opML Bet on Inference. Verifiable AI Moved to Training

The money moved upstream

On 7 October 2026, Nous Research confirmed a $90 million Series B at a $1.5 billion valuation, money raised to push an open-source AI agent that it trains on a network of scattered GPUs rather than inside one data center (TechCrunch). Three months earlier, Prime Intellect closed a $130 million Series A at a $1 billion valuation to aggregate GPUs across the internet for large training runs (SiliconANGLE). On the same 2026 calendar, the token that was meant to be crypto’s answer to verifiable AI, ORA, changes hands at about $0.002480 for a market value near $412,000 on roughly $28 of daily trading volume (CoinGecko).

Those three figures sketch where the trust problem in AI actually lives. opML, short for optimistic machine learning, was crypto’s most quotable attempt to make an AI model prove it had computed honestly. It was aimed squarely at inference: ask a model a question on-chain, publish the answer, and let anyone dispute it. Two years on, the teams raising real money are not checking answers. They are training the models, and they are learning that a trustless check earns its keep not in the cheap, repeatable act of inference but in the expensive, one-shot act of training. opML’s core idea did not fail. It was pointed at the wrong floor of the building.

This piece follows that idea upstream. It traces how the fraud-proof, challenge-window design opML made famous is quietly reappearing inside the decentralized-training stack, where the stakes are higher and the economics make more sense, and why the token that carried the opML banner became a rounding error while the problem it was built for moved somewhere else.

opML in sixty seconds

opML borrows its name and its nerve from optimistic rollups. The founding paper by KD Conway, Cathie So, Xiaohang Yu and Kartin Wong places an “interactive fraud proof protocol, reminiscent of the optimistic rollup systems” at its center (arXiv). A node runs the model off-chain, posts the result together with a cryptographic commitment, and a challenge window opens. If a watcher recomputes the same inference and disagrees, the two sides play a bisection game that narrows the dispute down to a single low-level operation, which a smart contract can then referee for a few cents. No challenge inside the window, and the answer is treated as final.

The pitch was efficiency. Ora, the company that commercialized opML through its Onchain AI Oracle, argued the approach could run a 7-billion or 13-billion parameter model on an ordinary machine with no GPU, and cut the overhead of a zero-knowledge proof by a factor it put in the millions (Ethereum Research). The oracle itself lives at a single shared contract address across more than a dozen chains, taking a request for a model such as Llama or Stable Diffusion, running it off-chain, and calling the answer back to the requesting contract once the window closes (GitHub). The trade was explicit: give up instant cryptographic certainty, accept a waiting period, and in return verify models far larger than a zero-knowledge circuit could touch.

Verification in this sense is not an audit you pay a firm to sign off once; it is a check any sufficiently motivated party can run themselves, a more fragile kind of trust than the audit-firm model crypto leaned on through its worst quarters. Everything in opML hangs on that one assumption: that somebody is watching.

Inference was the easy floor to guard

Here is the awkward thing about verifying inference: it is the computation you can most easily just redo. If a single forward pass costs fractions of a cent, the honest watcher who is supposed to catch a lie has almost no economic reason to bother, and the liar has little to gain from a one-off fib any observer can cheaply expose. The security model leans on at least one diligent, equipped watcher being present at all times, yet the reward for watching is close to zero. That is the verifier’s dilemma, and inference is where it bites hardest.

The second problem is speed. A challenge window is latency by design, and the 2026 application everyone wanted, autonomous agents transacting on-chain, cannot wait minutes to act on a verified answer. A research team proposing Optimistic TEE-Rollups put the objection bluntly, writing that optimistic approaches “impose prohibitive dispute windows, preventing real-time interactivity” (arXiv). When Ora raised $20 million in 2024, founder Kartin Wong said the money was to “unlock the design space for AI dapps” (CoinDesk). The design space was real. The demand for paying to verify a cheap, repeatable computation was not.

The usage data points the same way. An independent on-chain review of Ora’s oracle counted only a few thousand lifetime requests against circulated marketing figures in the hundreds of thousands, and could find no record of a fraud proof ever being filed (ownyourmind.ai). Zero challenges is consistent with flawless honesty; it is just as consistent with nobody checking. The ORA token’s slide toward illiquidity is the market settling that ambiguity in the least flattering way.

In 2026, decentralized training grew up

While verifiable inference stalled, a different corner of crypto-AI started shipping. Prime Intellect trained INTELLECT-1, a 10-billion parameter model, across GPUs in five countries, then followed with INTELLECT-2, a 32-billion parameter reasoning model trained through globally distributed reinforcement learning with permissionless contributors (Prime Intellect). Its stack splits the job into named parts: PRIME-RL for the training loop, SHARDCAST for pushing updated weights out to workers, and TOPLOC for checking that the work came back honest. Nous Research built Psyche, a training network that coordinates scattered GPUs through the Solana blockchain and uses compressed gradient updates to make training over ordinary internet links practical (Nous Research); its Hermes models are now post-trained on that network rather than a rented cluster.

