Who Mines the Proofs? Inside zk-ML’s Proving Markets in 2026

A day ago, this site ran a piece on how zero-knowledge machine learning can hand you a flawless cryptographic receipt for a model that is only an approximation of the one you meant to run. That story was about what a proof says. This one is about who makes it.

The quiet fact of zk-ML is that the proof is the expensive part. Running an AI model is cheap and fast. Wrapping that run in a zero-knowledge proof, a cryptographic object small enough to verify on-chain in milliseconds yet heavy enough to compute that it can take a rack of hardware minutes or hours to produce, is the bottleneck that kept the whole field academic for years. In 2026 the proofs finally got fast enough to matter. The open question is no longer whether an AI inference can be proven. It is whether anyone can generate the millions of proofs the vision requires at a price the market will pay.

The answer taking shape this year is the most crypto answer imaginable: turn proving into a market. Instead of every application buying its own racks of provers, a handful of networks now sell proof generation the way miners sell hashes. Some call it a proof marketplace. Some call it a prover network. One runs it as a subnet on Bittensor and literally calls the reward Proof of Inference. They are all betting that the supply side of verifiable AI, not the demand side, is the business worth owning. Here is how that bet works, who is placing it, and why their tokens still say the market is not convinced.

The half of zk-ML nobody puts on the slide

Every zk-ML explainer draws the same diagram: a model runs, a proof pops out, a smart contract checks it, everyone trusts the result without rerunning the work. The diagram is honest about the verifier side. Checking a proof really is cheap, a small and roughly fixed on-chain cost regardless of how large the model was. The diagram lies by omission about the other side. Generating the proof is where the cost lives, and that cost is enormous.

Vitalik Buterin put rough numbers on it in his essay on crypto and AI: the matrix multiplications at the heart of a neural network can be proven with under four times overhead, but the non-linear operations (the activations, the softmax, the attention) blow up to roughly two hundred times, and a full zero-knowledge proof of something as ordinary as an Ethereum block, he noted, can take hours. Translate that to a language model and the arithmetic becomes brutal. A survey of verifiable machine learning frameworks describes the field as caught in a three-way tension between model accuracy, prover time, and the soundness of the numeric model; you can usually have two.

That is why proving is not something you do on your laptop between API calls. It wants specialized software, a lot of memory, and increasingly GPUs or custom silicon dedicated to nothing but churning out proofs. For a single application, standing up that infrastructure to serve bursty, unpredictable demand is wasteful: you pay for peak capacity and idle most of the time. The economics beg for someone to pool the hardware and sell proving as a service. That someone, in 2026, is a market.

Why proving turned into a market

Pull apart the verifiable-AI stack and you find three jobs that used to be bundled together. There is the model, which does the thinking. There is the application, which decides what to prove and consumes the result. And there is the prover, which does the expensive cryptographic grunt work of turning a specific execution into a succinct proof. For most of zk-ML’s short history, whoever built the application also ran the prover, because there was no one to outsource it to. That is the bundle 2026 is breaking.

The logic is the same logic that unbundled cloud computing from owning servers. Proving demand is spiky and specialized; proving hardware is expensive and general. Separate the two and you can run the hardware at high utilization across many customers, driving the unit cost of a proof down toward the cost of the electricity and the silicon. A dedicated prover can justify GPUs, FPGAs, eventually ASICs, and the engineering to keep them busy. An application developer who just wants a receipt that the right model ran should, in theory, be able to buy one the way they buy an S3 bucket.

Three market structures are competing to be that S3 bucket. The first is the open proof marketplace, where anyone can post a job and any prover can bid to fill it. The second is the prover network, a more curated set of operators who stake capital to guarantee they will deliver proofs correctly and on time. The third is the DePIN subnet, which borrows Bittensor’s incentive machinery to pay a swarm of competing miners for verified inference. They overlap at the edges, but their answers to one question (how do you make a stranger prove honestly and cheaply?) are genuinely different.

Proof marketplaces: Boundless and the auction for compute

The purest version of the market idea is Boundless, the proving marketplace incubated by RISC Zero that launched its mainnet on Base in September 2025. The pitch is that any chain or application can post a proving job to a shared order book, and a competitive field of independent provers races to fulfill it. Boundless is deliberately protocol-agnostic: it does not care whether your proof comes from one zkVM or another, only that it verifies. Its backers include the Ethereum Foundation, and dozens of protocols have integrated against it rather than build proving in-house.

