Gensyn in 2026: Who’s Really Winning Decentralized AI Training?

Gensyn was built on one audacious bet: that you can train artificial intelligence across a scattered network of untrusted machines, prove the work was done correctly, and never once route it through a hyperscaler’s data center. Five years after Ben Fielding and Harry Grieve founded the company in London in 2020, that bet is being tested in the open, and the scoreboard is messy. The protocol’s flagship training product, RL Swarm, sits paused. Its only revenue-generating application is a prediction market. And its $AI token, after printing an all-time low on 16 September 2026, trades near $0.02 with a market capitalization under $30 million, according to CoinGecko.

While Gensyn regrouped, the category it helped define got crowded and, in places, genuinely commercial. Prime Intellect closed a $130 million Series A in July at a $1 billion valuation and says it runs past $100 million in annualized revenue. Nous Research trained a 40-billion-parameter model over the open internet. Pluralis is chasing a design in which no single participant ever holds a model’s full weights. The question for 2026 is no longer whether decentralized training can work. It is who is shipping it, and whether Gensyn, the company that made the idea investable, is still setting the pace.

This is a map of the decentralized-training race as it stands in late September 2026: what Gensyn actually built, why its training network went quiet, who is reaching customers first, and what the $AI token is really pricing in.

The Bet That Started a Category

Gensyn’s premise is simple to state and brutally hard to execute. Machine learning is bottlenecked by access to GPUs, and that supply sits concentrated inside a handful of clouds. Meanwhile, idle accelerators are everywhere: in gaming rigs, university labs, and half-empty data centers. Pool that latent supply into one trustless market, the thesis goes, and you expand the compute available for training by an order of magnitude or more.

That vision is what drew a16z crypto to lead Gensyn’s $43 million Series A in 2023. General partners Ali Yahya and Guy Wuollet argued in their investment note that “Gensyn can potentially 10-100x the available compute power for machine learning,” framing the network as infrastructure for an open, permissionless AI economy.

The catch is verification. Renting remote GPUs is a solved problem; Akash and io.net already do it. The unsolved problem is proving that an anonymous machine on the far side of the internet actually ran the training step it billed you for, rather than returning plausible garbage to collect the reward. Fielding has called this the heart of the company. “That’s the big secret sauce behind Gensyn, we’ve solved that problem for machine learning training specifically,” he told Decrypt. Grieve, in the same interview, put it more plainly: “We have a very acute machine learning problem that needed a decentralized trust layer.”

Investors kept backing the thesis. Gensyn has raised roughly $78 million since 2020, including the 2023 Series A, an a16z-led venture round in October 2025, and a December 2025 token sale, with Galaxy Digital and CoinFund among the backers, according to The Block.

The Verification Stack Gensyn Actually Owns

Strip away the token and the prediction market, and Gensyn’s most defensible asset is a verification stack that almost no rival is building. The protocol is organized into four layers: a Reproducible Execution Environment (REE) that produces bitwise-identical results across different hardware, a trustless verification layer built on a system called Verde, an Agent eXchange Layer for encrypted peer-to-peer messaging, and an on-chain coordination layer that runs as an OP Stack Layer 2.

The clever piece is how disputes get settled. RepOps, Gensyn’s reproducible-operators library, forces a fixed order of floating-point operations so that a matrix multiplication returns the exact same bits whether it runs on a consumer card or a data-center accelerator. That removes the hardware non-determinism that would otherwise make any two machines disagree for innocent reasons. On top of that, Verde plays a referee game: when two providers submit different results for the same computation, a referee bisects the computational graph down to the single operation where they diverge, then re-runs only that one operation to see who lied. As Gensyn’s research team put it, “if two compute providers disagree on an ML execution, a referee can determine which provider is honest by executing a single operation rather than running the entire task.” The guarantee holds as long as at least one provider is honest, a design spelled out in the team’s Verde paper.

This is no longer just theory. Verde and RepOps are deployed inside Judge, Gensyn’s verifiable model-evaluation system, and the smart contracts underneath have been through external review. Trail of Bits has published four audit reports covering Gensyn’s Buyback-and-Burn Vault, its bridged token, the Delphi dynamic parimutuel markets, and the ERC-20 token itself, per Trail of Bits. One caveat worth keeping in mind: those audits examined the Solidity contracts, not the AI-verification math, so the referee game’s guarantees rest on the research, not on a security firm’s sign-off.

