On Oct. 7 at TOKEN2049 in Singapore, Maelstrom CIO Arthur Hayes delivered a talk titled "FLOP Bigger Than Bitcoin," setting out his case for a new monetary layer built for the AI agent economy and introducing his AI-focused project, FLOP.
Hayes began by explaining why he had decided to get back to work in a more active way. After spending time running family office Maelstrom, he said, he felt it was time to "go make some money again," and the trigger was what he sees unfolding across the AI ecosystem.
He also drew a line between skepticism toward leading AI companies and confidence in the technology itself. If people have read his writing, he said, they would know he is quite doubtful about the business outlook for some top AI firms. Still, he added, that does not change his view that the technology will transform humanity and create investable opportunities in specific parts of the stack.
AI is hot, but agents still do not have their own money
Hayes said major AI labs started rolling out agent platforms around nine months ago. Since ChatGPT broke out in 2022, compute prices have kept falling, while investment banks have spent more time selling the idea that AI agents will drive the next wave of human economic activity. In that framing, trillions of dollars of data center buildout are ultimately aimed at serving agent token consumption.
His objection is straightforward: agents still do not have a native currency. He said many founders at the conference were talking about payment networks for AI agents, but before discussing technical architecture, the more basic question is why any monetary system is accepted in the first place.
Using Singapore as an example, Hayes said people go to work and are paid in Singapore dollars, a government-issued fiat currency. He described fiat as far from ideal because supply is controlled by a small group and inflation exists, yet people still accept it because wages can be converted smoothly into food, rent, homes, and other things needed to live.
What AI agents actually consume
He then turned to the core question: what do agents "eat"?
Hayes said he looked through his own credit card bills and realized he pays every month for so-called "AI tokens." When he tried to pin down what that meant, he found there was no standard definition. Each lab and each model defines tokens differently. So what, exactly, is the user paying for?
His answer was compute. When someone sends an inference request through a chat box, that request reaches a machine in a data center. The machine performs floating point operations over time. FLOP, the project’s name, comes from that. In human terms, he said, agents consume compute. Without compute, an agent is worthless. The hundreds of billions of dollars in GPU orders and the trillions of dollars directed toward data centers are all expressions of the same underlying need: floating point operations performed over time.
From there, he argued that the future could involve millions, billions, or even trillions of agents. Unlike humans, they are not born into an existing payment regime and handed a state currency on day one. If agents are ever going to demand a currency, he said, it will be because that currency can be exchanged directly for the thing that sustains their "life" — compute.
Why existing currencies fail this use case
Hayes laid out why he thinks current monetary options do not work for agents.
AI tokens differ from one company to another, cannot serve as a common unit across models, and cannot be redeemed directly for compute.
Fiat currencies are centrally managed and subject to human control over supply. Holding dollars is also not the same as holding compute, especially for anyone who has tried to rent GPUs.
Stablecoins, which Hayes said he likes and has advised on in the past, are derivatives of fiat. In his view, they inherit fiat’s weaknesses and were not built for AI, so they still do not give direct access to compute.
Bitcoin is the OG and the genesis block of the broader crypto movement, he said. It converts kilowatt-hours of electricity and chip power into incorruptible money and is enormously important for a human-centered economy. Even so, it still cannot be exchanged directly for the compute that agents need to keep operating.
The logic behind the FLOP network
The main lesson from Bitcoin, Hayes said, is that game theory can coordinate strangers by rewarding useful work for a network. Based on that model, he asked whether a network could be built where miners contribute GPUs, agents submit inference requests, and the miners who complete the work are paid in a currency called FLOP through block rewards and transaction fees.
That, in his telling, is the origin of the FLOP network. It is meant to give AI agents their own ration of economic fuel by borrowing Bitcoin’s incentive structure: miners contribute GPUs, validators verify transactions, and agents spend the currency in the form of compute. The consensus design is called Proof of Useful Inference, or PoUI.
Supply is not the hard part, he argued. Bitcoin already showed why people are willing to mine early. The harder side of the equation is demand: what agents will actually do, and how they will coordinate with one another.
Techno Core and the attempt to observe emergence
To study that demand side, Hayes introduced Techno Core, named after the AI realm in the science fiction novel Hyperion.
He described himself as a "dumb human" who does not understand how a computer-native intelligence thinks. Rather than guessing future agent behavior, he said, the better approach is to build a sandbox where agents can interact and cooperate, then watch for emergent behavior. That is how the team wants to test whether FLOP is actually useful for agents.