The mechanism Boundless uses to keep provers honest and hungry is called Proof of Verifiable Work. Rather than reward raw computation the way Bitcoin rewards hashing, it rewards the delivery of valid, useful proofs, scaling payouts to the quantity, speed, and complexity of what a prover produces. A prover that finishes more hard jobs faster earns more. It is an auction and a leaderboard at once, and the token (ZKC) is the unit both of payment and of the stake a prover puts at risk.

Shiv Shankar, Boundless’s chief executive and a former chief information officer at RISC Zero, framed the ambition in almost utility-grid terms when the network went live. For the first time, he told CoinDesk, developers on any chain can access what he called abundant zero-knowledge compute, enough to build applications that scale across ecosystems without surrendering decentralization. Abundant is the operative word. The whole premise of a marketplace is that competition turns a scarce, artisanal good into a commodity you can buy by the unit. Whether proofs are abundant yet is a question the demand section will return to.

Prover networks: Succinct, Lagrange, and staking the work

The second structure keeps the market but tightens the membership. A prover network is a coordinated set of operators who post a bond before they are allowed to fill proof requests, so that failing to deliver, or delivering garbage, costs them money. Succinct is the archetype. Its SP1 zkVM lets developers prove ordinary programs written for a RISC-V processor, which means machine-learning code can be proven without hand-building a bespoke circuit for every model. Around SP1 sits a prover network whose participants stake the PROVE token to take on jobs; the stake is the honesty mechanism and the liveness guarantee in one.

Lagrange is the network that bet most specifically on AI. Its DeepProve library is purpose-built for zk-ML, and the company open-sourced it in June 2026 after generating more than 12 million cryptographic proofs and verifying over three million AI inferences, claiming proof generation up to sixty times faster than the prior state of the art and verification hundreds of times faster. DeepProve-1 was the first system to prove a full language-model inference end to end, albeit a small one (GPT-2 class), with larger open models named as the next target. Behind it runs the Lagrange Prover Network, again staked, again slashing operators who misbehave.

Ismael Hishon-Rezaizadeh, Lagrange’s founder and chief executive, cast the open-sourcing as a deliberate surrender of control. His company did not build DeepProve to own verifiable AI, he wrote; it built it so nobody had to, because the verification layer should be a public good. “Today, DeepProve becomes a primitive,” he said on the day of the release. “The black box is open.” It is a generous sentiment with a hard commercial edge: if the library is free, the money is in running the network that generates the proofs. The code is the loss leader; the proving is the business.

The DePIN bet: who mines proofs on Bittensor

The third structure is the one that makes the market metaphor literal, and it is the least covered. Inference Labs runs Subnet 2 on Bittensor, a network it describes, without much modesty, as the largest decentralized zk-ML proving cluster in the world. The subnet launched in 2024 and had generated more than 160 million zero-knowledge proofs by mid-2025, a figure that has only climbed since. The project began life branded Omron and has since been reworked under the name DSperse, but the machinery is the same: a crowd of independent operators competing to prove that AI models ran correctly, paid in the subnet’s share of Bittensor emissions.

Bittensor’s design maps neatly onto proving. In a subnet, miners do the work and validators score it; the validators’ rankings decide how the block reward is split, so a miner that produces faster or better proofs earns a larger slice. Inference Labs calls its protocol Proof of Inference, and the incentive is spread across the whole pipeline: operators are rewarded not just for generating proofs but for distilling models into provable form, designing better circuits, and optimizing the prover itself. It is, in effect, a permanent open competition to drive the cost of a proof down, with a token pot for whoever does it best.

Inference Labs raised 6.3 million dollars in June 2025 from investors including DACM, Delphi Ventures, Arche Capital, and Lvna Capital, with a community round led by Native Capital. Co-founder Ron Chan framed the stakes in the language of inevitability. “AI agents will soon outnumber human users online, and AI is going to be part of every decision in our lives,” he said; “cryptographic trust in AI outputs is non-negotiable.” As he put it to The Block, Proof of Inference is the answer: privacy-preserving, provable inference at scale. The academic literature is moving the same way, treating a distributed network of provers, not a single trusted server, as the default shape of the problem.

