Gensyn in 2026: A Training Network Running on Prediction Markets

In June 2023, Andreessen Horowitz led a $43 million round into a London startup that promised to do for machine learning what Bitcoin did for money: turn the world’s idle graphics cards into a single permissionless training supercomputer. Two of the firm’s partners wrote at the time that the network could “10-100x the available compute power for machine learning.” Three years on, the company has a live mainnet, a token listed on major exchanges, and a product that generates real fees. That product is not AI training. It is a prediction market.

The distance between the pitch and the shipped product is the single most important thing to understand about Gensyn in September 2026. RL Swarm, the training system meant to prove the thesis, has been paused for months. Delphi, an AI-settled information market that barely featured in the original story, is the only thing on the network with meaningful economic activity. The $AI token trades near its all-time low, worth roughly a fifth of its April debut price. This is the story of how a decentralized-training company became, in practice, a prediction-market company, whether that is a detour or a destination, and what it means for anyone holding the token.

What Gensyn was built to be

Gensyn was founded around 2020 in London by Ben Fielding and Harry Grieve, who met through the Entrepreneur First talent program. Their starting point was not a token idea looking for a use. It was a specific, hard machine-learning problem: if you rent compute from a stranger’s GPU somewhere on the internet, how do you know they actually ran your training job correctly rather than returning plausible-looking garbage to collect the fee? Centralized clouds solve this with reputation and contracts. A permissionless network has neither.

“We came from a machine learning and deep learning background,” Grieve told Decrypt when the Series A closed. “We have a very acute machine learning problem that needed a decentralized trust layer.” Fielding put the claim more bluntly in the same interview: “That’s the big secret sauce behind Gensyn, we’ve solved that problem for machine learning training specifically.” The mechanism behind that claim is Verde, a verification scheme described in a February 2025 paper co-authored with NYU’s Joseph Bonneau, which lets the network re-check a tiny slice of a disputed computation and catch a dishonest worker cheaply, paired with RepOps, a library of reproducible operators that force bitwise-identical math across different hardware. Verifiable compute is precisely the difference between renting GPUs and having to trust them, a distinction we cover in depth in our guide to verifiable compute as the trust layer for AI agents.

Investors bought the vision. Gensyn raised about $1.1 million in a 2021 pre-seed, $6.5 million in a 2022 seed, then the $43 million Series A led by a16z crypto in 2023, and finally a roughly $16 million community sale in December 2025, according to fundraising trackers, for something near $67 million in total. Grieve has described the end goal as nothing less than “the economic foundation of a parallel machine civilization.” a16z’s Ali Yahya and Guy Wuollet, in their investment memo, framed the upside in raw capacity terms: by tapping smaller data centers, gaming rigs, and personal devices, Gensyn “can potentially 10-100x the available compute power for machine learning.” That is the company Gensyn raised money to be. It is not, at least not yet, the company that exists.

RL Swarm went quiet

The flagship demonstration of the training thesis was RL Swarm, a testnet system that let volunteers around the world collaborate on reinforcement-learning post-training, improving a model together over the open internet. It was iterated on repeatedly: the original Reasoning Gym environment was replaced in November 2025 by CodeZero, which recast the work as programming challenges with Solver, Proposer, and Evaluator roles, while a separate component called Judge, launched in August 2025, added cryptographically verifiable AI evaluation on top of Verde.

All of that engineering is now idle. Gensyn’s own documentation states plainly: “There are no official swarms running right now.” The page then points would-be participants somewhere else entirely, suggesting they explore Delphi instead. Read that sequence again, because it is remarkable. The canonical instructions for the product that justified a nine-figure valuation now redirect users to the prediction market. Gensyn frames this as focus rather than retreat: RL Swarm was always a testnet experiment, and the team’s attention is on shipping Delphi as the first real application on mainnet. But the practical reality is that the training network is not training anything for the public today, and the company is telling its most technical users to go trade markets instead.

