opML’s Squeeze: Caught Between Cheaper Proofs and Faster Chips

The bet opML made

opML, short for optimistic machine learning, was supposed to be the pragmatic one. While the zero-knowledge crowd chased cryptographic proofs that could take hours to generate, and the hardware camp asked users to trust a chip vendor’s attestation, opML offered a cheaper third path: run the model off-chain, post the result on-chain with a small commitment, and let anyone challenge it during a dispute window. If nobody objects, the answer stands. It borrows its entire security posture from optimistic rollups like Arbitrum and Optimism, and that familiarity was the point. You could ship it in 2024 without waiting for a cryptographic breakthrough.

Two years later, the sensible bet is getting squeezed. On one flank, zero-knowledge machine learning (zkML) has started doing the thing everyone said it could not: proving real language models. On the other, trusted execution environments (TEEs) have gone from lab demos to serving tens of billions of tokens a day. opML sits in the middle, and the middle is getting thin.

The scoreboard is blunt. ORA, the token of Ora Protocol, the team that coined opML and still runs its flagship deployment, trades around $0.0058 with a market capitalization under $1 million and roughly $360 of daily volume, per CoinGecko. In the same week, Phala Network’s confidential-compute fleet reported processing 23.1 billion model tokens in a single day through its TEE workers, according to figures on Phala’s own site. One of these is a research-grade idea with a near-dead token; the other is production infrastructure. Both claim to answer the same question: how do you know the AI actually ran the model it says it ran?

This is not an obituary. opML is a genuinely clever design with a real niche. But the story of 2026 is that its two neighbors moved, and opML’s structural problems (a dispute window that breaks real-time use, a challenger nobody pays to show up) did not get fixed. This piece is about the squeeze: why it happened, and what it would take for opML to climb back out.

What opML actually does

opML applies a very old blockchain trick to a very new problem. The old trick is the verification game. Instead of forcing every node to re-run an expensive computation, you let one party assert a result and post a bond, then give everyone else a window to object. If a challenger appears, the two sides run an interactive bisection: they repeatedly narrow the disagreement down to a single computational step, and only that one tiny step gets replayed on-chain to decide who is right and who forfeits their bond. The pattern comes from Truebit, described by Jason Teutsch and Christian Reitwiessner in their 2017 paper (people.cs.uchicago.edu), years before optimistic rollups made it famous.

opML applies that game to model inference. The formal write-up, opML: Optimistic Machine Learning on Blockchain (arXiv 2401.17555), is explicit that the design is “reminiscent of the optimistic rollup systems.” A node runs the model off-chain, publishes the output plus a Merkle commitment to the full execution trace, and the result is treated as final unless someone disputes it before the challenge window closes. Disputes are resolved inside a Fraud Proof Virtual Machine (FPVM), a constrained, deterministic environment where any single step of the computation can be replayed and checked.

The clever engineering is the split between single-phase and multi-phase proving. In single-phase mode every step runs inside the slow, GPU-free VM, which is simple but brutally slow for a large model. Multi-phase mode lets undisputed computation run natively on a GPU and only drops into the VM for the one bisected step that is actually in dispute, cutting the Merkle-tree overhead from roughly O(mn) to O(m+n). Ora’s opML repository reports MNIST-scale inference in about two seconds with a full challenge cycle around two minutes on a local testnet, and claims that a 7-billion-parameter LLaMA model (roughly 26GB) can run on an ordinary PC with no GPU. Those are self-reported benchmarks, not independently audited numbers, and it is worth keeping that caveat in mind.

Here is the point to hold onto: opML never proves the computation was correct in the cryptographic sense. It makes the computation cheaply checkable and bets that someone will bother to check. That bet is the source of both its low cost and its central weakness.

Why opML looked like the smart bet in 2024

To understand the squeeze, you have to remember why opML looked so attractive in the first place. In early 2024 the honest assessment of verifiable AI was bleak. Zero-knowledge proofs, the only approach offering pure cryptographic trustlessness, were catastrophically slow on anything resembling a real model.

Vitalik Buterin laid out the math in his January 2024 essay on crypto and AI (vitalik.eth.limo). Verifying an Ethereum block takes a few hundred milliseconds, he noted, but generating a ZK-SNARK of that same computation could take hours; for the non-linear operations that dominate neural networks, he pegged the overhead at roughly 200 times. Elena Burger of a16z put the practical version bluntly in the firm’s write-up on machine learning and zero-knowledge proofs (a16zcrypto.com): “zero-knowledge proofs today can’t represent 32-bit floating point operations in the necessary arithmetic circuit format without massive overheads.” To fit a model into a ZK circuit, you had to quantize it down to low-precision integers, trading away accuracy to buy provability.

