out to be.

opAgent: agents that own themselves

The second application is opAgent, which Ora describes as an Onchain Perpetual Agent framework and which the company said went live on 31 January 2025 (Decrypt). The pitch aims squarely at the weaknesses of ordinary AI agents: they run on someone’s server, they can be switched off, and they are not provably autonomous. opAgent’s answer is to move the agent’s control logic into a smart contract and give it an AI agent wallet with no private key at all, so there is no seed phrase to steal and no operator who can quietly seize the funds.

opML is what makes the agent’s decisions trustworthy. Every action an agent takes, a swap on a DEX, a token transfer, a contract deployment, can be tied back to a verifiable inference: the model said to do this, and here is the fraud-proofable record that it did. Ora describes two modes, a Singleton agent that operates on its own and a Swarm of specialized agents that collaborate, plus a companion service called Resilient Model Services for the off-chain side.

The vision rhymes with where the rest of the industry is heading. Exchanges are racing to give AI agents trading rails, a contest HOGE Wire tracked in the AI agent race between Coinbase, Binance, Kraken and OKX. An agent that can prove its own reasoning before it moves money is a genuinely useful primitive in that world, and it is arguably the most defensible use case opML has. The problem is not the concept. The problem is the clock.

The latency tax that fights the agent dream

opML’s core mechanism, the challenge window, is also its core liability for anything that needs to happen now. A result is only final after the dispute period elapses, and that period has to be long enough for honest watchers to notice and act, which means minutes at least, not milliseconds. For a revenue-accounting record that settles in the background, that is fine. For an agent that wants to react to a moving market, it is a tax on every decision.

This is not a fringe complaint. A December 2025 paper on Optimistic TEE-Rollups (arXiv 2512.20176), from a team including Aaron Chan and colleagues, puts it bluntly: optimistic approaches like opML “impose prohibitive dispute windows, preventing real-time interactivity.” The authors frame the whole field around a verifiability trilemma, the claim that a system cannot simultaneously deliver high computational integrity, low latency and low cost. opML buys integrity and cost by spending latency. An agent economy that runs on it inherits that trade: it can prove what it did, eventually, but it cannot act instantly and prove instantly at the same time. The same delay quietly breaks composability, because a second contract cannot safely build on a result that is not yet final.

Ora knows this, which is why its own research has drifted toward hybrids. The company co-authored a paper on opp/ai (arXiv 2402.15006), which stitches zero-knowledge privacy onto opML’s efficiency, and the broader 2026 trend is a scramble to combine fraud proofs, zero-knowledge spot checks and trusted hardware to escape the trilemma. The tell is that opML in its pure form is increasingly a component, not a product.

So, does anyone actually use it?

Here the story turns from architecture to adoption, and the numbers get thin. Ora has shipped a working oracle on more than a dozen chains, a finalized token standard, a tokenized model, and a live agent framework. What it has not shown, publicly and in a way that would settle the argument, is volume. There is no headline figure of millions of paid OAO inferences, no disclosed count of active opAgents managing real capital, no market cap for $OLM that signals a functioning market in model revenue. The Dune dashboard is public precisely so anyone can check; the enthusiasm, judging by everything downstream of it, is not there.

The token is the bluntest evidence. A protocol whose applications were finding product-market fit would not trade at a $792,000 market cap on $305 of daily volume. That is not the price of failed technology, because the technology works. It is the price of technology nobody is paying to use yet. And this is a pattern across the entire verifiable-compute sector, so it is worth seeing the whole picture before drawing any conclusion about opML specifically.

The whole sector is underwater, but ORA is the deepest

Every project that promises to make AI verifiable, whether through fraud proofs, zero-knowledge math, trusted hardware or staking, has shipping technology and a token trading at a fraction of its peak. The gap between what these networks can do and what their tokens are worth is the defining feature of the sector in 2026.

Figures from CoinGecko, 22 August 2026. Two things stand out. ORA is the smallest by an order of magnitude, a sub-million-dollar market cap where its peers sit in the tens or hundreds of millions. And the least crypto-native trust models are holding up best: EIGEN, backed by restaked ETH and a crypto-economic guarantee, is the largest of the four, even as the restaking sector it belongs to works through its own great unwinding. The market, for now, is paying more for economic security and for cryptographic proof than for optimistic fraud proofs. opML’s pragmatic middle position, cheaper than zkML and more trustless than a TEE, has not earned a premium; it has produced the smallest token in the group.

The securities question a revenue-share token can’t dodge

There is a reason most crypto tokens go to great lengths to look like utilities rather than investments, and IMO walks straight into it. An ERC-7641 token is, by its own description, a claim on a revenue pool: you buy it, other people’s effort generates fees, and you redeem a proportional share. Read that sentence next to the four prongs of the Howey test, the US Supreme Court standard the SEC leans on to decide what is a security (an investment of money, in a common enterprise, with an expectation of profit, derived from the efforts of others), and the overlap is hard to miss. A gas-like utility token has a fighting argument that it is consumed rather than invested. A token whose entire selling point is sharing the revenue of an AI model has a much weaker one.

