zkML’s Demand Problem: The Proofs Work, Now Who Pays?

Something strange happened to zero-knowledge machine learning in 2026. For three years the field carried one nagging question: could a cryptographic proof ever wrap a real neural network, or was zkML condemned to prove toy models forever while the actual AI ran on a server you simply had to trust? That question is mostly settled now. In June, Lagrange Labs open-sourced its DeepProve system after it had generated more than 12 million cryptographic proofs and verified over 3 million AI inferences end-to-end, including full transformer inference for models such as GPT-2 and Gemma-3. The wall that defined the sector cracked. And yet LA, the token that carries Lagrange’s name, trades near $0.05 for a market capitalization under $10 million, roughly 97% below its 2025 peak, according to CoinGecko. That gap, between technology that works and a market that prices it close to zero, is the real zkML story this year. The engineering problem is retreating. The demand problem has arrived, and it is the harder of the two.

What zkML Promises, and Why 2026 Is Different

zkML, short for zero-knowledge machine learning, produces a cryptographic receipt for an AI computation. The proof shows that a specific model ran on a specific input and produced a specific output, and anyone can check that claim in milliseconds without rerunning the model and without seeing the model weights or the private input. Swap the mental model from “trust the server that says it used the right model” to “verify the math that says a particular model produced this answer.” For a technology that wants to sit underneath high-stakes automated decisions, that is a powerful promise, and it is fundamentally different from merely renting decentralized GPUs, which proves nothing about honesty.

The reason 2026 matters is that the promise used to collapse the instant you pointed it at anything real. The non-linear operations that make a neural network more than a spreadsheet, the ReLUs and softmaxes and attention layers, are brutally expensive to express in the arithmetic circuits that zero-knowledge proofs require. Early benchmarks were sobering. Work from Daniel Kang and collaborators showed that proving a single inference of a social-media recommendation model with the EZKL toolkit could take many hours per example. Modulus Labs framed the whole issue in its 2023 “Cost of Intelligence” report: verifiability was achievable, but the overhead ran to many orders of magnitude. For most of its life, zkML in practice meant proving handwritten-digit classifiers and gesturing at everything larger. The change in 2026 is that the cost curve finally bent, hard.

The Supply Side Is (Almost) Solved

Two developments did most of the work. The first is DeepProve, which Lagrange open-sourced on 3 June 2026 after roughly a year of production use. By the project’s own accounting, DeepProve had generated more than 12 million proofs and verified over 3 million inferences end-to-end, producing large-language-model proofs up to 60 times faster than the prior state of the art and verifying them up to 671 times faster, all while preserving model accuracy rather than degrading it. It proved full transformer inference for GPT-2 and Gemma-3 from token embeddings through to the final output, a first for the field, with Llama-class models described as still in active development. That framing matters: this is a milestone, not a finish line, and the largest frontier models remain out of reach. Lagrange chief executive Ismael Hishon-Rezaizadeh put the open-sourcing in deliberately load-bearing terms: “Today, DeepProve becomes a primitive. The black box is open.” The DeepProve code and the underlying full-LLM proof milestone are both public.

The second development is Jolt Atlas, described in a February 2026 paper. It extends the Jolt proving system, which was born at a16z crypto and later adapted by ICME Labs and NovaNet, from general-purpose computation to machine learning by working directly on ONNX tensor operations. Its central trick is to lean on lookup arguments and the sumcheck protocol, which handle exactly the non-linear functions that used to explode circuit size. An optimization the authors call neural teleportation shrinks the lookup tables while preserving accuracy, and the approach can prove on ordinary devices without specialized hardware. The through-line across both projects is the same: replace expensive circuits with cheap lookups for the operations that hurt most, and the cost of proving a model falls by one or two orders of magnitude. Put DeepProve and Jolt Atlas together and the supply side of zkML looks healthier than at any point in its history. Which makes the token charts all the more jarring.