Against that backdrop, opML’s pitch was hard to argue with. No proving overhead at all in the happy path, because there is no proof; you only pay for on-chain work when a dispute actually happens, which the designers assumed would be rare. Latency in the optimistic case is just the model’s normal inference time plus the wait for the challenge window. And unlike a TEE, you are not asked to trust Intel or NVIDIA’s silicon; you only need one honest party willing to watch and challenge. That one-honest-actor property is genuinely powerful, the same property that makes optimistic rollups work, and it let Ora ship a live on-chain AI oracle across more than a dozen chains while the zkML crowd was still publishing benchmarks. The logic was sound for 2024. The problem is that 2024’s constraints were not permanent.

The left flank: zkML cracked the LLM wall

The first pillar of opML’s case was that zkML simply could not handle real models. In 2026 that pillar cracked.

The clearest break came from Lagrange, whose DeepProve system was open-sourced in June 2026. According to the release (news.eandtnews.com), DeepProve has generated more than 12 million proofs and verified over 3 million inferences end-to-end, running proof generation up to 60 times faster and verification up to 671 times faster than the prior state of the art. Crucially, it proved full inference for OpenAI’s GPT-2 and Google’s Gemma-3, not toy classifiers, with Llama-class models described as in active development. Lagrange chief executive Ismael Hishon-Rezaizadeh framed the milestone as a category shift: “AI needs a verification layer, and that layer should be open,” he said, adding, “Today, DeepProve becomes a primitive. The black box is open.”

The deeper change was architectural. For years, the thing that made zkML slow was representing a neural network’s non-linear operations (the ReLUs, the softmaxes) as arithmetic circuits. The 2026 breakthrough was to stop doing that and use lookup arguments instead. Jolt Atlas, an extension of a16z’s Jolt zero-knowledge virtual machine to tensor operations built by ICME Labs and NovaNet, swaps circuit representation for lookups and sumcheck. Wyatt Benno of ICME Labs put it plainly (blog.icme.io): “Lookups eliminate the need for circuit representation entirely,” with benchmarks showing a zkML inference proven in roughly 0.7 seconds versus 4 to 5 seconds for the older EZKL pipeline.

None of this means zkML is free; proving a frontier LLM end-to-end is still out of reach, and the honest framing is a wall that is cracking, not one that has fallen. But opML’s entire cost advantage rested on the assumption that the cryptographic alternative was impossibly expensive. Every order of magnitude zkML claws back is an order of magnitude of that advantage gone.

The right flank: TEEs hit production volume

If zkML squeezed opML on cost, trusted execution environments squeezed it on speed and, increasingly, on sheer adoption.

A TEE is a hardware enclave (Intel TDX, AMD SEV, or a confidential-computing GPU) that runs a computation in an encrypted, tamper-resistant environment and produces a signed attestation that a specific program ran on genuine hardware. NVIDIA’s confidential-computing mode for its H100 GPUs reports inference overhead under 7 percent (nvidia.com), close enough to native speed that, for many workloads, the trust trade-off is worth it. There is no dispute window, no bond, no challenger required; the answer comes back attested in real time.

And the volume is real. Phala Network, which runs confidential AI across a fleet of TEE workers, reported 23.1 billion model tokens processed in a single day as of mid-August 2026, drawn from its OpenRouter provider data (phala.com). That is not a testnet metric or a projection; it is production inference at a scale opML’s flagship has never approached. The catch, and it is a real one, is the trust model: a TEE asks you to trust the chip manufacturer and to assume the enclave has no exploitable side channel, an assumption that a steady drip of hardware vulnerabilities keeps testing. Trustlessness maximalists hate it. But for an AI agent that needs an answer now, trust-the-silicon beats wait-for-the-window.

That is the crux of the squeeze. For latency-sensitive work, TEEs already win, and they win at billions of tokens a day. For trust-maximal work, zkML is closing the cost gap. opML’s home was the space in between, and that space is being colonized from both directions.

The middle stayed thin: the problems that did not get fixed

opML could have defended the middle by fixing its own weaknesses. It mostly did not, and the sharpest articulation of why came from a rival design.