Running its own Layer 2 also plants Gensyn in the middle of crypto’s broader fee economics, where rollups are fighting a running battle to compress transaction costs. Cheap settlement matters for a network that wants machines, not just people, transacting thousands of times a day.

RL Swarm: The Training Product That Went Quiet

If the verification stack is the moat, RL Swarm was supposed to be the castle. It is an open framework, published on Gensyn’s testnet, that lets anyone anywhere join a distributed reinforcement-learning swarm where models, in the project’s phrasing, learn faster together than alone. The concept matured over 2025: an early Reasoning Gym gave way in November to CodeZero, which extended the swarm toward cooperative coding agents with Solver, Proposer, and Evaluator roles.

Here is the uncomfortable part. As of 22 September 2026, the RL Swarm documentation still reads, “There are no official swarms running right now,” and points visitors toward Delphi instead. The two experimental workloads that once ran alongside it, BlockAssist and CodeAssist, have been sunset, their data left on chain. The product that most directly embodies Gensyn’s founding mission, decentralized training across untrusted machines, is currently switched off.

That is not the same as failure. RL Swarm proved that thousands of heterogeneous nodes could join a coordinated training run over the internet, which is a real engineering result. But proving the plumbing works is different from operating a persistent, paying training service, and Gensyn has not crossed that line. For now, RL Swarm is a demonstration in standby, not a business.

Delphi and the Quiet Pivot to Prediction Markets

What Gensyn does operate, and monetize, is Delphi. Launched on mainnet on 22 April 2026, Delphi is an AI-settled information market: anyone can spin up a market on any question, from a Bitcoin price target to a sports result to a geopolitical event, and AI models resolve the outcome instead of a human oracle committee, as Bitcoin.com News reported at launch.

Settlement is tiered. Fast, long-tail markets can be resolved by a hosted foundation model, while higher-stakes markets use the REE to produce a settlement receipt (essentially the model, the full prompt, and the appended data) that anyone can re-run to check the verdict. Pricing uses a dynamic parimutuel design rather than a fixed order book. The fee model is where the token comes in: Delphi charges 2% of volume, of which 1.5 points go to the market creator and 0.5 points flow to an AI BuyBack Vault that swaps fees into $AI and burns 70% of what it collects, sends 29% to a community treasury, and pays 1% to the executor.

Delphi’s contracts, in Gensyn’s words, “do not distinguish between humans and agents,” which makes it a natural venue for autonomous trading bots. That places Delphi squarely inside the wider move toward AI agents that run DeFi positions without a human at the keyboard, and it leans on the same account infrastructure explored in our smart-account field guide. Fielding has been careful to say Delphi is not trying to beat Polymarket or Kalshi head-on; he told The Block it targets “niche, creator-owned markets those platforms would never build.”

The strategic tension is hard to miss. A company that raised money to decentralize AI training now earns its keep from a prediction market. Delphi does use the verification stack, so it is not a clean break from the mission, but it is a long way from Yahya and Wuollet’s promise of 10 to 100 times more compute for machine learning.

Meet the Field: Four Roads to Decentralized Training

Step back and the decentralized-training race in 2026 has roughly four serious contenders, each betting on a different hard part of the problem. Gensyn bets on verification. Prime Intellect bets on commercial breadth. Nous Research bets on the training algorithm itself. Pluralis bets on a radical ownership model. Around them sits a cohort of compute marketplaces (Akash, io.net, and Bittensor’s subnets) that rent GPUs or coordinate incentives but do not solve trustless training the way these four aim to.

The table flattens a lot of nuance, so the sections that follow unpack where each contender is genuinely ahead and where the marketing outruns the engineering.

Prime Intellect, the Commercial Frontrunner

If any project has turned decentralized training into a business, it is Prime Intellect. The San Francisco company raised a $130 million Series A on 8 July 2026, led by Radical Ventures with NVIDIA Ventures, Intel Capital, Dell Technologies Capital, and Iconiq participating, at a $1 billion valuation, according to TechCrunch. The same reporting pegs the company at a $100 million annualized revenue run rate reached in under a year.