The contrast with opML’s commercial fate is stark. Prime Intellect’s Series A valued it at $1 billion, with the venture arms of Nvidia, Intel and Dell on the cap table; Nous reached $1.5 billion days before this was written. These are not governance-token memes but priced bets by strategic investors that aggregating and coordinating compute to train models is a real business. The verification token built for the adjacent problem, meanwhile, is worth less than a used car. Money, it turns out, follows the computation that is hard to reproduce.

Why training is where a trustless check earns its keep

Training inverts every economic assumption that made inference verification a weak market. A frontier run is not a fraction-of-a-cent event you can casually repeat to check; it is a multi-week, multi-million-dollar process you cannot un-bake once the weights are set. If you are paying thousands of anonymous GPUs to contribute gradients or reinforcement-learning rollouts, you need to know two things before you fold their work into the model: that they actually did the computation they billed for, and that they did it correctly rather than slipping in poisoned or lazy updates.

The failure modes are concrete. A free-riding node can submit recycled or random gradients and collect rewards for compute it never spent. A malicious one can poison the run, nudging the model toward a backdoor or simply degrading it. A Sybil operator can split one machine into many identities to farm emissions. Every one of these is a verification question, and every one carries real money, which is exactly the condition inference lacked.

This is the setting where “assume honest, punish the liar” finally has teeth. A contributor who fakes work can be denied payment and have a staked bond slashed, so honesty becomes the cheapest strategy and a single honest checker can protect the whole run. The challenge-window latency that killed opML in real-time inference barely matters here, because a training run already unfolds over days. Optimistic verification is a poor fit for the fast, cheap thing and a natural fit for the slow, expensive thing. opML simply launched on the wrong one.

Verde: opML’s verification game, aimed at training

The clearest evidence that the idea moved upstream is Gensyn’s Verde. Gensyn runs a network for decentralized machine-learning compute, and Verde is how it decides whether a remote GPU told the truth. The mechanism will sound familiar: the system dispatches a job to more than one provider, and, in Gensyn’s own description, “if two compute providers disagree on an ML execution, a referee can determine which provider is honest by executing a single operation rather than running the entire job,” after the referee runs “a bisection game to pinpoint the first operation where they disagree” (Gensyn). That is opML’s fraud proof, almost beat for beat, aimed at computation on untrusted nodes instead of at a public oracle.

The framing differs in a telling way. Gensyn traces Verde to a cryptographic technique it calls refereed delegation rather than to the optimistic-rollup heritage opML cites, and it dispatches each job to several providers at once instead of relying on a single submitter and a crowd of hoped-for watchers. But the guarantee has the same shape. The Verde paper, whose authors include the cryptographer Joseph Bonneau, promises a design that lets “a computationally limited client to delegate a program to multiple untrusted compute providers, with a guarantee of obtaining the correct result if at least one of them is honest” (arXiv). One honest party, a cheap referee, a dispute narrowed to a single operation: these are the exact bones of Truebit and opML, reassembled for a market where the computation under dispute is worth catching.

Nor is Verde a research toy. In a late-2025 deployment note, Gensyn said the system and its companion library already run in production inside Judge, its model-evaluation product, which makes it one of the first optimistic-style ML verifiers deployed in the wild rather than merely described in a paper (Gensyn). The irony is hard to miss: the clearest working deployment of opML’s idea is not opML.

RepOps: fixing the determinism opML waved past

Moving the verification game upstream exposed a problem the inference pitch had mostly asserted away. A bisection game only works if two honest parties who run the same step land on the same answer, bit for bit. GPUs do not naturally cooperate here: floating-point arithmetic is not associative, parallel reductions reorder operations, and the same model can produce subtly different numbers on different hardware or drivers. In an inference demo you can wave at this. In a dispute that hinges on whether operation number four million matches, a one-bit difference between two honest machines breaks the referee entirely.

Gensyn’s answer was to build the determinism rather than assume it. Its RepOps library, it says, “ensures that the same ML operation produces identical bitwise results, regardless of GPU or environment” (Gensyn). That is unglamorous systems engineering, and it is precisely the part the optimistic-inference story tended to skip. It is also a quiet rebuke to the idea that opML was ever close to production-grade verification: the reproducibility that makes a fraud proof meaningful had to be constructed, operator by operator, by the people who actually needed it to hold.