Why decentralizing training is so hard

To understand why RL Swarm is paused while Delphi hums along, it helps to see what makes decentralized training uniquely brutal. Training a large model is not one big calculation; it is millions of small ones that constantly exchange results. Gradients have to be synchronized across every device on every step, and that synchronization is bandwidth-hungry and latency-sensitive. Inside a single data center, GPUs are wired together with high-speed interconnects measured in hundreds of gigabytes per second. Scatter those same GPUs across gaming rigs and small data centers on ordinary internet links, and the communication overhead can swamp the compute you were trying to harvest.

This is why even well-funded distributed-training efforts have tended to focus on smaller models or on specific phases like reinforcement-learning post-training, which tolerate looser coordination, rather than pretraining a frontier model across the open internet. Gensyn’s Verde solves the trust half of the problem, proving that a remote worker did the math honestly. It does not repeal the physics half, moving oceans of gradient data between far-flung machines. Settlement and inference, by contrast, are far less coordination-heavy: resolving a market is a single, self-contained model call. That asymmetry is a large part of why the prediction market shipped and works while the training network sits quiet. RL Swarm being paused is not just a scheduling footnote; it is a signal of how genuinely difficult the founding mission remains.

Delphi is the whole story now

Delphi went live on Gensyn’s mainnet on 22 April 2026, the network’s first application with real money moving through it, as The Block reported at the time. Gensyn is careful to call it an information market rather than a prediction market, and the distinction is not marketing fluff. On Polymarket or Kalshi, a small team decides which markets exist. On Delphi, anyone can create a market on any question, own it like a small business, and collect 1.5% of everything that trades through it. Audiences are not just bettors; they are information traders, and the person who framed the question gets paid. The platform ran on testnet from December 2025 and, by Gensyn’s account, already recorded millions of dollars in trading volume before mainnet. It is built as an OP Stack Layer 2 on Ethereum.

The timing is not an accident. Prediction markets have become one of crypto’s few genuine consumer hits: markets tracked by the data provider Artemis were on pace for close to $500 billion of annualized volume based on mid-2026 figures, per FalconX. Gensyn’s wager is that the binding constraint on that market is not liquidity or user experience, both of which incumbents have largely solved, but settlement: the messy, disputable act of deciding who won. That is the problem Gensyn already claimed to solve for training, repackaged for markets. In the company’s framing, quoted by news.bitcoin.com, “AI models can participate directly as predictors, earning from accurate resolutions.”

How an AI actually settles a market

This is the technical heart of the pivot, and it is worth being precise about. FalconX’s July 2026 write-up describes two distinct settlement paths in Delphi. The first is non-verifiable: a market can be resolved by an off-the-shelf cloud model (FalconX names a frontier Claude model as the example) that reads live data from the internet and returns a verdict. It is fast and convenient, but the run is not reproducible and there is no cryptographic proof of how the answer was reached, so users have to trust the operator.

The second path is where Gensyn’s original technology re-enters. Using the Reproducible Execution Environment (REE), a creator runs the settlement prompt through a locally hosted model that, in FalconX’s words, “produces a settlement receipt, which is a record containing the model used, the full prompt and appended data, and a hash of the computation.” Anyone can take that receipt, re-run the same model on the same inputs, and confirm the result independently. Settlement is automatically scheduled for 24 hours after a market closes. This is Verde and Judge redeployed from verifying training runs to verifying oracle calls, and it is the reason Gensyn can argue Delphi is more than another prediction-market front end. Because the contracts, as the team likes to note, do not distinguish between humans and software, the same rails support agentic trading, where autonomous bots create, trade, and settle markets without a person in the loop, a pattern that fits neatly with the broader move toward native account abstraction.

The credibly neutral AI oracle bet

Strip away the tokenomics and the branding and Gensyn is making one genuinely interesting bet: that a large language model can be a more neutral judge of reality than a crowd of token holders. Prediction markets live or die on resolution, and the incumbent model has a known weakness. Polymarket settles disputes through the UMA token’s optimistic oracle, and 2025 produced a run of contested resolutions in which large UMA holders were accused of voting their positions rather than the facts, the kind of on-chain incentive problem we examine in our guide to MEV and the bots that reorder your trades. When the people who settle a market also hold stakes in its outcome, the notion of truth becomes negotiable.

