Gensyn Explained: Verifiable AI Compute, Delphi and $AI

Training a large language model still comes down to one scarce resource: enough GPUs, wired together in the right place, for long enough. Nvidia’s newest data center chips sell out years ahead of shipping, hyperscalers lock up capacity in multibillion dollar blocks, and everyone else, from university labs to solo researchers, competes for what is left. Gensyn is a London-founded protocol that has spent five years arguing this is a solvable problem rather than a permanent one. If a network can prove that a stranger’s laptop, gaming PC or spare data center rack actually did the machine learning work it claims to have done, that idle hardware becomes usable compute, with no cloud contract or trusted intermediary required.

Three months after Gensyn’s mainnet and $AI token launch, that argument has moved from research paper to a live, still small, network with a real fee-paying application. This piece covers how Gensyn’s verification stack actually works, what its Delphi and Judge systems do with it, where the $AI token stands after a rocky first quarter, how the project compares with rivals like Bittensor and Akash Network, and the technical and regulatory questions still open.

• Mainnet live since April 22, 2026, an OP Stack layer-2 anchored to Ethereum
• $AI token generation event on April 29, 2026, with near-simultaneous listings on Binance, Coinbase, OKX, Kraken, KuCoin and Gate.io
• Delphi, Gensyn’s AI-settled information market, is the network’s first fee-generating application
• $AI supply is fixed at 10 billion tokens, with roughly 13% circulating as of July 2026
• Judge, a verifiable AI evaluation system built on Gensyn’s Verde research, has been live since August 27, 2025

What Problem Is Gensyn Trying to Solve

Machine learning research is bottlenecked less by ideas than by access to hardware, a frustration Gensyn’s founders experienced firsthand as researchers who had working algorithms but no GPU time to run them.

At the same time, enormous amounts of computing hardware sit idle: gaming PCs between sessions, small data centers with spare capacity, personal workstations doing nothing overnight. The issue is not that this hardware could not technically run machine learning workloads; it is that nobody renting it out has a reliable way to trust the result. A dishonest or careless compute provider could return fabricated output and still collect payment. Gensyn’s bet, laid out across a series of research papers and now a live network, is that this is a cryptography and incentive-design problem rather than an unsolvable one.

a16z crypto, which led Gensyn’s Series A, framed the upside in similarly structural terms. General partners Ali Yahya and Guy Wuollet wrote that by tapping “the long tail of idle, machine-learning-capable compute around the world,” from smaller data centers to personal computers, “Gensyn can potentially 10-100x the available compute power for machine learning”, in a16z crypto’s investment thesis. That estimate has not been independently tested at network scale, but it is the thesis the entire protocol is built around, and it places Gensyn in a wider 2026 category of crypto-incentivized compute networks, alongside names like Akash Network and Bittensor, all trying to route idle or underused hardware toward AI workloads through different trust mechanisms.

From Entrepreneur First to a $43 Million Series A

Gensyn was founded in 2020 in London by Ben Fielding and Harry Grieve, who met through the Entrepreneur First accelerator program. Both came from machine learning backgrounds rather than crypto ones. Grieve later described the company’s origin as a technical problem looking for a solution, not a token looking for a use case: “We have a very acute machine learning problem that needed a decentralized trust layer,” he told Decrypt, contrasting Gensyn’s approach with projects that start from crypto and search for relevance afterward.

The company raised a $43 million Series A in June 2023, led by a16z crypto with participation from CoinFund, Canonical Crypto, Protocol Labs and Eden Block, bringing total funding to more than $50 million. Fielding described verification, rather than raw distributed computing, as the project’s real breakthrough: “That’s the big secret sauce behind Gensyn, we’ve solved that problem for machine learning training specifically,” he told Decrypt at the time of the raise.

It took nearly three more years of research and testnet work before any of that verification technology reached a live mainnet with a trading token attached, a notably slower path than many crypto projects that raised money in the same 2023 cycle.

Inside the Stack: Four Layers That Make Up Gensyn

Gensyn’s documentation breaks the protocol into four layers, each solving a distinct piece of the problem of trusting compute you did not personally run.