Proof of Verifiable Work: crypto rediscovers mining

If the names keep echoing Bitcoin (Proof of Verifiable Work, Proof of Inference, miners, difficulty, rewards) that is not an accident. The proving market is mining with the waste engineered out. Bitcoin pays its network to compute hashes that are useless except as a lottery ticket. A proving network pays its operators to compute proofs that are intrinsically valuable, because someone downstream actually needs them. The security model rhymes (do measurable work, get paid in a token), but the work is a product rather than a sacrifice.

That shift changes the incentives in a way readers of this site will recognize from the hashrate beat. Bitcoin miners compete on energy price and chip efficiency because the output is fungible. Proof miners compete on something richer: the quality of their circuits, the speed of their GPUs, the cleverness of their batching and aggregation. Two provers given the same job can produce the same proof at wildly different costs, and the gap is engineering, not just electricity. That is good news for a market that wants prices to fall, and it is the reason every one of these networks pours its token incentives into prover optimization as much as into raw capacity.

It also imports mining’s oldest problem. The moment proving rewards are large enough, the operators with the cheapest hardware and the best engineers win, and winners in a work market tend to consolidate. The same dynamic that turned Bitcoin mining into a handful of industrial pools is latent in proving, and it sits uncomfortably beside the word decentralized that every one of these projects uses to sell itself.

A map of the proving market in 2026

The structures shade into one another, and most projects hedge across more than one. But the organizing question (how does a stranger get paid to prove, and what keeps them honest?) sorts them cleanly enough to put on a page. The table below maps the main venues, the token that fuels each, and what they actually prove.

The bottom two rows are a reminder that not every market for verifiable AI sells cryptographic proofs at all, a point the piece returns to below. The top four are the proving economy proper, and they are converging on the same supply-side playbook from different directions.

What actually makes a proof cheap

A market only helps if there is real headroom to cut costs, and in 2026 there is. The single biggest lever is the one the previous generation of zk-ML lacked: lookup arguments and sumcheck-based proof systems that retired the brute-force circuits used to approximate a model’s non-linear steps. That is the engineering that let DeepProve claim its order-of-magnitude speedups and let newer toolkits prove real network architectures rather than toy ones, and a wave of 2026 research is racing to turn those primitives into production-minded verifiable-inference pipelines.

Hardware is the second lever. Proving is embarrassingly parallel, so GPUs already dominate, and the roadmap runs through FPGAs toward dedicated proving ASICs, the same arc Bitcoin mining walked a decade ago. The third lever is aggregation: recursive proofs that fold many small proofs into one, so a network can batch a flood of individual inferences and amortize the on-chain verification cost across all of them. For a consumer application generating thousands of proofs an hour, aggregation is the difference between viable and absurd.

Put the three together, competition plus specialized hardware plus aggregation, and you get the deflation the marketplaces are selling. The honest caveat is that the starting point is high. Even after 2026’s gains, proving a frontier-scale language model end to end remains economically out of reach; the systems that ship prove small models, or a single important step, or a heavily compressed version of a bigger one. The market can make a proof cheaper. It cannot repeal the arithmetic that makes a proof of a hundred-billion-parameter model cost more than the answer is worth.

The recentralization trap

Here is where the mining analogy turns into a warning. A proving market rewards whoever makes proofs cheapest, and cheapest means scale: the biggest GPU fleets, the best-tuned kernels, the capital to buy next year’s hardware. In a determinism-sensitive setting the pressure is even sharper, because bit-for-bit reproducible proving tends to want identical hardware across operators, which quietly nudges a supposedly open network toward a single canonical chip. The endpoint of an efficiency race is concentration, and concentration is exactly what the decentralization pitch was supposed to avoid.

The pull is not just economic; it is also where the paying customers are. Lagrange spent late 2025 and 2026 chasing not crypto protocols but governments and defense primes. In November 2025 it joined the Oracle Partner Network to deploy DeepProve inside Oracle’s sovereign and mission-cloud environments for defense, aerospace, and government, and it has positioned DeepProve as an approved component for Lockheed Martin programs and integrated it into Anduril’s Lattice software. Its named ecosystem partners run to IBM, NVIDIA, Qualcomm, Intel, and AWS. That is a very different customer than a Bittensor subnet miner, and it tugs the proving supply chain toward permissioned, high-assurance clouds.

So the plausible 2026 outcome is not one proving market but two. A permissionless crypto economy (Boundless, Succinct, the DSperse subnet) sells abundant, censorship-resistant proofs to on-chain applications and agents. A sovereign-enterprise lane sells high-assurance proofs to institutions that want mathematical certainty and a vendor they can sign a contract with. Both are the supply side of verifiable AI. Only one of them needs a token, which is a thought to keep in mind when the tokens come up.

