In June 2023, when a16z crypto led a $43 million Series A funding round, what Gensyn was doing was still abstract in the eyes of most crypto investors: a decentralized machine learning training protocol. That year, ChatGPT had only been online for six months, the shortage of H100 had not yet become a geopolitical issue, the entire AI infrastructure sector had not yet become the "sector" it is today, Bittensor's token $TAO was less than $10 at the time, and Akash was still in the niche understanding of cloud computing power.
Three years later, things look completely different. The shortage of AI computing power has become a global narrative, with hyperscale cloud vendors' orders booked until 2027, and OpenAI and Anthropic vying for chips for their next-generation models. On April 22, 2026, the Gensyn mainnet officially launched; on April 29, $AI opened simultaneously on Coinbase and OKX; and on May 14, Binance spot trading was launched under the code AIGENSYN. From a protocol white paper that was difficult for outsiders to understand to launching on three major exchanges within three weeks, Gensyn hit a near-perfect window of opportunity.
It's either an excellent plan or exceptional luck. Judging from the backgrounds of the team's two founders, it's highly likely that they have both.
Two people who opposed centralized AI before ChatGPT
Gensyn's two founders, Ben Fielding and Harry Grieve, met in 2020 at the Entrepreneur First accelerator in London. Eight weeks later, they decided to go all-in on one thing: building a decentralized machine learning network.
This decision wasn't particularly appealing in 2020. GPT-3 had just been released, and "AI" was still an occasionally mentioned keyword in the crypto world; most people were chasing the tail end of DeFi summer and the upcoming NFTs. The two made this decision because they each brought specific, firsthand experiences and problems with them.
Ben Fielding holds a PhD in Computer Science, specializing in neural network architecture search and using evolutionary algorithms to automatically optimize the structure of deep networks. In short, what he did during his PhD was "let AI optimize itself." Before his PhD, he worked in database management, and after his PhD, he co-founded a data privacy company. He is someone deeply committed to both "data sovereignty" and "algorithmic autonomy."
Harry Grieve comes from a completely different background: an econometrics major who switched to machine learning during his master's studies, and later led a data research team at a London-based AI startup, where he spearheaded a disaster risk prediction system for the insurance industry. His immediate motivation for starting his own business wasn't seeing a trend, but rather his personal experience of the "extreme difficulty of acquiring large-scale computing power and high-quality data"—a personal experience of being blocked by the computing power barrier.
The two shared two points of consensus: first, machine learning was the next technological wave (this was in 2020; ChatGPT would have to wait another two years); second, they both opposed centralization, Ben from a data privacy perspective and Harry from a computing power monopoly perspective. They initially researched federated learning, which involves training models on distributed data. As they worked on it, they discovered that the core bottleneck wasn't data but computing power, and distributed computing power needed to solve the problem of "how to distribute tasks and verify results among mutually untrusted machines," which led them to introduce blockchain.
This is a path from AI-native to crypto, and this distinction will become increasingly important amidst the "AI x Crypto" hype of 2025-2026. Most DeAI projects' founding teams are web3 veterans who transitioned after seeing the trend emerge, their technical narratives built on a secondhand understanding of deep learning. Gensyn's core team is the opposite: they first became AI researchers, then spent a decade discovering that "centralization" is the core problem of AI, and finally decided to use crypto to solve it.
The team's academic output supports this point. The paper on the core verification protocol Verde has been published on arXiv ( arxiv.org/abs/2502.19405), and others include NoLoCo (communication efficiency optimization), CheckFree (distributed training fault recovery), SkipPipe, SAPO, RL Swarm, and the latest "Hail to the Thief" (a study on the attack and defense of decentralized GRPO). In a field where white papers are commonly used instead of academic papers, having more than seven cited research papers is itself a rarity.
What Gensyn is doing goes on a much larger scale than "decentralized computing power".
In its highest-level summary, Gensyn is a network for machine intelligence. This may sound abstract, but it essentially boils down to one thing: for AI training, inference, and collaboration to run in a decentralized environment, three things must happen simultaneously: machines must be able to find each other, participants must have traceable identities, and computational results must be verifiable. Gensyn doesn't aim to be just one layer; it aims to build all three simultaneously.
To use a familiar but effective analogy, if the upper-layer AI applications—agents, models, and various prompt-packaged products—are like electrical appliances, then Gensyn aims to be the underlying power grid. The power grid may not be the most visible element normally, but without it, scale and efficiency would be severely hampered. Currently, the entire AI x Crypto industry's attention is almost entirely focused on the appliances; few are digging the trenches and laying the cables. Gensyn's target is this trench.
The entire network consists of a three-layer infrastructure: AXL handles communication, Chain handles identity and settlement, and REE handles authentication. Structurally, it looks like this:
The communication and identity layers are relatively easy to understand; the challenge lies in their engineering implementation. The truly difficult, and indeed worthwhile, layer is verification.