Gensyn’s research argues for a different anchor. In an essay on what it calls credibly neutral AI oracles, the team frames the dilemma directly: “Centralize the oracle and you’ve reintroduced a single point of failure; decentralize and you’ve opened the system to adversarial manipulation.” Its proposed escape is an AI adjudicator whose output depends only on the ground-truth evidence, not on who holds what, with the process designed so that overturning a correct call requires an overwhelming supermajority rather than a simple bought majority. Whether that holds up under real adversarial pressure is unproven; prompt injection, ambiguous evidence, and model drift are all live risks. But it is a serious attempt at a real problem rather than a slogan. Delphi’s pricing and settlement contracts were audited by Trail of Bits ahead of launch, with the full technical reports promised alongside a longer writeup; readers who want to understand what that kind of review can and cannot guarantee should see our explainer on formal verification and the audits that prove code.

The buy-and-burn machine

Every Delphi trade feeds a token mechanism that Gensyn has built with some care. The protocol takes roughly 2% of trading volume as a fee. Of that, 1.5 percentage points go to the market creator, usually paid in stablecoins at settlement, and the remaining 0.5 percentage point flows into what Gensyn calls the AI BuyBack Vault. The vault then splits three ways, according to news.bitcoin.com: 70% of it is used to buy $AI and permanently burn it, 29% goes to a Community Treasury for development and grants, and 1% pays whoever executes the operation. The buyback swaps Delphi’s collected fees into $AI on Uniswap V3, which was deployed to the Gensyn L2 in March 2026 precisely to give the vault somewhere to trade.

It is an elegant loop on paper: usage buys and burns the token, so more usage should mean a scarcer asset. The problem is scale. Only half a percentage point of volume reaches the vault, and only 70% of that is actually burned, which means the burn captures about 0.35% of Delphi’s trading volume. Set that against a token whose fully diluted value sits near $195 million and whose emission schedule will release billions of new tokens over the coming years, and the arithmetic is unforgiving. To burn enough to counterbalance future supply, Delphi would need to sustain trading volumes in the many billions of dollars per year, orders of magnitude above where a young platform with a few million in tracked activity sits today. The machine is built, and at a trickle it is spinning. It is nowhere near fast enough to matter yet.

A token stuck at all-time lows

The market has noticed the gap between ambition and traction. $AI changes hands at about $0.019, giving it a market capitalization near $25 million and a rank around #714, according to CoinGecko. That is roughly 82% below the $0.1073 it touched on 29 April 2026, days after its token generation event, and it printed a fresh all-time low near $0.019 on 30 August. There has been no shortage of distribution: the token has been widely listed, and in late August it surfaced in Robinhood’s Daily Pulse and Trending discovery feeds, a sign of retail attention that arrived even as the price kept sliding. Visibility is up and the chart is down, which is its own kind of signal.

Gensyn is not alone in that divergence. The entire cohort of AI and decentralized-compute tokens has spent 2026 grinding lower even as the underlying networks shipped product. Bittensor’s TAO trades around $229, off roughly 70% from its 2024 peak; Akash’s AKT sits near $0.51, down about 94% from its 2021 high; io.net’s IO is close to $0.13, a punishing 98% below its 2024 debut. Thin risk appetite has not helped, with a hawkish macro backdrop keeping speculative capital cautious through the autumn, a dynamic we track in our coverage of the September rate-decision countdown. But even within a bruised sector, Gensyn stands out for how little of its supply is actually in the market.

The supply overhang nobody is pricing

The all-time-low headline actually understates the risk, because the more important number is not the price but what sits above it. Only about 13% of the 10 billion maximum supply of $AI is circulating today. The rest is a schedule. Per Gensyn’s published allocation on tokenomist.ai, insiders hold a majority of the token: the team’s 25% and investors’ 29.6% add up to 54.6% of everything that will ever exist, and those allocations were locked for twelve months from the April 2026 token generation event, then vest linearly after that. In plain terms, the first large insider unlock wave begins around April 2027, and it keeps flowing for two years thereafter.