Its pitch is breadth: compute access, a reinforcement-learning framework, training environments, evaluations, and deployment, sold as a modular stack rather than a single product. In its own Series A announcement, Prime Intellect says it serves more than 6,000 customers, including leading AI startups and enterprises, and highlights a case study with the fintech Ramp, whose subagent “beat the frontier models on accuracy while running at faster speeds and a fraction of the cost.”

David Katz of Radical Ventures framed the investment around economics rather than ideology. “They’ve stitched this together and built it in such a way that they’re operating at the frontier in a way that’s affordable,” he said, describing Prime Intellect as a one-stop shop rather than a partial toolkit. Chief executive Vincent Weisser casts the mission in almost civic terms: “It shouldn’t just be a few nerds in a glass tower in San Francisco that have the capability to train AI models. It should be every enterprise, every nation state.”

There is a revealing wrinkle in the flagship model, though. Prime Intellect’s INTELLECT-3, a 106-billion-parameter mixture-of-experts model, was trained on a concentrated, high-bandwidth cluster of NVIDIA H200 GPUs, not an internet-wide swarm of volunteer machines. Even the company most aggressively commercializing decentralized training reached for a centralized cluster to build its best model. And, tellingly, it did all of this without a token.

Nous Research and Training Across the Open Internet

Nous Research is the contender closest to Gensyn’s original dream of training over the open internet, and it is arguably furthest along at doing it. Its Psyche network, built on Solana, coordinates fault-tolerant training across globally distributed machines, using in-house optimizers called DisTrO and DeMo that slash the bandwidth normally required to keep thousands of GPUs in sync.

The proof point is Consilience, a 40-billion-parameter model that Nous says was pre-trained over the internet on 20 trillion tokens, which the lab calls the largest distributed pre-training run ever, in a write-up on the next phase of Psyche. Nous also ships the widely used Hermes family of open models, giving it a developer following that most compute networks lack.

On money, Nous raised $50 million in a round led by Paradigm in 2025 at a $1 billion valuation, per SiliconANGLE, and has been reported to be raising more at a higher mark since. The philosophical contrast with Gensyn is sharp. Nous attacks the communication problem, making distributed training fast enough to be practical, and largely trusts its own coordination. Gensyn attacks the trust problem, assuming the machines are adversarial and building proofs accordingly. Both could be right; they are solving different halves of the same puzzle.

Pluralis and the Long Tail of Protocol Learning

Pluralis Research is the smallest and most conceptually ambitious of the four. Its approach, called Protocol Learning, is low-bandwidth, model-parallel training in which the model is split across many participants and, crucially, is never fully materialized in any single place. No one holds the complete weights, which the team frames as a way to keep a model open and un-seizable while still letting contributors earn a programmatic share of the result.

The design is the most crypto-native of the bunch: contribute compute, own a slice of the model. Pluralis announced a $7.6 million seed round co-led by Union Square Ventures and CoinFund, with Topology, Variant, Eden Block, and Bodhi Ventures joining, plus angel checks from Balaji Srinivasan and Hugging Face co-founder Clement Delangue, according to a GlobeNewswire release. It has no product in market yet, which makes it a bet on research rather than traction, but the ownership mechanism is the most direct answer to the question of why a training network needs a token at all.

The Bandwidth Wall, and Why Everyone Still Uses Data Centers

To understand why RL Swarm can be paused without the sky falling, you have to understand the bandwidth wall. Training a large model means constantly synchronizing gradients across every GPU, and over the public internet, latency and limited bandwidth throttle that synchronization far below what an InfiniBand-linked cluster achieves. This is the physical reason decentralized training has lagged the frontier for years.

The field’s best ideas attack the synchronization frequency. Google DeepMind’s DiLoCo showed that machines can train mostly independently and reconcile only occasionally. Nous’s DisTrO and DeMo push the same principle further. These methods genuinely narrow the gap. But the ceiling is still there, and it shows up in the choices even the leaders make: Prime Intellect built INTELLECT-3 on a concentrated cluster, not a swarm. The honest state of the art in 2026 is that decentralized methods are competitive for smaller models, reinforcement-learning post-training, and fine-tuning, while frontier-scale pre-training still wants a data center.