The Demand Side Tells a Different Story

If verifiable AI were about to eat the world, you would not know it from the price of the assets meant to represent it. Every token tied to zero-knowledge proving or verifiable compute is deep underwater, most of them 90% or more below their all-time highs, even as the underlying software ships real proofs daily. The pattern is not a single project stumbling; it is the entire category. Lagrange’s LA, Succinct’s PROVE, Boundless’s ZKC and Giza’s GIZA all trade as though the market has concluded that proving AI is a solved research problem with no paying customer attached. It is the same punishment that has fallen on the wider decentralized-AI trade, where even Gensyn’s $AI slid to new lows despite live infrastructure. Framework quality and token performance have decoupled almost completely.

The most brutal line is Giza. GIZA changes hands near $0.00088 for a market cap under $300,000, some 99.8% below its high, according to CoinGecko. Succinct’s PROVE, the healthiest of the group at roughly $0.15 and a $29 million cap per CoinGecko, is still down around 91%. When a sector’s best-performing asset has lost nine tenths of its value while its technology hits its stated milestones, the problem is not the technology.

Why Most Applications Still Do Not Verify

To understand the demand gap, start with the arithmetic every potential buyer runs. Vitalik Buterin laid it out in his 2024 essay on the promise and challenges of crypto plus AI. Verifiable computation is far more expensive than the ordinary kind. An Ethereum block, he noted, takes a client a few hundred milliseconds to verify directly, but generating a zero-knowledge proof of that same block can take hours; for the non-linear layers of a neural network, the best implementations in practice still show overhead around 200 times. That is the tax. A buyer only pays it when the value of a trustworthy answer clearly exceeds the cost of proving it. For a customer-service chatbot reply, it never does. For a decision that moves a large sum of money or triggers a legal obligation, it might.

There is a second cost hiding inside the first: accuracy. Elena Burger of a16z crypto has pointed out that zero-knowledge proofs today cannot represent 32-bit floating-point operations in the arithmetic-circuit format they need without massive overhead, which pushes teams toward 8-bit quantized models that are, in her words, crude approximations of the original. So the buyer is often asked to pay a large premium to prove a slightly worse model. Newer lookup-based systems soften this, but they do not erase it. This is why zkML demand is narrow rather than broad: the honest market for it is the set of computations where a wrong or unverifiable answer is genuinely catastrophic, not the vast ocean of everyday inference where cheap and unchecked is good enough.

Giza’s Pivot and the Hard-Business Lesson

No single story captures the demand problem better than Giza. The team arrived at the 2023 Ethereum Community Conference in Paris positioned squarely as a zkML project, building tooling to prove machine-learning models with zero-knowledge cryptography. Over the following two years it quietly became something else. Giza’s flagship product today is ARMA, an autonomous stablecoin-yield agent that went live on Base and moves user funds across lending venues such as Aave, Morpho, Compound and Moonwell. The proving story faded; the agent story took over. A team with real cryptographic talent and venture backing looked at the market and concluded that the thing people would actually pay for was an agent that earns yield, not a proof that an inference was correct.

That is the uncomfortable lesson sitting under the whole sector. Pure zkML, sold as verification-for-its-own-sake, has struggled to find a customer who will write a recurring check. The buyers who show up want an outcome, and verifiability is at best a feature of that outcome, not the product itself. Giza’s token, now worth a rounding error, is what happens when a project prices in a zkML future and then has to live in a zkML present. The good news for the field is that Giza did not die; it adapted. The sobering news is what it had to adapt into. The question every remaining proving project now faces is whether verification can be the product, or whether it is destined to be an ingredient someone else sells.

Three Places Real Demand Is Emerging

The picture is not all gloom. Demand for verifiable AI is real, but it is concentrated in three specific places where the value of a proof plausibly clears its cost: where a regulator wants an audit trail, where an autonomous agent has to be trusted with money, and where a decision inside a financial protocol moves funds on the strength of a model’s output. None of these is the mass consumer-inference market that early token prices assumed. Each is a genuine, defensible niche. Taken together they suggest zkML’s route to a business is narrow and vertical, not wide and horizontal, and the projects that survive will be the ones that stop selling proofs in the abstract and start selling into these lanes.

