Gensyn’s First 100 Days: Delphi Thrives, RL Swarm Goes Quiet

Gensyn’s mainnet went live on April 22, 2026, and its $AI token followed a week later on April 29. That puts the network more than one hundred days into the phase that was supposed to prove out its core idea: that machine learning training can be verified the way blockchains verify transactions, without re-running the whole computation or trusting a single operator.

The number that matters most after those hundred-plus days is not the price of $AI, though that has moved plenty. It is which product is actually generating activity. Delphi, an AI-settled information and prediction market, has become Gensyn’s flagship application by default, while RL Swarm, the crowdsourced reinforcement learning network the entire verification stack was originally built to secure, has gone quiet. This piece looks at what changed, what the numbers actually show, and what the split means for a network that raised money to fix decentralized training and ended up shipping a betting market first.

What Gensyn Actually Set Out to Build

Ben Fielding and Harry Grieve founded Gensyn in London in 2020, after meeting through the Entrepreneur First program. That timing predates the current generative AI boom by a couple of years, which is part of why Gensyn is often described as an infrastructure bet on machine learning generally rather than a reaction to the large language model wave that came later. Their starting observation was straightforward: a large share of the world’s GPU capacity sits idle at any given moment, while machine learning researchers queue for scarce, expensive cloud compute. Renting out idle hardware is the easy part. The hard part is trust: how does a network verify that a stranger’s GPU actually ran the training job it was paid for, rather than returning fabricated results, without forcing every node to redundantly re-run every computation and erase the cost advantage of decentralizing in the first place.

That verification problem attracted serious money. Gensyn raised a $43 million Series A in 2023 led by a16z crypto, part of more than $78 million raised since founding, according to The Block. a16z crypto general partners Ali Yahya and Guy Wuollet made the firm’s thesis explicit, writing that Gensyn “can potentially 10-100x the available compute power for machine learning.” Grieve has described the founding motivation in blunter terms: “We have a very acute machine learning problem that needed a decentralized trust layer,” he told Decrypt around the time of that raise.

Gensyn is not the only project chasing this problem. It sits in a small cluster of protocols trying to make AI compute checkable rather than just cheap, alongside cryptographic approaches like zero-knowledge machine learning and optimistic schemes that assume correctness unless challenged. Where those approaches lean on math or dispute windows, Gensyn’s verification stack, built around a research system called Verde, leans on game theory: pay someone to check the work, and make cheating more expensive than it is worth. Whether that bet has paid off is easier to judge now, one hundred days into mainnet, than it was at launch.

The Mainnet and Token Debut, in Numbers

Gensyn’s mainnet is an OP Stack rollup, the same rollup framework behind Optimism, anchored to Ethereum. By the end of launch day, the network reported hashrate equivalent to more than 5,000 Nvidia H100 GPUs, a fast ramp for a chain that had spent roughly a year running its OP Stack testnet before flipping the switch. The $AI token generation event followed on April 29, 2026, after an English-auction public sale that valued the network between a $1 million floor and a $1 billion fully diluted cap.

The token’s all-time high came on debut day itself: $0.1073, per CoinGecko. It did not hold. By June 25, $AI had fallen to an all-time low of $0.02035, a drop of more than 80 percent from the opening spike. As of early August, $AI trades around $0.022, giving the network a market capitalization near $29 million against a fully diluted valuation above $220 million, with roughly 13 percent of the fixed 10 billion token supply in circulation.

The table below lays out the sequence from founding to the current snapshot.

RL Swarm Goes Quiet: What Happened to the Training Thesis

RL Swarm was the product meant to prove Gensyn’s core pitch: an open source framework that let anyone with a consumer GPU join a swarm of language models cooperatively improving through reinforcement learning over the open internet. Nodes answered, critiqued, and revised each other’s outputs in structured rounds, earning points for participation and training performance rather than fabricating results, at least by design.

The swarm’s task environment evolved over its testnet life. In November 2025, Gensyn replaced the original Reasoning Gym environment with CodeZero, which reframed training around cooperative coding tasks, with models taking on Proposer, Solver, and Evaluator roles and improving through a model-based reward system rather than direct code execution. Participation numbers were never continuously published in an official dashboard, but independent trackers at various points estimated testnet participation in the thousands of nodes, running on hardware as modest as a single consumer GPU.