The first two layers, reproducible execution and verification, are where most of Gensyn’s original research effort has gone. AXL and on-chain coordination are newer additions that turn the verification primitive into something applications can actually use, culminating so far in Delphi, covered later in this piece.

RepOps and Verde: Verifying Machine Learning Without Rerunning It

The technical obstacle Gensyn had to clear is subtle. Standard floating-point math does not reliably produce identical results across different hardware, different parallelization strategies, or even different orderings of the same operations; a sum computed in a different sequence, or a matrix multiplication tiled differently across GPU cores, can legitimately land on a slightly different final number. That means two honest compute providers running the exact same model on the exact same input can return different results without either one lying, which makes it hard to tell honest variance apart from a provider cutting corners.

Gensyn’s answer is RepOps, a library of reproducible operators engineered to enforce a fixed reduction order and correctly rounded math so identical calculations return bitwise-identical outputs regardless of the chip, driver or parallelism setup underneath. RepOps is the practical foundation for Verde, described in an academic paper co-authored by Fielding and seven other researchers, including NYU cryptographer Joseph Bonneau, titled “Verde: Verification via Refereed Delegation for Machine Learning Programs.” Verde lets a computationally limited client hand a job to several untrusted compute providers at once, with a guarantee of getting the correct result as long as at least one of them is honest. When providers disagree, Verde does not need to rerun the entire job to work out who is right; a dispute protocol pinpoints the exact step in the computational graph where results first diverged, then resolves just that step.

That check-only-the-disputed-step property is what makes the system economically viable. Rerunning an entire large training job to catch one dishonest provider would erase the cost savings of decentralized compute in the first place; refereed delegation keeps verification cheap even when the underlying job is not.

RL Swarm: Crowdsourced Reinforcement Learning

Gensyn’s most visible pre-mainnet product was RL Swarm, an open source framework, published on GitHub, for training models collaboratively across a peer-to-peer network of volunteer machines. Anyone could clone the repository and run a node, from a home GPU to a laptop, and connect it to Gensyn’s testnet to build an on-chain identity tracking their contributions over time.

RL Swarm’s active training task has changed as Gensyn iterated. An earlier reasoning-focused environment was replaced on November 12, 2025 by CodeZero, which extends RL Swarm from math and reasoning into collaborative coding tasks. In CodeZero, models take on distinct roles, Solvers, Proposers and Evaluators, and are scored through a rule-based reward function guided by a local evaluator model rather than direct code execution.

RL Swarm itself has been paused at times as Gensyn’s engineering focus shifted toward shipping mainnet and Delphi, and there is no permanently running swarm today; participation happens in scheduled or community-organized runs. That is a reminder that RL Swarm currently functions more as a research and onboarding testbed than as a dependable production training pipeline.

Judge: Teaching the Network to Grade Its Own Work

Launched on August 27, 2025, Judge is Gensyn’s attempt to solve a related but distinct problem: how do you score an AI model’s output when the obvious shortcut, using a more powerful model like GPT-5 as a judge, relies on a closed, opaque API that can change without notice and cannot be independently reproduced. Gensyn’s blog post introducing Judge argues that in high-stakes contexts, an unreproducible judgment is not a trustworthy one, however capable the underlying model.

Judge is built directly on Verde, using deterministic CUDA kernels and a custom compiler that lowers computational graphs from the ONNX format while tracking the provenance of every operation back to its source. That lets any evaluation Judge produces be independently checked rather than taken on faith. Its first real workload folded Judge into RL Swarm itself: a reasoning task structured as a prediction market, where RL Swarm models place bets on the correct answer as evidence is revealed progressively, rewarding both accuracy and speed of reasoning.

Mainnet Day: An OP Stack Rollup Anchored to Ethereum

Gensyn’s mainnet went live on April 22, 2026, as an EVM-compatible layer-2 network built on the OP Stack and anchored to Ethereum, following roughly a year of testnet operation that began in April 2025. Gensyn reported hashrate-equivalent capacity on day one above 5,000 Nvidia H100 GPUs, a figure the company itself published rather than an independently audited benchmark, worth treating as directional rather than precise.