That reframes the entire race. The prize is not replacing hyperscalers overnight. It is capturing the fast-growing slice of workloads (reinforcement learning, post-training, agent tuning, and inference) that do not need a monolithic cluster and can tolerate a distributed, verifiable backend. On that narrower and more realistic battlefield, Gensyn’s verification bet looks less like a detour and more like a wager on where value accrues once compute is genuinely untrusted and spread out.

Token vs Equity: Two Ways to Finance a Compute Network

The clearest fault line in this race is not technical, it is financial. Gensyn launched the $AI token, with a token generation event on 29 April 2026, to bootstrap a two-sided market and reward compute providers directly. Prime Intellect took the opposite path, building a nine-figure revenue business on equity with no token at all. Nous sits in between, financing with equity plus token-style instruments. Pluralis makes the token the whole point, since contributors are paid in ownership of the model they help train.

Each choice carries a tax. A live token can jump-start a network and align incentives, but it also creates a price to defend and a supply schedule to survive. Gensyn’s $AI has a maximum supply of 10 billion, and its allocation leans heavily toward insiders, with investors and the team together holding well over half the supply.

Only about 13% of the supply is circulating today, which means the burn from Delphi fees is fighting against a much larger future unlock. Insider tokens face a cliff roughly a year after the token generation event, in the spring of 2027, followed by a linear release. At current Delphi volumes, the buy-and-burn simply is not large enough to offset that overhang, a dynamic that separates a token with real cash-flow-backed yield from one that mostly runs on narrative. Prime Intellect’s $100 million revenue run rate, earned without any token, is a pointed reminder that the token itself is a choice, not a requirement.

What the $AI Token Is Pricing In

The market’s verdict on all of this is unforgiving but not unique to Gensyn. As of 22 September 2026, $AI trades around $0.02121, for a market capitalization of roughly $27.7 million and a fully diluted valuation near $212 million, on 24-hour volume of about $3.3 million, per CoinGecko. The token printed an all-time low of $0.01790 on 16 September and has since bounced, sitting more than 18% above that floor and about 8.6% higher over the trailing week, though still down roughly 80% from its April 2026 peak of $0.1073.

The gap between a $27.7 million market cap and a $212 million fully diluted valuation is the supply overhang made visible: the market is discounting heavily for tokens not yet in circulation. Yet the drawdown is a cohort story, not a Gensyn-specific indictment. Every AI-compute token has been repriced hard from its highs (cohort figures via CoinGecko).

Read together, the cohort tells you the market has soured on AI-compute tokens broadly, from the $3.6 billion incumbent Bittensor down to sub-$60 million minnows. $AI is the youngest and smallest of the group, which cuts both ways: more room to grow if Gensyn executes, and less liquidity to cushion the coming unlocks if it does not.

The Regulatory Picture and the SEC

For US readers, Gensyn sits in a genuinely unsettled corner of the rulebook, and it helps to separate two very different questions. The first is whether the $AI token is a security. Gensyn is a UK-domiciled company and $AI carries no US securities registration; the Securities and Exchange Commission has not brought any action against Gensyn, and under the SEC’s softer 2026 posture, token enforcement has cooled from its peak. That does not settle the classification question, which still turns on how a token is marketed and whether buyers expect profit from the efforts of others.

The second question is about Delphi, and it points at a different agency. In the United States, prediction markets and event contracts fall largely under the Commodity Futures Trading Commission, not the SEC, as the running fights over Kalshi and Polymarket have made clear. An AI-settled market that resolves real-world events looks a lot like an event-contract venue, and the fact that Delphi’s contracts do not distinguish between humans and autonomous agents raises novel questions about market integrity and surveillance that no regulator has fully answered.

Layered on top is the global compliance backdrop, where anti-money-laundering enforcement has grown teeth through measures like the ones we examined in our look at Section 311 and FATF pressure. A permissionless market that anyone, or any bot, can trade is exactly the kind of surface that draws that scrutiny. None of this is an active problem for Gensyn today, but it is the regulatory weather the project is building in.