A December 2025 paper, Optimistic TEE-Rollups: A Hybrid Architecture for Scalable and Verifiable Generative AI Inference on Blockchain (arXiv 2512.20176), frames the field around a “verifiability trilemma”: no system can simultaneously deliver high computational integrity, low latency, and low cost. On opML specifically, the authors are unsparing: “optimistic approaches (opML) impose prohibitive dispute windows, preventing real-time interactivity.” That is the latency problem in one sentence. Because a result is only final after the challenge window closes, any application that needs to act on an inference immediately (an agent making a trade, a game resolving a move, an oracle feeding a live price) either waits or accepts an unfinalized answer that could still be reverted.

The second unfixed problem is the verifier’s dilemma. opML’s security depends on at least one honest, competent watcher re-running computations and challenging bad ones. But watching costs money (you have to run the model yourself), and in the normal case, when everyone is honest, the watcher earns nothing. Truebit anticipated exactly this and answered it with forced errors and a jackpot that pays challengers to stay alert; opML has no equivalent mechanism deployed. So the economic reality is a courtroom that is always open and almost never visited. The same incentive puzzle HOGE Wire has documented in validator economics, where security that leans on under-rewarded vigilance tends to decay toward nobody-is-watching, shows up here in sharper form.

Add the composability break (a downstream contract cannot safely consume an inference until it is final), and the requirement that the full execution trace be public (so opML offers no privacy on its own), and you have a design whose theoretical elegance keeps running into practical friction. None of these are fatal in isolation. Together, in a year when the alternatives improved, they are why the middle got thin.

Four ways to verify AI, side by side

It helps to see the four approaches together. Each answers the question did the AI really run this model? with a different guarantee, and each pays a different price for it.

The pattern the table makes obvious is that opML is nobody’s best option on any single axis in 2026. It is not the most trustless (zkML is), not the fastest (TEEs are), and not the one shipping the most inference (TEEs again). Its historical edge, good-enough trust at the lowest cost, is exactly the property the other two are eroding. That does not make it useless, but it does explain the token chart.

The tell: opML is borrowing from its rivals

The clearest sign that opML cannot hold the middle alone is that its own proponents keep bolting on the very techniques it was meant to replace.

Ora’s own research points this way. In opp/ai (arXiv 2402.15006), a team including Ora’s Cathie So, KD Conway, and founder Kartin Wong proposes a framework that, in their words, “integrates Zero-Knowledge Machine Learning (zkML) for privacy with Optimistic Machine Learning (opML) for efficiency.” That is opML admitting it needs zkML to cover its privacy hole. Separately, academic work on zk-OPML, published in a 2026 issue of the Journal of King Saud University, uses zero-knowledge techniques specifically to shrink opML’s dispute window, attacking the latency problem with the very cryptography opML was designed to avoid.

The Optimistic TEE-Rollups paper goes further. It proposes to keep opML’s fraud-proof safety net but run the fast path inside an H100 TEE for sub-second finality, then sprinkle in stochastic zero-knowledge spot checks. In other words, the 2026 frontier design uses all three of the other approaches, TEE for speed, ZK for spot-verification, and optimistic fraud proofs as a backstop, with pure opML relegated to the backstop role. When your best future is as one ingredient in someone else’s recipe, you have lost the argument that you are a standalone category. It is the same convergence HOGE Wire traced in cross-chain security, where optimistic and cryptographic verification stopped being rival camps and became layers in the same stack.

opAgent and the agent-economy bet

If opML has a lane where the squeeze could still be survived, it is the one Ora has bet its product roadmap on: autonomous on-chain agents.

The thesis is clean. As AI agents move from chatbots to economic actors that hold assets and sign transactions, someone has to answer a hard question: when an agent moves money, how do you know the model actually produced the decision the agent claims? A private key can sign anything; it cannot prove the reasoning behind it. Ora’s answer is opAgent, an Onchain Perpetual Agent framework that went live on 31 January 2025 (Decrypt). opAgents are described as “truly autonomous digital entities that exist indefinitely through the blockchain,” controlled by smart-contract logic rather than a private key, with opML acting as the bridge that makes each action (trading on a DEX, transferring tokens, deploying a contract) verifiable rather than merely asserted.