US regulators have not addressed these instruments directly. The joint SEC and CFTC interpretation issued on 17 March 2026 named sixteen crypto assets as digital commodities, Bitcoin, Ether, Solana and Chainlink’s LINK among them (SEC Release 33-11412), but it said nothing about AI-oracle or verifiable-compute infrastructure tokens, let alone revenue-share tokens on individual AI models. That silence leaves ORA itself in a case-by-case gray zone, and it leaves IMO-style RevShare tokens in a far more exposed one. A revenue-sharing instrument sits closer to the center of securities law than almost anything else in crypto, and shipping it as a permissionless standard does not change the economic substance a regulator would look at. The wider fight over who is even responsible when autonomous on-chain code breaks the rules runs through HOGE Wire’s coverage of DeFi compliance and the fight over the gatekeeper; a real market in tokenized models would land in the middle of it. None of this is legal advice, and none of it means enforcement is coming tomorrow. It means the single most novel thing opML enables is also the thing most likely to draw a regulator’s attention if it ever grows large enough to matter.

The verdict: a real primitive still hunting a market

Add it up. opML is a legitimate piece of engineering, and IMO is the most creative use anyone has proposed for it, a way to turn a model into an income-producing asset backed by cryptographic accounting instead of a trusted operator. The ERC-7007 standard is finalized, the oracle is live on a dozen chains, the agent framework shipped, the first model is tokenized. On the supply side, Ora has delivered what it promised.

The demand side is empty. There is no evidence of the volume that would make a model-revenue market real, the token trades like an afterthought, opML’s latency undercuts the agent use case that might have driven adoption, its honest-watcher assumption remains an unpatched economic hole, and the one application that could genuinely matter, tokenized model revenue, is also the one most likely to invite a securities challenge. Ora built the design space Kartin Wong promised to unlock. Almost nobody has moved in. Vitalik Buterin anticipated the difficulty in his 2024 essay on crypto and AI, filing the use of AI as “the rules of the game,” where on-chain money hinges directly on a model’s output, under the category to “tread very carefully” in (vitalik.eth.limo); you only pay for a proof when being wrong costs more than the proof does.

What would change the picture is not more technology; it is a single application with real users. A model whose IMO token traded on genuine revenue, an opAgent managing meaningful capital, an OAO integration doing millions of paid inferences a month. Any one of those would reprice the whole thesis overnight. Until then, opML remains the answer to a question the market has not quite asked: it can prove the AI did its job, for whoever turns out to need that badly enough to pay for it.

Frequently Asked Questions

What is opML in simple terms?

opML, or optimistic machine learning, is a way to run an AI model off-chain and then prove on-chain that it ran honestly. Instead of re-running the model on the blockchain, which would be far too expensive, opML posts the result with a cryptographic commitment and assumes it is correct unless someone challenges it within a set window. If there is a dispute, a fraud proof pinpoints the exact step where the two sides disagree and settles it cheaply on-chain. It borrows the whole approach from optimistic rollups.

What is an Initial Model Offering (IMO)?

An Initial Model Offering is Ora’s mechanism for tokenizing an AI model so that its earnings flow to token holders. It pairs an ERC-7641 RevShare token, which represents ownership and pays holders a proportional share of the fees the model earns each time it is used on-chain, with ERC-7007 tokens, which represent individual verifiable outputs such as images or videos. Ora’s first IMO tokenized an open language model called OpenLM under the ticker $OLM.

Is the ORA token a good investment?

HOGE Wire does not give investment advice, but the facts are worth knowing. As of 22 August 2026, ORA trades around $0.004764 with a market cap near $792,000 and daily volume near $305, down roughly 99.9% from its November 2024 high, according to CoinGecko. The technology works, but the token reflects almost no measurable adoption of the applications built on it, and revenue-share tokens raise unresolved securities questions in the US. Weigh any position accordingly.

How is opML different from zkML and TEEs?

All three make AI verifiable, but they trust different things. zkML uses zero-knowledge cryptography to prove correctness mathematically, which is trustless but computationally heavy. TEEs use trusted hardware, such as secure chips, which is fast but requires trusting the chip vendor. opML uses economics: it assumes honesty and relies on a challenge window plus fraud proofs, which is cheap but adds latency and needs at least one honest watcher. Each buys two of integrity, low latency and low cost at the expense of the third.

Does opML work with closed models like GPT-4o?

Only partially. opML’s fraud proof depends on a challenger being able to independently re-run the model and reproduce its output. That is possible with open-weight models like Llama or Stable Diffusion, but not with a closed API like GPT-4o, whose weights are private. Ora’s oracle can relay results from closed models, but those are attested rather than fraud-proof verified, which is a weaker guarantee than opML provides for open models.

By Marcus Okafor, senior AI and crypto correspondent at HOGE Wire.