The Bull and Bear Case for Gensyn

The bull case for Gensyn is that it owns the hardest, most durable problem in the category. Cheap GPUs are a commodity; trustless proof that remote compute ran correctly is not, and it is exactly what a world of autonomous agents and untrusted machines will need. Verde and RepOps are in production inside Judge, the contracts are audited, and Delphi is generating real fees that flow back into buying and burning the token. With a16z crypto, Galaxy Digital, and CoinFund behind it, Gensyn has the balance sheet to keep building while the market is cold.

The bear case is just as concrete. The founding product is paused, the revenue comes from a prediction market rather than training, and rivals are reaching customers and frontier-scale models without leaning on a token at all. A large insider unlock arrives in 2027 against a market capitalization under $30 million, the token is down roughly 80% from its high, and Delphi has to carve out room in a prediction-market niche while regulators circle the whole category. The stack may be brilliant and still lose to teams that simply ship models people pay for.

Both cases can be true at once. Gensyn may well have built the best verification technology in decentralized AI and still find that the market rewards Prime Intellect’s revenue and Nous’s models first. Owning the hardest problem is only an advantage if the market decides that problem is the one worth paying to solve.

What to Watch Between Now and 2027

A handful of concrete signposts will tell you which way this is breaking well before the 2027 unlock forces the issue.

• A relaunch of RL Swarm as a persistent, paying training service, which would revive the founding thesis rather than leave it in standby.
• The first external integration of REE, Verde, or Judge by a project other than Delphi, which would prove the verification stack is a platform and not just an internal feature.
• Delphi volume growth large enough for the buy-and-burn to meaningfully offset emissions, the single cleanest signal that the token flywheel works.
• The spring 2027 insider unlock, and whether the market absorbs it or is swamped by it.
• Any concrete CFTC or SEC signal on AI-settled prediction markets, which would reprice Delphi’s regulatory risk overnight.
• Whether rivals like Prime Intellect or Nous adopt verification of their own, which would either validate Gensyn’s bet or commoditize it.

Until then, the honest summary is that Gensyn built the most sophisticated answer to a problem the market has not yet decided to prioritize, while nimbler rivals win the customers in front of them. The technology is not the question. Timing is.

Frequently Asked Questions

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

Gensyn is a decentralized computing protocol, founded in London in 2020, that aims to let anyone train and verify machine-learning workloads across a network of untrusted machines instead of a centralized cloud. The $AI token is the network’s native asset: it pays for and rewards compute, coordinates the on-chain layer, and absorbs a share of fees from Gensyn’s live product, Delphi, through a buy-and-burn mechanism.

Is Gensyn’s RL Swarm still running in 2026?

No. As of 22 September 2026, Gensyn’s documentation states that there are no official RL Swarm training runs operating, and it directs users to Delphi instead. RL Swarm proved that thousands of machines could join a coordinated training run over the internet, but it has not become a persistent, paying service, and the two related workloads, BlockAssist and CodeAssist, have been retired.

How is Gensyn different from Prime Intellect and Nous Research?

The three attack different parts of the problem. Gensyn focuses on trustless verification, proving that remote compute ran correctly. Nous Research focuses on the training algorithm, using low-bandwidth optimizers to make internet-scale training practical, and has trained a 40-billion-parameter model that way. Prime Intellect focuses on commercial breadth, selling a full stack of compute, reinforcement learning, and evaluation, and reports over $100 million in annualized revenue without issuing a token.

What is Delphi and how does it make money?

Delphi is Gensyn’s AI-settled information market, live on mainnet since April 2026, where anyone can create a market on any question and AI models resolve the outcome instead of a human oracle. It charges a 2% fee on volume, with 1.5 points going to the market creator and 0.5 points flowing to a vault that buys and burns $AI, sending most of the remainder to a community treasury.

Why has the $AI token fallen so much?

$AI reached an all-time low of $0.01790 on 16 September 2026 and trades around $0.02 today, down roughly 80% from its April 2026 peak. The decline reflects a broad repricing of AI-compute tokens, a large supply overhang with only about 13% of tokens circulating and a major insider unlock due in 2027, and the reality that Gensyn’s flagship training product is paused while its revenue comes from a still-small prediction market.

Marcus Okafor covers artificial intelligence and crypto infrastructure for HOGE Wire.