This is what the ratio of market cap to fully diluted value, roughly 0.13, is quietly telling you: the market is pricing 13% of the tokens, and the other 87% is an overhang waiting in the wings. The buyback burn, as we saw, offsets a fraction of a fraction of that. And the buyers most exposed are the ones who showed up last: participants in the December 2025 community sale paid about $0.0473 per token, so at today’s $0.019 they are sitting on losses of roughly 60% before the largest unlocks have even started. None of this makes Gensyn a bad protocol. It does mean the token carries a structural gravity that product traction alone will struggle to overcome in the near term.

Where Gensyn sits in the compute cohort

Grouping $AI with TAO, AKT, IO, and RENDER, as most trackers and index products do, obscures more than it reveals. Those networks rent or coordinate GPUs; their revenue, such as it is, comes from compute. Gensyn’s live revenue comes from a prediction market, and its actual compute product is switched off. That makes it an awkward comparison in both directions. Measured as an AI-infrastructure token, it is the smallest and least utilized of the group, with no live training network to point to. Measured as a prediction-market business, its natural peers are Polymarket and Kalshi, not Bittensor, and against those it is tiny.

The generous reading is that Gensyn is two option bets in one wrapper: a near-term information-market business that is actually shipping, and a longer-term decentralized-training protocol that remains dormant but could reactivate if the economics and the demand line up. The skeptical reading is that the market is being asked to fund a training thesis while the company earns its keep doing something else, and that the two may never converge. Either way, the token’s tiny circulating float and thin liquidity mean it trades more on narrative and unlock anxiety than on Delphi’s fee run rate. Compared with peers that at least sell a metered product, Gensyn’s revenue engine is both newer and structurally different, and that difference is not yet reflected in how the token is categorized.

Delphi against the prediction-market incumbents

If Delphi is the business, the competition is fierce and well capitalized. Polymarket is the volume leader, settling through UMA’s oracle and operating under intensifying regulatory attention. Kalshi is a regulated event-contract exchange in the United States that has won important court fights over its right to list contracts on real-world outcomes. Both settle disputes with humans and rules. Delphi’s differentiators are threefold: permissionless market creation, so the long tail of niche questions can exist at all; creator fees, which turn market-making into a revenue stream for ordinary users; and AI settlement, which promises faster, cheaper, and less corruptible resolution than a token vote or a compliance desk.

The risks map directly onto those features. Permissionless creation invites junk, ambiguous, and manipulable markets. Creator fees can be gamed through wash trading to farm volume. And AI settlement, the whole pitch, is exactly where the danger concentrates: an adjudicating model can be wrong, can be fed misleading evidence, or can be steered by a cleverly written market description. The non-verifiable mode reintroduces the trust problem Gensyn set out to eliminate, and the verifiable REE mode, while cryptographically honest, is manual and comparatively slow. Delphi is a real answer to a real weakness in prediction markets, but it trades one set of settlement risks for another rather than making them disappear.

The regulator question

An AI does not launder the legal questions a prediction market raises; if anything, it sharpens them. In the United States, event contracts and prediction markets are principally the domain of the Commodity Futures Trading Commission, which is where Kalshi’s fights have played out and where Polymarket’s US access has repeatedly run into trouble. A platform that lets anyone, anywhere, spin up a market on any real-world event and settle it with a model is close to the archetype that draws both federal derivatives regulators and state gambling authorities. Gensyn is a UK company and the protocol itself is permissionless, but US-facing access, front ends, and fiat on-ramps are where enforcement tends to land.