When the market skips the math

Not every market for verifiable AI sells a zero-knowledge proof. The two bottom rows of the table are the honest competition, and they win on the dimension zk-ML is worst at: speed and scale. The crypto-economic approach, pioneered by EigenCloud, does not prove an inference at all. It bonds operators with restaked capital, has them re-execute the work, and slashes anyone caught lying. Trust comes from money at risk rather than from mathematics, which is why it inherits the same questions this site has raised about who actually stands behind restaked guarantees; the guarantee is only as good as the cost of corruption.

The hardware approach, led by Phala, leans on trusted execution environments: the model runs unmodified inside a sealed enclave on an Intel or NVIDIA chip, which signs an attestation that the right code ran on the right inputs. It is fast enough to serve tens of billions of tokens a day, orders of magnitude more throughput than any ZK prover can touch, and it needs no circuit and no quantization. The catch is that you are trusting the chipmaker and the enclave, and 2025’s hardware attacks on confidential computing showed those seals can be broken with physical access. A TEE market sells convenience and performance; it does not sell the trust-no-one guarantee that is zk-ML’s entire reason to exist.

That is the real competitive landscape for the proving economy. The ZK marketplaces are not only racing each other to the cheapest proof; they are racing two rival trust models that are already cheaper and faster, and that are good enough for a lot of buyers who do not actually need cryptographic certainty. The proving market has to find the customers who do.

The tokens tell a different story

If the technology is shipping and the market structures are live, you would expect the assets that power them to be worth something. They are not, or at least the market does not think so. Every token tied to the proving economy trades a long way below where it started, and the ones closest to the pure zk-ML thesis trade worst of all. The figures below are from CoinGecko on 19 October 2026.

ZKC is the sharpest line. Boundless shipped a working marketplace on Base, courts real protocols, and its token sits near its all-time low, down about 98 percent from its peak with a market cap under ten million dollars. Lagrange, with the most credible AI-proving product and a defense-contractor client list, is in the same ditch. Even against a broader market that has spent 2026 stuck in a holding pattern, these drawdowns are extreme, and the volumes are thin enough that a token like ZKC lives one quiet decision away from the kind of unexplained delisting that has quietly thinned the long tail this year.

The gap between working technology and worthless token is the sector’s defining feature, and it has a simple cause. A proving network’s token is paid to the supply side, to the provers, mostly out of emissions, before there is enough paying demand to absorb it. Until buyers show up with real money for real proofs, the token is a subsidy to miners, and the market prices it as one. The supply side has been built. The demand side is the bet that has not yet paid.

The demand a cheap supply cannot summon

A commodity market needs buyers, and the honest read is that proof demand remains mostly prospective. The bullish case rests on a handful of lanes. The first is compliance. Lagrange, citing McKinsey research, likes to note that seventy-one percent of enterprise executives say they will not scale AI systems without proof of correctness, and the risk backdrop is real: Stanford’s AI Index counted a sharp rise in AI incidents alongside adoption near ninety percent of organizations. If even a fraction of that caution converts into a willingness to pay for cryptographic receipts, the proving market has its customers.

The second lane is autonomous agents. As software agents start to transact, trade, and settle on their own, the party on the other side has every reason to demand proof that the agent ran the model it claims; AI-settled information and prediction markets, where bots increasingly set the odds, are a natural early buyer. The third is DeFi, where any model that moves money (a dynamic fee, a risk score, a liquidation trigger) is worth proving because the cost of a tampered model is measured in drained pools. These are plausible, growing, and still small. Against them sits the caveat the companion piece laid out: buyers also want to know the proof covers the real model, not a lossy approximation, and that fidelity question has not gone away. Abundant supply does not create demand; it only waits for it.

Washington is quiet, Brussels is loud

Regulation is the wildcard that could turn prospective demand into mandatory demand, and the two biggest jurisdictions are pulling in opposite directions. In the United States, the SEC and CFTC issued a joint interpretation in March 2026 that sorts digital assets into five buckets (digital commodities, collectibles, tools, payment stablecoins, and securities) and leans on the familiar Howey question of whether buyers expect profit from an issuer’s managerial efforts. It says nothing at all about AI, verifiable compute, or infrastructure tokens.