It is a genuinely good fit for opML’s strengths. Agent decisions are often asynchronous, tolerating a delay before finality better than a live trading oracle would, and the one-honest-watcher model maps neatly onto a world where an agent’s counterparties have every incentive to challenge a fabricated action. This is the demand story the whole verifiable-AI sector keeps pointing at, and it connects opML to the broader question HOGE Wire raised in zkML’s demand problem: the proofs work, but who actually pays for them?

So far, the market’s answer is: almost nobody, yet. opAgent has been live for more than a year and a half, and the token that meters its economy still trades at a sub-$1-million valuation on negligible volume. The right market for opML may well be autonomous agents. The evidence that agents are paying for opML at any scale is not yet on the chain.

Follow the money: the token tells the story

Tokens are a noisy signal, but across the verifiable-compute sector the noise rhymes, and opML’s token is the loudest note of distress. Every project in the category is deep underwater from its all-time high, a reminder that shipping credible infrastructure and capturing token value are very different achievements. What sets ORA apart is the degree. The figures below are from CoinGecko as of 18 August 2026: ORA, LA, PROVE, PHA, and EIGEN.

Read down the market-cap column and the hierarchy is clear. The crypto-economic approach, riding restaking’s larger capital base, is an order of magnitude bigger than the rest. The zkML and TEE projects cluster in the tens of millions. opML’s flagship is alone below a million, ranked outside the top 2,600, turning over a few hundred dollars a day. You can argue that token price is a bad proxy for technology, and you would be right: PHA’s drop of more than 98 percent from its peak coexists with 23 billion tokens of daily throughput. But that is exactly the point. Phala has the drawdown and the usage. ORA has the drawdown and, publicly at least, not much usage to show for it. When a token is this thin, it also becomes hard to list, hard to use as a real fee unit, and easy for exchanges to ignore, a dynamic HOGE Wire mapped in its work on what it actually costs to get a token listed.

The closed-model ceiling

There is a structural limit on opML’s addressable market that rarely gets stated plainly: opML can only really secure models a challenger can independently re-run.

The fraud proof is the whole security model, and a fraud proof requires that a watcher be able to reproduce the disputed computation step for step. That works for open-weight models, Llama, Mistral, Gemma, Stable Diffusion, where anyone can download the weights and replay the inference inside the FPVM. It does not work for the closed, hosted models most users actually want. Ora’s Onchain AI Oracle supports frontier open models across sixteen-plus chains, but it also relays proprietary APIs like GPT-4o, and those, per Ora’s own OAO documentation and repository, are fundamentally different: with no public weights, there is nothing for a challenger to re-execute, so a closed-model result is attested or relayed rather than opML-verified in the same sense. The optimistic guarantee quietly evaporates for exactly the models with the most demand.

This is not a bug Ora can patch; it is intrinsic to fraud proofs. And it caps the market. The workloads where opML’s guarantee is strongest, open-weight, reproducible, non-real-time, are a real slice of the AI world, but they are a slice, not the whole. zkML shares part of this constraint, but a TEE does not care whether the model is open or closed; it attests whatever ran inside the enclave. In the fight for the enterprise workloads that involve proprietary models, opML starts several steps back.

Where opML still wins

It would be a mistake to write opML off. The squeeze is real, but so is the residual niche, and the honest version of this analysis has to state the steelman.

Start with cost. In a world where zkML is closing the gap but has not closed it, opML remains the cheapest way to attach a credible correctness guarantee to a large, open-weight model, because in the happy path there is no proof to generate at all. For high-volume, latency-tolerant, cost-sensitive inference (batch scoring, periodic model-driven oracle updates, asynchronous agent decisions) that economics still matters. Second, the trust model is genuinely attractive to purists: one honest watcher, no trusted hardware vendor, no cryptographic trusted setup. Third, opML is easy to reason about because it reuses the optimistic-rollup mental model the entire industry already understands, which lowers the integration cost for developers who have shipped optimistic systems before.

And the hybrids cut both ways. Yes, opp/ai and Optimistic TEE-Rollups show opML leaning on its neighbors, but they also show that fraud proofs remain a useful backstop even in designs led by TEEs or ZK. Optimistic verification did not disappear when rollups matured; it became a layer. opML’s most realistic future is probably the same: not the headline act, but the safety net underneath faster primitives, the thing that catches a compromised enclave or a buggy prover. That is a smaller role than being the way we verify AI, but it is a durable one. Categories that become plumbing do not trend on token charts; they just quietly keep working.

The regulator’s shadow and the demand question

Hanging over all of this is a question the market has not answered: does verifiable AI need a token at all, and will regulation create the demand that adoption has not?