Then, ahead of the mainnet transition, Gensyn’s own documentation confirms that RL Swarm and all Gensyn-hosted nodes have been paused, with engineering focus consolidated on Delphi as the network’s first mainnet application. That is a striking sequencing choice for a project whose entire verification architecture, and a large share of its early credibility, was built around proving that crowdsourced training could be trusted. Grieve’s own framing of the company’s mission was about unlocking scale through decentralization, not about shipping a market app: “Scale is the only answer,” he said in an interview with Bitget News shortly before mainnet. “Decentralization is to unlock an unprecedented scale of computing and data resources.” A hundred days into mainnet, the network’s actual crowdsourced training product is not running.

Delphi Becomes the Network’s Only Real Product

Delphi is a permissionless market platform where anyone can create a market on a resolvable question (crypto price targets, sports outcomes, awards shows, geopolitical events) and have it settled by an AI model instead of a human moderator or a decentralized oracle committee. Pricing runs on a symmetrical logarithmic market scoring rule, an automated market maker design that guarantees continuous liquidity from the first trade through settlement without needing an order book or a counterparty on the other side of every trade.

The fee structure is where $AI’s buyback-and-burn mechanism actually gets fed. Total protocol fees run at 2 percent of trading volume: 1.5 percentage points go to the market’s creator, and 0.5 percentage points route into a buyback vault that purchases and burns $AI, according to reporting from Bitcoin News. That creator cut is deliberate. Fielding has described it as a genuine revenue line for people who build a following around a niche market, not just an engagement gimmick, and has been explicit that Delphi is not trying to out-compete the big regulated players on their own turf. “Delphi isn’t directly competing with Polymarket and Kalshi for the same markets,” he told The Block. “The strategy is to open up an entirely new category of niche, creator-owned markets that those platforms would never build. The long tail is the point.”

Before mainnet, Delphi’s testnet drew real numbers: a sports-outcome market pulled in more than 87,000 traders and $4.88 million in volume, and an Oscars market attracted over 45,000 traders. Those are testnet figures, not mainnet ones, and neither Gensyn nor an independent tracker had published a comparable cumulative mainnet volume figure as of this writing. That gap matters for judging how much of the testnet enthusiasm has actually carried over now that trades involve real money instead of points.

Not every Delphi market even uses Gensyn’s verification technology by default. Verifiable settlement is opt-in: a creator who selects a verifiable model gets a cryptographic receipt, built on Gensyn’s Reproducible Execution Environment, that lets anyone independently rerun the settlement and confirm the answer. Creators who skip that option get a faster, cheaper market settled the way most AI products work today: you trust the output because you trust the operator.

Verde, RepOps and Judge: A Verification Stack Waiting for a Workload

Underneath the branding, Gensyn’s verification research is genuinely inventive. The core paper, published on arXiv with cryptographer Joseph Bonneau among its co-authors, describes a refereed delegation scheme called Verde: instead of re-running an entire training job to check it, which would defeat the point of decentralizing compute, a validator only needs to find and verify the first point where two nodes’ computational graphs diverge. Pair that with RepOps, short for Reproducible Operators, a fixed reduction order and rounding scheme that forces identical calculations to produce bitwise-identical results across different GPU hardware, and the result is a dispute-resolution system that is cheap most of the time and only expensive when someone is actually caught cheating.

Judge, launched in August 2025 and built on top of Verde, extends that idea to grading rather than re-execution: a cryptographically verifiable way to score whether a model’s output was actually good, not just whether the compute was run honestly. That distinction matters because grading open-ended reasoning is a softer problem than checking arithmetic, and most decentralized AI projects still lean on an opaque, unreproducible call to a closed third-party model to do it.

In practice, the verification layer that is actually live and load-bearing today is REE, the settlement mechanism behind Delphi, and even that is opt-in per market. Verde’s dispute game and Judge’s evaluation primitive were built to secure a constant stream of distributed training jobs running through RL Swarm. With RL Swarm paused, the most rigorous parts of Gensyn’s verification stack do not currently have a comparable production workload to secure. That is not the same as saying the technology does not work; nobody has reported it failing. It is a narrower claim: a hundred days into mainnet, Gensyn’s most distinctive intellectual property is better tested in papers and testnets than in daily production use.

Security and Bridge Risk: What an OP Stack L2 Inherits

Running as an OP Stack rollup gives Gensyn a well-trodden security model, and also its liabilities. The canonical route between Ethereum and Gensyn is secured by the standard OP Stack bridge contracts, with withdrawals back to Ethereum subject to the usual seven-day challenge period, per Gensyn’s own documentation. For anyone unwilling to wait a week, Gensyn recommends third-party bridge providers such as Stargate for moving $AI and USDC in and out faster.