REE: Making AI Computation Reproducible
GPU parallel computing is inherently nondeterministic; the same model and the same input will produce different results on different hardware. This may sound counterintuitive, but it's an engineering fact: the order of floating-point operations, the parallel scheduling of the underlying CUDA kernel, and even minor hardware differences between different GPUs can all lead to observable deviations in the final output. This means you can't verify the correctness of someone else's calculation by "running it again yourself."
REE (Reproducible Execution Environment) solves this problem. Through a self-developed deterministic operator library, RepOps, REE enables the same model to produce bit-level consistent outputs on any supported hardware, supporting models with up to 72 bytes of parameters. If two nodes disagree on the computation results, any third party can use REE to rerun the operation on their own hardware to determine who is right and who is wrong; the underlying dispute arbitration protocol, Verde, can further locate the first operation in the computation graph that causes a disagreement and recalculate only the least controversial part.
REE's position becomes clearer when viewed within the broader context of the decentralized AI field:
Decentralized AI has been trying to answer the same question: How can on-chain AI computation results be trusted off-chain? There have been four main approaches to this question before.
One is the decentralized computing power market , such as Akash and io.net , which provides cheap GPUs but does not solve the verification problem.
The second type is zkML , such as Ritual and Giza, which use cryptography to prove the correctness of computations. The downside is that the performance overhead is extremely high, and it cannot run large models.
Thirdly, there are TEEs like Phala, which use hardware to isolate the execution environment. The downside is that you have to trust the hardware vendor not to do anything malicious (the side-channel vulnerability of Intel SGX has been repeatedly proven).
Fourthly, there is economic game theory , represented by Bittensor. It uses an incentive mechanism to allow nodes to compete with each other in order to approximate the correct result. The drawback is that it is probabilistic and does not prove the correctness of any single calculation.
Gensyn chose the fifth path: instead of proving whether the computation is correct, it allows everyone to reproduce and execute it on their own hardware to verify it. This path is philosophically closer to the spirit of open source, not relying on any single party's trust, and making verification capabilities open to everyone. In Ben Fielding's own words, "The digital world has new judges, and they cannot be corrupted."
After the mainnet launch, Delphi was the first target.
On April 22, 2026, the same day the mainnet launched, Gensyn's first application, Delphi, went live. This is an information marketplace that uses AI for settlement; anyone can create a marketplace, the AI model is responsible for determining the results, and REE guarantees that the settlement can be reproduced and verified by anyone. Marketplace creators receive a 1.5% commission on the transaction volume. There is no KYC requirement, and AI agents can participate without permission.
In terms of product form, it resembles Polymarket, but its target users are not limited to humans. The Gensyn team defines Delphi more deeply as an "AI reinforcement learning environment," where market prices naturally serve as the reward function. AI agents learn continuously by participating in transactions and receiving reward signals. This flywheel is currently in its early stages, but the logical chain is sound.
It's worth noting several product evolution signals since Delphi's launch. On May 7th, the official team released a lengthy article titled "Building Delphi: Pricing, Settlement, and Agentic Trading," explicitly including "agentic trading" in the roadmap. The Delphi SDK continues to be updated synchronously on GitHub, and the API for agent calls is now open. The settlement models include a mix of open-source and closed-source models such as Claude Opus 4.6 and Qwen3-32B, and market themes have expanded from sports, crypto assets, and cultural entertainment to academia (such as "the number of people achieving perfect scores in the 2026 IMO"). A public, AI-settled, long-tail problem market infrastructure is now running on the mainnet.
Why is the window now?
Returning to the initial question, when a16z invested $43 million in 2023, Gensyn's work was neither a hot topic at the time nor had a predictable implementation timeline. However, their judgments on two things became increasingly clear three years later: First, AI wouldn't remain confined to centralized labs; there would always be people wanting to run it outside, without KYC, API keys, or anyone's approval; second, that "outside" required a complete infrastructure to function, not just a single tool.
When the shortage of H100 became a global narrative, when MCP and A2A became buzzwords in the agent industry, and when a16z personally tweeted to spread the word when the mainnet went live on April 22, the market's attention finally turned to what this team had been doing for five years.
For those looking at the AI x Crypto sector, there are a few subtle but strong indicators to help determine whether Gensyn is worth tracking long-term: a16z doesn't have many projects in the decentralized AI field, and Gensyn is their biggest bet; the team's academic output is real and verifiable, not just based on white paper narratives; from deciding to go all-in in 2020 to the mainnet launch in 2026, they haven't jumped to any hotter sectors along the way.
In a market accustomed to consuming trending narratives until they're exhausted before moving on to the next one, a team that spends five years digging trenches and laying cables underground isn't necessarily the most attractive. But when everyone is scrambling to sell appliances, the power grid will eventually need to be connected, and those who are hardest to replace are often the ones connecting the grid.