Building on the OP Stack was a pragmatic choice: it let Gensyn reuse a well-tested rollup framework and inherit Ethereum’s base-layer security rather than bootstrap consensus from scratch. It also means Gensyn inherits the general risk profile of any Ethereum layer-2, including reliance on a bridge contract to move assets between layers, the same category of infrastructure behind some of crypto’s largest thefts, as detailed in HOGE Wire’s breakdown of how bridge hacks actually happen. Nothing about Gensyn’s own bridge has been reported as compromised, but the architecture carries the category’s inherent exposure regardless of the network’s own code quality.

Delphi: Markets Settled by AI, Not Oracles

Delphi is Gensyn’s first live application built on top of its verification stack, and the thing that turned $AI from a governance-adjacent token into one with actual fee flow. It launched on testnet in December 2025, initially focused narrowly on AI model performance benchmarks, then expanded to let anyone create a market on virtually any resolvable question, before moving to mainnet on April 22, 2026.

Delphi deliberately avoids the term prediction market. In an interview about the product, Fielding drew a sharp distinction: “Prediction markets in our view are unidirectional. All you can do is sell information. Information markets to us are very specifically bidirectional, requiring you to be able to sell information,” he said, describing a model where creators earn revenue both from framing a popular question and from the trading volume it generates, becoming what he called “shelling points” for a topic. On settlement, Fielding was direct about what the AI judges are actually doing: “The models that are settling the markets very explicitly are not predicted. All they need to do is reason about what has happened,” evaluating evidence of events that already occurred rather than forecasting the future.

Mechanically, Delphi runs on a symmetrical LMSR, or logarithmic market scoring rule, automated market maker, which continuously reprices outcomes as capital flows in and out without needing an order book or a counterparty on the other side of every trade. Optional REE-based cryptographic receipts let anyone independently verify how an AI model actually settled a given market, tying the application layer back to Gensyn’s core verification research.

On testnet, engagement was substantial for a niche product: one sports-focused market alone drew more than 87,000 traders and $4.88 million in volume, while an Oscars-style market pulled in more than 45,000 traders.

The 0.5% routed to the AI BuyBack Vault is then split further: 70% of the vault is permanently burned, reducing $AI’s circulating supply, 29% funds a Community Treasury, and the remaining 1% pays executor rewards for the accounts that trigger the buyback-and-burn transactions.

The $AI Token: Supply, Burn Mechanics and a Rough First Quarter

$AI launched on April 29, 2026, listing within days on Binance, Coinbase, OKX, Kraken, KuCoin and Gate.io, with Upbit adding KRW, BTC and USDT trading pairs on June 30, 2026. The token has a fixed maximum and total supply of 10 billion, of which roughly 1.305 billion, about 13%, was in circulation as of July 21, 2026, according to CoinGecko.

Unlike Bitcoin’s fixed, disinflationary issuance schedule, where new supply halves on a fixed clock, a pattern HOGE Wire has tested for statistical validity beyond Bitcoin itself, $AI’s scarcity mechanism runs in the opposite direction: nearly all of its supply already exists on paper, and the network is instead trying to shrink circulating supply gradually through Delphi’s burn mechanism, while roughly 87% of tokens remain to unlock over time. That is a meaningfully different bet: it depends on burn volume from actual product usage outpacing scheduled unlocks, rather than on a hard-coded issuance curve doing the work automatically.

Price action since launch has been volatile even by 2026 token-generation-event standards. $AI spiked to an all-time high of $0.1073 on its April 29 debut, before falling sharply; by June 25 it had bottomed at an all-time low of $0.02035. As of July 21, 2026, $AI traded around $0.02516, up about 24% from that low but still down more than 76% from its opening-day peak, with a market capitalization near $32.8 million against a fully diluted valuation near $251 million. That roughly eightfold gap between market cap and fully diluted valuation is the low-float, high-FDV pattern that has defined much of 2026’s token launches, and it means today’s price says relatively little about where $AI would trade once the bulk of supply actually unlocks.

Bringing Liquidity On-Chain: The Uniswap V3 Proposal

In March 2026, GFX Labs submitted a formal request for comment to Uniswap’s governance forum proposing a canonical deployment of Uniswap V3 on Gensyn’s layer-2, with Oku, GFX Labs’ existing Uniswap V3 trading interface, serving as the front end. The stated goal is to build deep on-chain liquidity for $AI alongside USDC and ETH, and to give Gensyn’s buyback mechanism a transparent venue to execute through rather than relying solely on centralized exchange order books.