Not everyone bets optimistic: the primitive bake-off

Optimistic verification is not the only horse in the training race. Prime Intellect leans on TOPLOC, which it describes as a lightweight scheme for checking the rollouts that untrusted inference workers send back, using locality-sensitive hashing to spot tampering or numerical drift rather than replaying the model or wrapping it in a zero-knowledge proof (Prime Intellect). It is cheaper than a fraud-proof game and weaker in its guarantees, a statistical tripwire rather than an adversarial court. That gap captures the live argument inside decentralized training: how much certainty is worth how much cost, when the thing you are guarding is a gradient rather than a payment.

The other two options bound the space. Zero-knowledge proofs give the strongest guarantee but remain too heavy for large models, so they tend to protect small, high-value outputs rather than whole training runs. Trusted hardware is fast and near-native but asks you to trust a chip vendor and the enclave’s side-channel posture, which is why it dominates latency-sensitive inference rather than long adversarial training jobs. The table below lays out the four families as they look in late 2026.

The honest asterisk: INTELLECT-3 plugged back in

A story this tidy deserves its counter-evidence, and decentralized training supplies plenty. When Prime Intellect built its largest model, INTELLECT-3, a 106-billion parameter mixture-of-experts with roughly 12 billion active parameters, it did not spread the run across the open internet. As reported by Implicator, the late-2025 training run happened on a centralized cluster of 512 Nvidia H200 GPUs wired with 400-gigabit InfiniBand, about $2.2 million of compute over two months. The reason was physics, not ideology: “When you’re running asynchronous RL at this scale, policy weights need to reach inference workers mid-generation,” the report notes, and “that kind of tight synchronization demands InfiniBand, not commodity broadband.”

That matters for the verification thesis in two ways. First, the frontier run still centralized, which means the trustless-verification machinery was exercised mostly on the inference-style rollouts feeding the training loop, not on the core gradient computation. Second, it is a reminder that decentralized training is a spectrum, not a switch: the networks verifying untrusted contributors today are doing it for reinforcement-learning rollouts, fine-tuning and smaller base runs, not yet for the biggest pretraining jobs. The optimistic check has found a real home. It has not conquered the whole house.

Who watches the trainer?

Carrying opML’s mechanism upstream also carries opML’s unsolved social problem. Optimistic security is only as good as the assumption that at least one honest, capable party is watching and willing to challenge. On Ora’s public oracle, as noted, there is no record of a fraud proof ever being filed, which tells you either that nothing was ever wrong or that nobody looked. A training network has to answer the harder version of that question: who runs the referees and validators, what hardware do they keep ready, and why would they stay honest once the novelty fades.

The answer training networks reach for is the one the rest of crypto reaches for: bonds and slashing, so that watching pays and cheating burns capital. That turns verification into an operator business with real supply-side economics, the same shape as the operator economics that decide who actually runs restaking infrastructure. It is a better-funded version of opML’s honest-watcher gamble, but it is still a gamble on incentives rather than a cryptographic guarantee, and it will be tested the first time a large, well-resourced contributor finds it more profitable to cheat than to comply.

The scoreboard: compute got priced, verification did not

Step back to the tokens and the gap is hard to miss. Every liquid asset built specifically to sell verification trades far below its peak, while the one broad decentralized-AI network token and the private training companies command real value. The market is paying handsomely to aggregate and sell compute, and to train models, and paying almost nothing for verification sold on its own. The snapshot below uses prices from 12 October 2026.

Token figures come from CoinGecko pages for ORA, Succinct, Phala, EigenCloud and Bittensor; the private valuations are from the Prime Intellect and Nous funding rounds cited above. ORA’s daily volume is so thin that its quote is barely a price at all, so read it as a verdict on demand rather than a tradable mark.

The hybrids are converging

If no single primitive wins outright, the obvious move is to blend them, and that is exactly what the 2026 research record shows. Three hybrids tell the story.

• opp/ai stitches zero-knowledge proofs for privacy onto opML for efficiency, written by some of the same authors behind the opML paper (arXiv).
• zk-OPML uses zero-knowledge proofs to shrink opML’s dispute window and scale it to larger models, trimming the latency that made the pure optimistic version unusable in real time (Springer).
• Optimistic TEE-Rollups run a trusted-hardware enclave for fast provisional finality, then layer fraud proofs and occasional zero-knowledge spot-checks on top, claiming most of a centralized system’s throughput at roughly seven cents a query (arXiv).