Demand One: Regulation Wants a Receipt

The strongest tailwind for verifiable AI is not crypto at all; it is compliance. As of 2 August 2026, the European Commission, acting through its AI Office, can formally enforce the obligations the EU AI Act places on general-purpose AI, alongside the rules on prohibited practices. From that date the Commission can demand documentation, run technical evaluations of a model, order compliance measures, restrict or withdraw a model from the EU market, and levy fines up to 15 million euros or 3% of global turnover for general-purpose-model breaches, with broader violations reaching 35 million euros or 7%. High-risk systems must keep logs sufficient for post-hoc audit. The official framework is explicit that auditability is now a legal requirement, not a nice-to-have.

This is exactly the shape of problem a zero-knowledge proof fits. A ZK audit trail lets the commitments be public and immutable while the sensitive content stays private, revealed only when a proof is generated for an auditor’s specific query, which satisfies data-protection law and the regulator at once. Lagrange framed its DeepProve launch around a related datapoint, citing McKinsey research that 71% of enterprise executives say they will not scale AI systems without proof of correctness. Whether that figure holds or not, the direction is clear: buyers who face audits have a reason to pay for proofs. In the United States the picture is murkier for the tokens. The SEC and CFTC issued a joint interpretation in March 2026 that classified sixteen assets as digital commodities, including Chainlink’s LINK, but said nothing about AI inference or verifiable compute, leaving infrastructure tokens such as LA, ZKC and PROVE in a case-by-case gray zone even as their use case gains a regulatory tailwind. The demand pull and the token treatment are pointing in different directions, which is its own kind of trap; getting the rules to reward the right behavior is a theme HOGE Wire has explored in how compliance rules get written into the code.

Demand Two: Agents That Must Prove Themselves

The second lane is autonomous agents. An on-chain agent that holds its own wallet and signs its own transactions is only as trustworthy as the model deciding what it does, and 2026 supplied a vivid reminder of how fragile that is. In the Grok and Bankr episode in May, an attacker combined a gifted membership NFT that escalated wallet permissions with an instruction hidden in a Morse-code social-media reply, and drained roughly $150,000 to $200,000 of tokens on Base before anyone noticed, an incident logged in the OECD AI incident monitor. When an agent can be talked into signing, “trust the model” is not a security posture.

zkML offers a partial answer: an agent can attach a proof that its output was produced by a specific, approved model running on a specific input, so a counterparty or a smart contract can gate the transaction on that proof rather than on faith. It narrows the trust surface, and it pairs naturally with the account-level guardrails emerging in smart-account wallets, where spending limits and session keys can be made conditional on a valid inference proof. The honest caveat is that zkML proves the computation, not the intent behind the prompt. It confirms the right model ran; it cannot confirm the input was not a malicious injection. Verifiable inference would not have stopped the Morse-code attack on its own. It is a real piece of the agent-security puzzle, and a real source of demand, but it is one layer among several, not a cure.

Demand Three: DeFi, Oracles, and Money on the Line

The third lane is the one where zkML has already shipped something live. OpenGradient, working with the EZKL toolkit, built a verifiable dynamic-fee model for Uniswap V3, letting an automated market maker adjust pool fees to market conditions with a proof that the fee came from the agreed model rather than someone’s discretion. The integration uses EZKL, an open-source compiler that turns an ONNX model into a Halo2 circuit and is available on GitHub; several verifiable-inference networks, including OpenGradient, Ritual and others, build on it. The logic generalizes cleanly. When a model’s output determines who gets paid, both sides want a receipt.