The $AI token raises a separate question for the Securities and Exchange Commission, namely whether it is an investment contract under the Howey test. Here Gensyn benefits from ambiguity: the March 2026 joint interpretation from the SEC and CFTC that named a slate of digital commodities was silent on AI and DePIN tokens, leaving assets like $AI, TAO, and IO in a gray zone rather than clearly inside or outside securities law. That silence is comfortable for now and precarious later, because a gray zone is a policy choice that can be revised. The net effect is that Delphi inherits every unresolved prediction-market legal risk and adds a novel one about machine settlement, while the token carries the same classification uncertainty as the rest of the sector.

Can the pivot work?

The bull case is coherent. Prediction markets are one of the few crypto products with real, growing consumer demand, and settlement is their genuine soft spot. If Gensyn’s AI oracle proves meaningfully harder to corrupt than a token vote, Delphi can take share, and permissionless creation plus creator fees is a real distribution advantage over curated incumbents. In that world the dormant training network is free optionality on top of a working markets business, and the buyback finally starts to bite as volume compounds. Gensyn would not be the first company to raise money for one thing and find product-market fit in an adjacent one.

The bear case is just as coherent, and shorter. Eighty-seven percent of the token supply is still to come, with the first big insider unlock due around April 2027 and a burn far too small to absorb it. The original moat, verifiable decentralized training, is switched off, so the thing that differentiated Gensyn from a dozen prediction-market clones is not currently running. AI settlement is unproven against determined adversaries. And the token already sits at all-time lows with retail only now discovering it. For the pivot to work, Delphi has to become a genuinely large business fast enough to outrun the supply schedule, which is a high bar.

There are concrete milestones that will tell the story better than the price chart. Watch for a permanent, paying RL Swarm that shows training can actually ship on mainnet; the publication of the full Trail of Bits audit reports; the first external, non-Delphi integration of Gensyn’s settlement oracle, which would prove the technology has value beyond the company’s own product; Delphi volume large enough to make the burn visible on-chain; and, above all, how the market absorbs the April 2027 unlock. Gensyn never announced a pivot. It simply shipped the thing that worked and quietly redirected its users toward it. Whether the market it built can carry the company it raised is, for now, the only question that matters.

Frequently Asked Questions

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

Gensyn is a decentralized network founded in London around 2020 that was designed to let people rent out idle GPUs for verifiable machine-learning work, using a verification system called Verde to prove computations were done honestly. The $AI token is its native asset, used for staking, protocol payments, verification incentives, and governance, and it captures value from the network’s live product, Delphi, through a buy-and-burn mechanism funded by trading fees.

Is Gensyn’s RL Swarm still running?

No. As of September 2026, Gensyn’s own documentation states that there are no official swarms running, and it points users toward Delphi instead. RL Swarm was a testnet system for collaborative reinforcement-learning training; it has been paused for months, which is why the network’s live economic activity comes from the prediction market rather than from AI training.

What is Delphi and how does it settle markets?

Delphi is Gensyn’s AI-settled information market, live on mainnet since April 2026, where anyone can create a market on any question and earn 1.5% of its trading volume. Markets can be resolved two ways: a fast non-verifiable mode using a cloud AI model that you have to trust, or a verifiable mode using Gensyn’s Reproducible Execution Environment, which produces a settlement receipt that anyone can independently re-run to confirm the outcome.

Why is the $AI token near its all-time low?

Several forces overlap. Only about 13% of the 10 billion token supply is circulating, with insiders holding 54.6% under locks that begin releasing around April 2027, so the market is bracing for future dilution. The buy-and-burn mechanism currently captures only about 0.35% of Delphi’s volume, too little to offset that supply. And the broader AI and decentralized-compute token sector has fallen 70% to 98% from its highs during a cautious macro period.

Is Gensyn’s $AI token regulated by the SEC?

Not clearly. The March 2026 joint interpretation from the SEC and CFTC named several digital commodities but said nothing about AI or DePIN tokens, leaving $AI in a regulatory gray zone rather than definitively inside or outside securities law. Separately, Delphi’s prediction markets touch the CFTC’s event-contract jurisdiction and potentially state gambling rules, so the legal picture around Gensyn spans several regulators and remains unsettled. None of this is investment advice.

By Marcus Okafor, senior AI and markets correspondent at HOGE Wire.