That third-party bridge layer is exactly where 2026’s hack wave has concentrated. As HOGE Wire has covered in detail, this summer’s bridge exploits have consistently hit faster, more convenient third-party rails rather than a chain’s slow-but-secure canonical bridge, because that is where liquidity concentrates and where the attack surface, custom message-passing logic, multisig key management, oracle assumptions, is least standardized. Most established OP Stack chains publish audit reports from firms that specialize in rollup and bridge code before or shortly after mainnet, precisely because bridge contracts are high-value, high-complexity targets. Gensyn’s public documentation covers bridge mechanics clearly but does not surface an equivalent audit report in the same place, which is a gap worth closing given how many of 2026’s largest crypto losses trace back to bridge contracts rather than to the base chains they connect.

Where the Compute Price War Leaves Gensyn

Gensyn competes for attention, if not always for the exact same customer, with a wider field of decentralized compute networks that have spent 2026 undercutting cloud hyperscalers on price. Networks like Akash, io.net, and Bittensor aggregate idle or purpose-built GPU capacity and rent it out well below on-demand hyperscaler rates, which is part of why AI developers priced out of scarce cloud GPUs have started looking at decentralized alternatives at all.

The distinction that matters is what each network is actually selling. Akash and io.net are, at their core, GPU rental marketplaces: post a job, a provider bids, compute gets delivered, and trust is enforced mostly through reputation and payment escrow rather than cryptographic or game-theoretic proof of correctness. Bittensor runs a different model again, an incentive marketplace where miners produce machine intelligence outputs and validators rank them, with value flowing to whichever subnet produces the best-ranked output. Gensyn’s pitch was never simply cheaper compute. It was verified compute, a narrower and harder problem that, per the sections above, is currently proven more in research and testnets than in a live, high-volume training market.

On pure market size, the gap between Gensyn and the largest of these networks is wide, as the table below shows using current data from CoinGecko.

Bittensor’s roughly $1.8 billion market capitalization against Gensyn’s $29 million is a reminder that the market, for now, is pricing “useful AI outputs, ranked and rewarded” well ahead of “AI training, verified end to end.” Whether that gap closes if RL Swarm restarts, or widens further if it does not, is one of the more concrete things to watch over the rest of the year.

AXL and the Bet on Machine-to-Machine Payments

One layer of Gensyn’s stack looks past both training and prediction markets toward a different customer entirely: autonomous AI agents. The Agent eXchange Layer, or AXL, is a peer-to-peer communication layer built to support the Model Context Protocol and Agent2Agent standards that have become common plumbing for letting AI agents call tools and talk to each other. The pitch is that agents running on or alongside Gensyn could discover each other, negotiate, and pay for services without a human approving every transaction.

It is the most speculative layer of the four described in Gensyn’s architecture documentation, and the one furthest from generating measurable revenue today. Machine-to-machine payments are a genuinely active area of crypto infrastructure work in 2026, with multiple competing standards trying to become the default rail for agent commerce. AXL’s success depends less on Gensyn’s own execution than on whether agent developers standardize around it instead of a rival payment rail backed by a larger ecosystem. It is a real, shipped product, not vaporware, but it is also the layer most likely to be judged in years rather than in the hundred-day window this piece otherwise uses.

Tokenomics and the Unlock Cliff Nobody’s Pricing In Yet

$AI has a fixed maximum supply of 10 billion tokens, of which roughly 13 percent was in circulation as of early August, per Tokenomist. The allocation splits roughly as follows: 40.4 percent to a community treasury released against on-chain milestones rather than a fixed calendar, 29.6 percent to investors, 25 percent to the team, 3 percent to the public community sale, and 2 percent to testnet reward claims.

The allocations that matter most for future sell pressure are the investor and team tranches, which together make up more than half of total supply. Both are reported to carry a 12-month cliff from the April 2026 token generation event, followed by a 24-month linear unlock, which places the end of the cliff period around April 2027. Nothing from those two buckets has hit the market yet, worth remembering given how far $AI has already fallen from its debut-day high without any unlock pressure at all. That gap between market cap and fully diluted valuation, roughly eight times, is itself a signal: a large majority of $AI’s theoretical value is not yet in tradeable hands, which is common for a young token but also means today’s price reflects a small, thin slice of eventual supply. What happens to price discovery once investors and team members can start selling into whatever liquidity exists a year from now is an open question, not a resolved one.

The buyback-and-burn mechanism funded by Delphi’s protocol fee is the network’s only current counterweight to that future unlock supply, and it is only as strong as Delphi’s trading volume. A prediction market that keeps growing can plausibly out-burn a slow trickle of vesting tokens. A prediction market that stalls cannot.