As of this writing, the proposal remains in community review on Uniswap’s governance forum rather than live. If it goes ahead, it would give $AI a genuine on-chain trading venue for the first time, something most of its volume currently lacks, since trading has stayed concentrated on the centralized exchanges that listed it around the April token generation event.

AXL and the Case for AI Agents That Pay Each Other

Of Gensyn’s four architectural layers, AXL is the one aimed least at today’s usage and most at where the team expects demand to come from next. Described in Gensyn’s own documentation as encrypted, application-agnostic peer-to-peer communication with no central server, AXL explicitly supports both MCP, or Model Context Protocol, and A2A, or Agent-to-Agent, communication standards, the same protocols increasingly used to let autonomous AI agents discover tools and transact with one another.

That positioning connects Gensyn to a broader shift already underway elsewhere in crypto, where AI agents are being given their own wallets and the ability to sign transactions autonomously, a trend HOGE Wire has covered in depth in its look at crypto’s autonomous agent economy. If that trend continues, the pitch for AXL is straightforward: an agent that needs verified compute, whether to run inference, fine-tune a model, or settle a Delphi-style information market, could theoretically discover and pay a Gensyn provider directly, with Verde providing the trust layer that lets one autonomous piece of software rely on another’s output without a human checking the work. None of that runs at meaningful scale yet; it is closer to a roadmap thesis than a shipped product.

Gensyn vs Bittensor vs Akash: Three Different Bets on Decentralized AI

Gensyn is often mentioned alongside Bittensor and Akash Network as decentralized AI infrastructure, but the three protocols are solving different layers of the same broad problem, and comparing their token metrics without that context is misleading.

Bittensor, covered in detail in HOGE Wire’s explainer on TAO, its subnets and 2026’s governance disputes, runs dozens of specialized subnets, each an incentive market where miners produce AI outputs, from text generation to price prediction, and validators rank the results, with TAO emissions flowing to whoever scores best. It is fundamentally an output-quality marketplace. Akash Network takes a different approach again, operating a general-purpose reverse-auction marketplace for cloud compute and GPUs with no built-in mechanism for verifying that a given AI computation was correct, only that capacity was rented. Gensyn sits a level lower than either: it is trying to verify raw compute itself, the actual correctness of a training or inference job, rather than ranking output quality or simply matching supply with demand.

The market cap gap is stark and worth sitting with. Bittensor’s roughly $1.9 billion valuation is nearly sixty times Gensyn’s, reflecting both TAO’s multi-year head start since its 2021 launch and a subnet ecosystem with far more visible day-to-day usage. Akash, smaller than Bittensor but still about five times Gensyn’s size, benefits from being immediately useful to any developer who just wants cheaper GPU rental, no verification story required. Gensyn’s pitch is the most technically ambitious of the three and, so far, the least proven in terms of market capitalization and trading liquidity.

Where $AI Sits Under US Securities and Commodities Law

Gensyn itself is a UK-registered company, which puts it outside the SEC’s direct jurisdiction, but $AI trades on exchanges serving US customers and is held by US-based traders, so American securities law still shapes how the token is marketed and listed. Under the SEC’s March 2026 interpretive release on crypto asset categories, which sorts tokens into five buckets, digital commodities, digital collectibles, digital tools, payment stablecoins and digital securities, a network fee token like $AI, whose value comes from programmatic buyback-and-burn activity tied to actual product usage rather than a promise of managerial profit-sharing, would most plausibly fall into the digital commodity or digital tool categories rather than the securities bucket. That is an analytical read of how the framework would likely apply, not a determination the SEC has actually made about $AI specifically, and the guidance itself is non-binding on courts. HOGE Wire’s broader mapping of how US crypto enforcement actually works in 2026 covers how that framework has played out case by case.