Gensyn’s Verde-plus-RepOps combination is itself a convergence: an optimistic dispute game welded to hard reproducibility engineering. The lesson is that opML was never going to be a standalone product so much as one ingredient. The cost argument underneath explains why. Vitalik Buterin noted early that zero-knowledge proving of neural networks can carry “around 200x” overhead on non-linear layers and that a full proof “can take hours,” which is why builders reach for cheaper optimistic or hardware checks and reserve heavy cryptography for the pieces that truly need it (vitalik.eth.limo). a16z’s Elena Burger made the same point from the modeling side, observing that zero-knowledge proofs “today can’t represent 32-bit floating point operations in the necessary arithmetic circuit format without massive overheads” (a16z crypto). On the hardware flank, Phala now serves tens of billions of confidential tokens a day, more than 30 billion at recent counts, a scale no optimistic oracle has approached (Phala).

Inference verification is not dead, just repriced

None of this means verifying inference was a mistake to pursue, only that it was mispriced as a standalone business. Autonomous agents really will need a way to prove that the model behind a decision computed what it claimed before they sign a transaction or move funds, a need that grows as wallets start spending on their users’ behalf. The honest read of 2026 is that this demand is still mostly ahead of us. A TRM Labs analysis of on-chain agent payments found that of roughly $52.7 million settled through the x402 standard, only 0.6% to 7.5% of the value looked genuinely agentic, with the rest being ordinary automation a plain script could run (PYMNTS).

So opML’s original turf did not vanish; it shrank to a niche with specific contours. Optimistic inference verification makes sense when the output is open-weight and reproducible, when the stakes justify a watcher, and when a delay between answer and finality is acceptable. That is a real slice of the map. It is just a far smaller slice than a general-purpose verification layer for all of on-chain AI, which is what the token was priced to become.

Oracles, regulators, and who is liable

There is a sharper edge to optimistic verification that the training framing should not bury. An opML oracle that finalizes an answer because nobody challenged it in time is, functionally, an oracle that can be wrong on the record. If that answer feeds a lending protocol or settles a prediction market, a stale or manipulated output becomes someone’s loss, the same class of risk crypto already knows from oracle extractable value and the exploits it quietly legalizes. The optimistic assumption is not just a technical footnote; it is a liability question dressed as a latency parameter.

Regulators have not drawn that line yet. The joint SEC and CFTC interpretation issued in March 2026 named sixteen digital assets as digital commodities but said nothing about AI-oracle or verifiable-compute infrastructure, leaving tokens like ORA in a case-by-case gray zone under the usual Howey analysis (SEC). The sharper exposure may sit with tokens that promise a cut of a model’s revenue rather than a gas-like utility, which raises the same questions aired in debates over who is allowed to keep the yield. Buterin’s own warning about using AI as “the rules of the game,” where a manipulated model becomes an incentive to attack rather than a neutral referee, reads as a caution written for exactly this moment. Verifiable AI is turning load-bearing before anyone has decided who answers when the proof does not fire.

Frequently Asked Questions

What is opML (optimistic machine learning)?

opML is a way to verify AI computation on a blockchain without re-running it in full. A node runs the model off-chain, posts the result with a cryptographic commitment, and opens a challenge window. If a watcher recomputes the inference and disagrees, a bisection game narrows the dispute to a single operation that a smart contract can cheaply referee. If no one challenges in time, the result is treated as final. The design is modeled on optimistic rollups and was pioneered by the company Ora.

How is opML different from zkML and TEEs?

zkML produces a cryptographic proof that is costly to generate but cheap to verify and gives the strongest guarantee, though it struggles with large models. A TEE relies on a hardware chip attesting that code ran untampered, which is fast but trusts the chip vendor. opML assumes honesty and leans on a challenge window plus fraud proofs, which is cheap and scales to large models but adds latency and depends on at least one honest watcher. Each option trades certainty, speed and cost differently.

Is the ORA token dead?

ORA is not formally shut down, but by October 2026 it trades near $0.0025 for a market value around $412,000 on only a few dollars of daily volume, more than 99.9% below its 2024 peak. At that level of volume the price is barely meaningful. The technology still exists and the ideas live on in other projects, but the token itself is effectively a rounding error.

Can optimistic verification be used for AI training, not just inference?

Yes, and that is where it looks most useful. Gensyn’s Verde system uses the same refereed, bisect-the-dispute approach to check training and inference performed by untrusted GPUs, and it pairs the dispute game with a reproducibility library so honest machines agree bit for bit. Training is expensive and hard to redo, which gives an honest checker a real reason to watch and a cheater real money to lose, the opposite of the weak incentives around cheap inference.

Who is actually using decentralized AI training in 2026?

The most visible names are Prime Intellect, which trained the INTELLECT series and raised a $130 million Series A at a $1 billion valuation; Nous Research, whose Psyche network trains its Hermes models and which reached a $1.5 billion valuation; and Gensyn, which runs a verifiable compute network. Their largest training runs still sometimes use centralized clusters for speed, so the field is maturing rather than finished.

By Marcus Okafor, who covers AI and crypto infrastructure for HOGE Wire.