That silence leaves the proving tokens in a gray zone with an interesting wrinkle. A token earned by doing measurable proving work, the Proof-of-Verifiable-Work model, looks less like a security than a token sold on the promise of a team’s future efforts; you are being paid for output, not betting on a roadmap. But the same token sold to speculators with staking yields and emission schedules can look very much like the thing the SEC is describing. How these assets are marketed may matter more to their legal fate than what they technically do.

Europe is the mirror image. The EU AI Act’s general-purpose obligations became enforceable on 2 August 2026, carrying the power to fine providers up to fifteen million euros or three percent of global turnover and a requirement that model providers publish a sufficiently detailed summary of training content. The law mandates that you declare how a model was built; it does not yet force you to prove any of it is true. That gap, declared versus proven, is precisely the hole a proving market wants to fill, which is why the clearest demand signal for verifiable AI is coming from Brussels, not Washington.

What it means for players and builders

For a gaming audience, the proving market is the unglamorous infrastructure that provably-fair AI will eventually run on. A game that uses a model to set loot odds, match players, or drive a non-player character could in principle prove it did not cheat, but only if generating that proof is cheap enough to do at the pace a game demands. That is a throughput problem a marketplace is built to solve, and public demos that verified millions of live inferences already hint at the bursty, high-volume demand a proving network would have to absorb. Cheap, abundant proofs are the precondition for provable fairness being a feature rather than a research paper.

For builders, the practical shift is that you no longer have to become a cryptographer to ship verifiable AI. You post a job to Boundless, stake into Succinct, call DeepProve, or route through a Bittensor subnet, the same way you once stopped running your own servers and started renting them. Expect the market to split: a permissionless lane for on-chain applications and agents, and a sovereign-enterprise lane for institutions that want a contract and a compliance story. Pick the lane that matches your trust model and your buyer.

The verdict on the proving economy is that it is real infrastructure built ahead of its market. The hard engineering problem, generating an AI proof fast enough to sell, has largely been solved, and three different market structures are competing to sell it. What has not arrived is the wave of paying demand the tokens were priced for, which is why the plumbing looks impressive and the assets look broken. The proofs are getting made. The open question, as it has been all year, is who will pay to read them.

Frequently Asked Questions

What is a zk-ML proving market?

A zk-ML proving market is a network that sells the expensive part of zero-knowledge machine learning, generating the proof, as a service. Instead of each application running its own proving hardware, specialized operators compete to produce proofs that an AI model ran correctly, and are paid in the network’s token. Boundless, Succinct, Lagrange and the Inference Labs subnet on Bittensor are the leading examples in 2026.

Which networks generate zk-ML proofs in 2026?

The main venues are Boundless, an open proof marketplace on Base; Succinct, a staked prover network built on its SP1 zkVM; Lagrange, whose DeepProve library targets AI inference specifically; and Inference Labs, which runs the DSperse proving cluster on Bittensor Subnet 2. EigenCloud and Phala sell related guarantees using restaking and trusted hardware rather than zero-knowledge proofs.

What is Proof of Verifiable Work?

Proof of Verifiable Work is the reward mechanism Boundless uses to pay provers for delivering valid proofs, scaling payouts to the quantity, speed and complexity of the work rather than to raw hashing. It is a close cousin of Inference Labs’ Proof of Inference on Bittensor. Both resemble Bitcoin mining, except the work produces something useful, a proof someone needs, instead of a throwaway lottery ticket.

Why are zk-ML proving tokens down so much if the technology works?

Because the token is paid to the supply side, the provers, mostly out of emissions, before there is enough paying demand to absorb it. The proving networks have shipped working technology, but on-chain demand for proofs is still small, so the market prices the token as a subsidy to miners. In October 2026, tokens like ZKC, LA and PROVE traded roughly 89 to 98 percent below their all-time highs.

Is a zk-ML proof the same as running AI in a TEE or using restaking?

No. A zero-knowledge proof gives a trust-no-one mathematical guarantee that a computation was done correctly. A trusted execution environment, as used by Phala, asks you to trust the chipmaker’s sealed enclave instead, and restaking networks like EigenCloud rely on money at stake that can be slashed. The hardware and crypto-economic approaches are faster and scale to larger models, but they do not offer the same cryptographic certainty.

Marcus Okafor is a staff writer at HOGE Wire covering the intersection of crypto and artificial intelligence.