Delphi raises a narrower but sharper question. Prediction and information markets sit inside an active regulatory fight in the United States: the CFTC published a notice of proposed rulemaking on June 10, 2026 covering how event-contract markets get reviewed and what counts as a prohibited gaming contract, while Kalshi, the best-known US event-contract exchange, has spent 2026 fighting more than a dozen lawsuits from state gaming regulators alongside its own offensive suits, plus the CFTC’s first-ever insider trading case tied to event contracts. Gensyn’s decision to market Delphi as an information market rather than a prediction market, built on Fielding’s bidirectional-versus-unidirectional distinction, reads as a deliberate attempt to sit outside that fight rather than inside it. Whether US regulators ultimately agree with that framing is untested, and nothing about Delphi has drawn regulatory action so far.

What Could Go Wrong

Gensyn’s risks split roughly into token-structure risk, execution risk and category risk. On tokens: with only about 13% of $AI’s 10 billion supply circulating, the bulk of unlocks are still ahead, a dynamic that has pressured price in most comparable 2025 and 2026 token launches regardless of underlying product progress.

On execution: RL Swarm’s history of pauses shows that Gensyn’s crowdsourced training layer is not yet a dependable, always-on pipeline, and Delphi, while genuinely live with real fees, is still a single application carrying the weight of the network’s entire revenue and burn story. A second and third application would meaningfully de-risk that concentration.

On category risk: Gensyn is not the only project chasing verifiable AI compute, and rivals using zero-knowledge proofs, trusted hardware enclaves or optimistic fraud-proof designs are pursuing the same underlying problem through different cryptographic tradeoffs; refereed delegation is one credible approach among several, not an obviously dominant one yet. The network also inherits ordinary layer-2 and crypto-wallet risks: anyone connecting a wallet to run an RL Swarm node, claim rewards, or trade on Delphi is a target for the same credential-phishing and malicious-approval scams that affect wallet holders across the wider industry, a risk that has nothing to do with Gensyn’s own code quality and everything to do with the ecosystem it operates in.

Frequently Asked Questions

What is Gensyn?

Gensyn is a decentralized protocol built as an OP Stack layer-2 network anchored to Ethereum that lets anyone contribute computing hardware to machine learning workloads and get verifiably paid for it. Its core innovation, a verification system called Verde, checks that computation was carried out correctly without re-running the whole job. Gensyn’s mainnet went live on April 22, 2026, alongside the $AI token, and its first live application is an AI-settled information market called Delphi.

What is the $AI token used for?

$AI is Gensyn’s native token, used to pay network and trading fees, including fees on Delphi’s information markets. A portion of Delphi’s trading fees permanently burns $AI through Gensyn’s AI BuyBack Vault, while another portion funds a Community Treasury. The token launched on April 29, 2026 with a fixed maximum supply of 10 billion, of which roughly 13% was circulating as of July 2026.

Is Gensyn the same as Bittensor?

No. Both are decentralized AI networks, but they solve different problems. Bittensor rewards miners for producing outputs that validators rank highest across specialized subnets, essentially a marketplace for AI output quality. Gensyn focuses one level lower, on verifying that raw machine learning computation, like training or evaluating a model, was actually executed correctly on someone else’s hardware. Bittensor’s TAO token also has a far larger market capitalization than Gensyn’s $AI.

How does Gensyn verify that AI computations were done correctly?

Gensyn uses a system called Verde, described in an academic paper co-authored by Gensyn cofounder Ben Fielding, based on refereed delegation. The same task is sent to multiple independent, untrusted compute providers, and if their results disagree, Verde pinpoints the exact step in the computation where results first diverged rather than re-executing the entire job. This relies on RepOps, a library of reproducible operators engineered to return identical results regardless of the underlying hardware.

Is the Gensyn $AI token a security?

Gensyn has not been charged or sued by the SEC or any other regulator over $AI’s status, and the token functions as a network fee and buyback-and-burn asset rather than a claim on company profits. Under the SEC’s 2026 interpretive guidance on crypto asset categories, tokens whose value comes mainly from a network’s programmatic operation generally fall outside the definition of a security, though that guidance is not binding on courts and Gensyn itself is a UK-registered company outside direct SEC jurisdiction.

Daniel Reyes is a contributing writer for HOGE Wire, covering AI-crypto infrastructure and decentralized compute.