Retail Speculation Finds a Serious Infrastructure Token

The token launched with listings across most of the major centralized exchanges within days of the token generation event, a distribution pattern usually reserved for higher-profile launches, and one that gave $AI a liquid, globally distributed trading footprint from day one rather than the slow, thin-liquidity ramp many infrastructure tokens experience. That liquidity cuts both ways. It made $AI easy to buy and sell immediately, which likely amplified both the debut-day spike to $0.1073 and the subsequent slide toward its all-time low. It has also attracted the kind of short-term, momentum-driven attention that infrastructure teams tend to have mixed feelings about: $AI has repeatedly surfaced in exchange “trending” and meme-adjacent trading features aimed at retail speculation rather than at anyone evaluating the network’s actual product usage. None of that changes what Delphi’s settlement volume looks like day to day, but it does mean $AI’s price action over its first hundred days reflects broader retail risk appetite at least as much as it reflects anything happening inside Gensyn’s own network.

Regulatory Footing: Where $AI and Delphi Sit Under US Rules

In March 2026, the SEC and CFTC jointly published an interpretive release sorting crypto assets into five categories: digital commodities, digital collectibles, digital tools, stablecoins, and tokenized securities, with digital commodities defined as tokens whose value comes from programmatic network operation rather than a team’s ongoing managerial effort, according to Forbes‘ coverage of the release. That framework named sixteen specific tokens as digital commodities. $AI was not one of them, and neither AI compute verification nor prediction-market infrastructure got a specific mention anywhere in the release, leaving newer infrastructure tokens like $AI in the same gray zone as much of the rest of the sector, a pattern HOGE Wire has tracked across the broader enforcement landscape: probably closer to a network utility token than a security under the framework’s own logic, but without a specific determination to point to.

Delphi faces a sharper and more specific regulatory question than the token does, because it settles bets on real-world outcomes, which puts it adjacent to a fight already underway between the CFTC and event-contract platforms such as Kalshi. The CFTC proposed a rule in June 2026 updating how it reviews event contracts and defines “gaming” contracts, against a backdrop of Kalshi itself facing more than a dozen lawsuits from state gaming regulators over whether its sports-related contracts are unlicensed betting products. Fielding has tried to pre-empt that comparison with a technical distinction: Delphi, in his framing, runs bidirectional information markets rather than the narrower, unidirectional prediction markets at the center of the CFTC fight, since users can trade information both ways rather than simply betting yes or no on an outcome. Whether US regulators would agree with that distinction is untested, since Delphi has not, as of this writing, been the subject of a public CFTC or state-level action either way.

Gensyn Versus the Other Verifiable AI Approaches

Gensyn is one of several competing answers to the same underlying question: how do you trust an AI computation you did not run yourself? Zero-knowledge machine learning takes the cryptographic route, generating a mathematical proof that a specific model produced a specific output on specific inputs, verifiable in milliseconds regardless of how long the original computation took. It is provably sound but computationally expensive to generate, which has mostly limited zkML to smaller models and narrower use cases so far, a tradeoff HOGE Wire has covered in the context of proving AI outputs without exposing the underlying data.

Optimistic machine learning, or opML, borrows the assume-correct-until-challenged model from optimistic rollups: results are accepted immediately, with a challenge window during which anyone can dispute a result and force an on-chain re-execution to settle the argument. That is conceptually close to what Gensyn’s Verde system does for training disputes, and HOGE Wire’s explainer on opML covers the mechanics in more depth. The difference is scope: opML projects have generally targeted inference verification, while Verde was purpose-built for the harder problem of verifying distributed training runs.

A third camp skips novel cryptography or dispute games altogether and instead borrows Ethereum’s existing economic security. EigenCloud, built on top of EigenLayer’s restaking infrastructure, lets operators stake real collateral behind AI compute claims, so a false result gets punished through slashing rather than through a bespoke verification protocol, an approach HOGE Wire examined in detail when EigenCloud launched. Gensyn’s own verification model actually has more in common with this camp than with zkML: both Verde and EigenCloud’s slashing model are fundamentally economic mechanisms, betting that honest behavior can be made cheaper than dishonest behavior, rather than mathematical guarantees that dishonesty is impossible.

None of the four approaches has definitively won the argument, and it is entirely possible more than one survives: cryptographic proofs for high-stakes, low-frequency verification, and economic or optimistic schemes for high-frequency, lower-stakes work where speed matters more than mathematical certainty. Gensyn’s specific bet is that training verification, not just inference verification, is where the hardest and most valuable version of this problem lives. A hundred days into mainnet, that bet is still waiting for its highest-value workload to come back online.