On the regulatory side, US agencies spent 2026 drawing clearer lines. In a joint interpretation issued on 17 March 2026 (SEC Release 33-11412), the SEC and the CFTC named 16 crypto assets, including Bitcoin, Ether, Solana, and Chainlink’s LINK, as digital commodities rather than securities. What the guidance did not do was say anything about AI-oracle or verifiable-compute infrastructure tokens. ORA and its peers sit in the gap, still subject to a case-by-case Howey analysis that turns on whether buyers are relying on a managerial team’s efforts, which for a sub-$1-million token with a live-but-quiet network is an uncomfortable place to be. Clearer rules helped the majors; they did nothing to resolve the status of the long tail, a theme that runs through HOGE Wire’s coverage of how compliance is getting written into the code itself.

The more interesting possibility is that regulation becomes the demand. If enterprises and regulators start requiring proof that an AI system produced a given output, the nice-to-have of verifiable inference becomes a must-have, and the cheapest credible option for open-weight models has a real business again. Lagrange has leaned hard on this framing, citing McKinsey research that a large majority of enterprise executives say they will not scale AI systems without proof of correctness. That is the bull case for the entire sector, opML included. But it is a bet on future rules, not current usage, and betting on future rules is a familiar way for crypto infrastructure to stay perpetually early.

What would un-squeeze opML

The way out is not mysterious; it is just hard. Three things would widen the middle again. First, latency: an opML variant that shrinks or hides the dispute window, whether through the zk-OPML route or by pairing a TEE fast path with an optimistic backstop, would let opML back into the real-time applications the OTR authors say it currently locks itself out of. Second, incentives: a deployed challenger-reward mechanism, Truebit’s jackpot or some staking-funded equivalent, would turn the always-open, never-visited courtroom into one that is actually staffed, which is the difference between security on paper and security in practice.

Third, and most important, a use case that clears at scale. opAgent is the right bet on paper; it needs to become a real one on-chain, with agents paying opML fees in volumes that show up in the data rather than in the pitch deck. Until then, opML’s honest position is a clever, cheap, structurally limited approach that had its clearest window in 2024 and 2025, before its neighbors caught up. It is not dead. It is boxed in, and the walls are its own unresolved trade-offs as much as anyone else’s progress. The optimistic bet still assumes someone will show up to check the work. In 2026, the harder question is whether anyone will show up to use it.

Frequently Asked Questions

What is opML (optimistic machine learning)?

opML is a way to verify AI computation on a blockchain without expensive cryptography. A node runs a model off-chain, posts the result on-chain, and anyone can challenge it during a dispute window; if a challenge succeeds, a fraud proof replays the disputed step to settle it. If no one challenges before the window closes, the result is accepted as final. The design borrows directly from optimistic rollups like Arbitrum and Optimism.

How is opML different from zkML?

zkML produces a cryptographic proof that the model ran correctly, so a verifier trusts math and re-runs nothing, but proving is slow and expensive. opML produces no proof up front; it assumes the result is correct unless someone disputes it, which is far cheaper but adds a challenge-window delay and relies on at least one honest watcher. In short, zkML pays cost to buy instant trust, while opML pays latency to buy low cost.

Is opML still used in 2026?

Yes, but narrowly. Ora Protocol runs a live opML-secured Onchain AI Oracle across more than a dozen chains and has built an autonomous-agent framework, opAgent, on top of it. Measurable usage and token value remain very low, and much of the 2026 frontier research folds opML into hybrid designs alongside TEEs and zero-knowledge proofs rather than using it alone.

Why is the ORA token worth so little if the technology works?

ORA traded near $0.0058 with a market cap under $1 million in August 2026, more than 99 percent below its 2024 peak, per CoinGecko. Working technology and token value are different things: the entire verifiable-compute sector is deep underwater, and opML in particular faces competition from faster TEEs and cheaper-than-before zkML, plus a structural limit on which models it can secure. Thin liquidity then tends to feed on itself.

Can opML verify closed models like GPT-4o?

Not in the full sense. opML’s security comes from letting a challenger re-run the disputed computation, which requires open, downloadable model weights. Open-weight models like Llama or Stable Diffusion can be verified this way; closed, hosted APIs such as GPT-4o have no public weights to reproduce, so they are relayed or attested rather than truly opML-verified. That caps opML’s reach over the proprietary models many users want most.

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