The candidate use cases follow the money. A lending protocol can run a credit-risk model on a borrower’s private financial history and let the borrower prove they qualify without exposing balances or income. An oracle can publish a data feed with a proof of how it was computed instead of asking users to trust the operator. A prediction market can resolve a subjective question with an AI judge whose verdict is provable rather than assertable. In every case the proof is worth paying for precisely because a wrong or manipulated answer would be expensive, and because the parties do not trust each other. That is the wedge for zkML: not cheap inference at scale, but costly inference where the stakes justify the receipt, the same economic logic that makes verifiability matter in adjacent fields like validator economics.

zkML Versus the Cheaper Rivals

Even inside its narrow niches, zkML has to win a bake-off. A buyer who wants verifiable AI has at least four ways to get it, and zkML is the most trustless but also the slowest and most expensive. Optimistic machine learning, or opML, publishes a result and a commitment and treats it as final unless someone disputes it within a challenge window, which is cheap but adds latency and assumes at least one honest watcher. A trusted execution environment leans on hardware attestation from a chip vendor, which is fast, with overheads that can sit under 10%, but replaces cryptographic trust with trust in the silicon and its side-channel resistance. Crypto-economic approaches stake capital and slash it for misbehavior, which scales well but offers an economic rather than a mathematical guarantee.

The practical result is that for most workloads a buyer picks the cheaper option. If sub-second speed matters, a TEE wins; if cost matters more than instant finality, opML or a crypto-economic network wins. zkML earns its premium only when the buyer needs the strongest possible guarantee and cannot accept trusting a chip vendor or a set of stakers, or when privacy of the input or the model is non-negotiable. That is a defensible market. It is not a large one, and the 2026 trend is toward hybrids that use a fast TEE for finality, fraud proofs for economy and occasional zero-knowledge spot-checks for assurance, which means zkML increasingly competes as a component inside a stack rather than as the whole answer.

The Proof-Marketplace Bet

One structural answer to the demand problem is to make proving a commodity market where scattered demand can aggregate and competition drives cost down. Boundless, which launched its mainnet on Base in September 2025 backed by the Ethereum Foundation, Base and Wormhole, runs a protocol-agnostic proving marketplace with a Proof of Verifiable Work mechanism that rewards provers for the quantity and complexity of proofs they produce. Its chief executive Shiv Shankar framed the pitch to CoinDesk: “For the first time, developers on any chain can access abundant zero-knowledge compute to build complex applications that scale across ecosystems without sacrificing decentralization.” Succinct, with its SP1 zkVM and PROVE token, plays a similar role and has secured billions of dollars of value across dozens of protocols.

The marketplace thesis is sound in theory and unproven in the market. A commodity exchange only works when buyers show up, and the token prices suggest they have not shown up in the volume the model needs. It is a chicken-and-egg bind: marketplaces need paying demand to justify their networks, demand needs cheap proofs to be viable, and proofs get cheap mainly through the scale that only real demand provides. DeepProve and Jolt Atlas attack the cost side of that loop, which is genuine progress. But cheaper proofs do not create buyers by themselves; they only lower the bar a buyer has to clear. Until a use case with real, recurring volume lands, the proof marketplaces are building highways ahead of the traffic.

Does Verifiable Compute Even Need a Token?

There is a deeper question the drawdowns are asking, and it is awkward for the whole sector. DeepProve is open-source. Its own chief executive called it a primitive, something anyone can pick up and run. EZKL is a public library. If proving is a piece of software you download and execute, where exactly does token value accrue? A zero-knowledge proof is trustless no matter who generated it; a proof produced by a single company, or by an ordinary cloud service, verifies just as cleanly as one produced by a decentralized network. That property is the technology’s greatest strength and, for tokenholders, its most inconvenient feature.

The counterargument is that coordinating a decentralized network of provers, with staking, slashing and payment for work, needs a native asset to function, and that a permissionless marketplace resists capture in a way a single vendor cannot. That may be true. But it is a claim about market structure, not about the proof itself, and the market is clearly skeptical that the structure requires the token. This is the tension behind valuations 91% to 99.8% below their highs: the technology can succeed on a large scale while the token attached to it captures very little of that success. Anyone weighing one of these assets should separate the engineering from the tokenomics, because the history of tokenized infrastructure is full of protocols that paid dearly to list a token the product never actually needed.