The Bull Case and the Bear Case

The bull case for Gensyn rests on three things holding up at once. First, Delphi’s actual product-market fit: creator-owned niche markets are a real category that Polymarket and Kalshi have not prioritized, and a 1.5 percent creator fee is a meaningfully better economic deal than most content platforms offer. Second, the buyback-and-burn mechanism gives token holders direct exposure to that volume instead of a vague promise of future utility. Third, Gensyn’s research bench and a16z-caliber backing mean an RL Swarm relaunch, or an entirely new flagship application, is a realistic possibility rather than wishful thinking, especially with a fixed 10 billion token supply that never inflates further.

The bear case is just as concrete. The network depends on a single live application for essentially all of its fee revenue, and that application competes in a category, event and information markets, that is drawing direct regulatory attention in the US, even if Delphi’s specific legal framing has not yet been tested. The core training-verification technology that justified the network’s early valuation is not currently running a production workload. More than half of total token supply is still locked behind a cliff that expires around April 2027, at which point investors and team members who bought in or vested at a much higher implied valuation get their first real chance to sell. And the wider decentralized compute sector, judging by Bittensor’s market capitalization relative to everyone else’s, currently rewards ranked AI outputs at scale far more than it rewards verified training. Put together, the bear case is less “Gensyn is broken” than “Gensyn has not yet had to prove its hardest technology under real load, while its easiest-to-build product carries risks that have nothing to do with machine learning at all.”

What to Watch Through the Rest of 2026

A short list of concrete, checkable signals for anyone tracking Gensyn from here:

• Whether RL Swarm resumes regular scheduled training runs, or stays paused indefinitely while Delphi absorbs engineering resources.
• Whether the Uniswap V3 deployment proposed for Gensyn’s L2, first raised on the Uniswap governance forum in March 2026, actually goes live, since it is meant to deepen on-chain $AI liquidity ahead of the 2027 unlock cliff.
• Delphi’s cumulative mainnet trading volume and active market count, once Gensyn or an independent tracker publishes a figure comparable to the testnet’s $4.88 million sports-market benchmark.
• Any move by the CFTC or a state regulator that touches AI-settled information markets specifically, given the parallel fight already underway with Kalshi over event contracts.
• Whether a completed third-party security audit of the mainnet deployment surfaces publicly, given how much value now sits behind a bridge design that HOGE Wire and others have flagged as one of 2026’s most consistently exploited categories of infrastructure.

Frequently Asked Questions

What is Gensyn and what does its $AI token do?

Gensyn is a decentralized machine learning network built on an OP Stack rollup anchored to Ethereum. Its verification research aims to let untrusted, geographically distributed hardware run AI training jobs that can still be checked for honesty without re-running the entire computation. The $AI token, launched April 29, 2026, pays network fees, backs a buyback-and-burn mechanism funded by Delphi’s trading fees, and carries governance weight over network parameters.

Why did Gensyn pause RL Swarm?

Gensyn’s own documentation confirms that RL Swarm and all Gensyn-hosted training nodes have been paused, with engineering focus consolidated on Delphi ahead of and after the mainnet launch. The company has not published a public timeline for resuming regular RL Swarm training runs.

How does Delphi decide who wins a bet?

Delphi lets a market creator choose between a standard AI settlement and a verifiable one. Verifiable markets use Gensyn’s Reproducible Execution Environment to generate a cryptographic receipt of the model, prompt, and computation used, which anyone can independently rerun to confirm. Non-verifiable markets settle the way most AI products work today, by trusting the operator’s reported output.

Is $AI a security under US law?

There is no specific SEC or CFTC determination on $AI. The two agencies’ March 2026 joint interpretive release sorted crypto assets into five categories, including a digital commodity bucket for tokens whose value derives from programmatic network operation rather than managerial promises, but the release named only sixteen specific tokens and did not address AI-compute or prediction-market infrastructure tokens directly. $AI’s fee-driven, buyback-funded design points toward the digital commodity side of that line, but that is an analytical read, not a regulatory ruling.

How is Gensyn different from Bittensor or Akash?

Akash and io.net are primarily GPU rental marketplaces where trust is enforced through reputation and escrow. Bittensor runs an incentive marketplace where miners produce machine intelligence outputs and validators rank them. Gensyn’s stated goal is narrower and more technical: cryptographically and economically verifying that a specific training computation happened honestly, without requiring the verifier to redundantly re-run the whole job.

Reporting by the HOGE Wire markets desk.