What Has to Be True for zkML to Find Its Buyers

Three things would turn zkML’s narrow niches into a durable business. First, the cost curve has to keep bending. The lookup-centric designs behind Jolt Atlas and the order-of-magnitude gains in DeepProve are the right trajectory, and on-device proving without specialized hardware would widen the buyer pool considerably. Second, a proof has to become load-bearing in some legal or financial sense: an EU AI Act auditor accepting a zero-knowledge attestation as evidence, an insurer requiring one before underwriting an AI system, or a court treating it as proof of what a model did. The moment a proof is the difference between a fine and a clean audit, the 200-times overhead stops looking expensive.

Third, one anchor application with real, recurring volume has to emerge, most likely somewhere money already moves on a model’s output: DeFi risk engines, verifiable oracles, prediction-market resolution, or agent-to-agent commerce where machines pay machines and neither side will extend trust. Any one of these could become the flywheel that the proof marketplaces are waiting for. What is unlikely, and what the early token valuations wrongly assumed, is that zkML becomes the default wrapper around ordinary inference. Cheap and unverified will remain good enough for the vast majority of AI calls for years. zkML wins by going deep in the places where being wrong is unacceptable, not by going wide.

The Bottom Line for 2026

The defining fact about zkML in 2026 is a role reversal. For years the sector’s bottleneck was the mathematics; the honest answer to “can you prove a real model” was “not yet.” That answer is now “mostly, and getting cheaper.” The bottleneck has moved downstream, from the proving system to the profit-and-loss statement. The question is no longer whether a proof is possible but whether anyone will pay enough for it, often enough, to sustain a network and a token. Regulation, agent security and money-on-the-line DeFi are the three places where the answer is plausibly yes. Everywhere else it is still no, and the market is pricing accordingly. 2026 is the year zkML stopped being a research problem and became a go-to-market problem, and go-to-market problems are the kind that separate the projects that endure from the tokens that do not.

Frequently Asked Questions

What is zkML in simple terms?

zkML, or zero-knowledge machine learning, is a way to attach a cryptographic proof to an AI computation. The proof shows that a particular model ran on a particular input and produced a particular output, and anyone can check it in milliseconds without rerunning the model and without seeing the private input or the model weights. It turns “trust that we used the right model” into “verify that we did.”

Can zero-knowledge proofs verify a large language model in 2026?

Partly, and for the first time. In June 2026 Lagrange Labs open-sourced DeepProve, which proved full transformer inference for GPT-2 and Gemma-3 end-to-end and generated large-language-model proofs up to 60 times faster than the prior state of the art. Llama-class models were still in active development at that point, and the largest frontier models remain out of reach, so zkML can prove real language models now but not yet the biggest ones.

Why are zkML and verifiable-compute tokens down so much?

Because the technology is shipping while paying demand has not arrived at scale. As of 17 August 2026 the leading proving tokens sit 91% to 99.8% below their highs even though their software works. Verifiable computation is far more expensive than ordinary computation, so most applications skip it, and the proofs themselves are trustless regardless of who generates them, which makes it hard for a token to capture the value the technology creates.

Is zkML better than a TEE or optimistic ML?

It depends on what you are optimizing for. zkML gives the strongest, most trustless guarantee and preserves privacy, but it is the slowest and most expensive option. A trusted execution environment is much faster but asks you to trust a chip vendor, and optimistic ML is cheaper but adds a dispute-window delay. Many 2026 systems combine them, using fast hardware for finality and zero-knowledge proofs for occasional high-assurance checks.

Will regulators require verifiable AI?

Not by name, but the pressure is building. Since 2 August 2026 the EU AI Act’s obligations on general-purpose and high-risk AI are enforceable, including audit-logging requirements and fines up to 3% of global turnover for some breaches. Zero-knowledge proofs are a natural fit for keeping audit trails that are both verifiable and private. In the United States, the SEC and CFTC have clarified how they treat crypto assets but have said nothing specific about verifiable compute, so the regulatory tailwind for the use case is stronger than the clarity around the tokens.

Marcus Okafor is a senior editor at HOGE Wire covering the crossover of artificial intelligence and crypto infrastructure.