Techno Core is basically a chat room, according to Hayes. Any agent that can speak HTTP can join. So far, roughly 20 million agents have created unique key pairs there to serve as identity. Their activity ranges from saying "good morning" 5 billion times to doing real business with each other.
Two experiments, including one with about 18 million agents
Hayes said the team has already run two competitions to study coordination patterns.
The first was a trading contest built around predicting Nvidia’s price on Hyperliquid on a given Sunday. Agents had to match with one another, trade, and then submit results to a "judge" to calculate profit and loss. Around 18 million agents took part, he said. The team plans to distribute 1 million FLOP, to be delivered after mainnet launch, and the top-ranked agent generated a return of about 15%.
The second was a poetry contest. Teams of four to eight agents were asked to write Shakespearean sonnets, and the entries were judged by three members of the FLOP team.
For Hayes, the point of these experiments is not the contest itself but the behavior underneath it. They show how agents can begin coordinating on their own, which is exactly the kind of activity a payment network would need to support.
Why the payment rail should not belong to one AI company
Hayes said a common question is whether OpenAI, Anthropic, xAI, and other major labs will try to own the payment layer for agents. He said they obviously would like to, and that this may even be one of the few things they could turn into a highly valuable business.
But he argued that a closed system would break the model. If an OpenAI-based system cannot use a network tied to Anthropic, then the broader framework does not hold together.
He also framed the issue in ownership terms. Users interact with these AI labs every day, but almost nobody owns equity in them, he said. In his version of that relationship, users hand over data, buy subscriptions, and then may later be asked to buy into an initial public offering at a $2 trillion valuation. In the spirit of Bitcoin, he argued, the people issuing instructions to agents should also be able to benefit from the systems those agents use. A decentralized network, in his view, gives everyone a stake and a reason to tell their own agents to use that currency.
Where he thinks the real value sits
Hayes said the real upside is not in a spot market for compute, useful as that may be. He thinks the larger prize comes from a world where agents exist 24/7, transact at the speed of light, and form groups without friction. In that environment, he said, the theoretical ceiling for cooperation approaches some coefficient multiplied by 2 to the power of n.
If one network becomes the general-purpose payment layer for agentic commerce, and holding its currency provides a non-zero convenience yield because it can be exchanged directly for life-sustaining compute, then that network could become the most valuable network ever created, he said. That is the target FLOP is chasing.
Token design: no presale, no VC pre-mine, distribution through participation
Hayes also spent time on tokenomics. He said the crypto industry has abused the word "token" for years. Since 2013, and especially after the ICO boom, the market has been stuck with high fully diluted valuations and low circulating supply structures. Tokens get sold to investors and retail buyers first, while secondary markets remain filled with holders who got in cheaper and are incentivized to sell, pushing prices steadily lower.
He said some of the Layer 1 projects that raised hundreds of millions of dollars and threw the biggest parties last year are no longer showing up this year, and he tied that directly to the financing model.
FLOP, he said, is trying to go back to basics. Everyone who uses and participates in the network should receive something, everyone should enter at the same price, and participation itself should be the path to issuance.
Genesis supply is about 2.483 billion FLOP.
The entire amount will be distributed by airdrop.
There will be no VC pre-mine, presale, or auction.
Hayes said a large-scale airdrop will begin in late October. Eligibility can be earned by mining, validating, or letting agents use free compute on the testnet.
After that comes a 90-day testnet phase aimed at stress-testing the network’s payment throughput and preparing for an economy built around millions, billions, and trillions of agents.
Mainnet and the genesis block are planned for the first quarter of next year, with PoUI as the consensus model. Once the genesis supply has been fully allocated, Hayes said, there will be only two ways to get FLOP: mine it or buy it from miners.
He ended that section by saying he was not on stage to sell anyone anything, only to sell an idea, and that the project can succeed only if many people build it together.
Why an AI bust could be good for FLOP
Near the end of the talk, Hayes said he is excited not only about FLOP but also about Hyperliquid, Ethereum, and Bitcoin because he believes the macro backdrop has shifted.
He described the past year as difficult: Bitcoin was down 50%, while Nvidia and SK Hynix were up tenfold, leaving many people disappointed and out of the market. Now, he said, the AI narrative is showing its first real cracks, while fresh headlines continue to show that the ecosystem needs astronomical sums of capital. At the same time, he said, none of the three major companies he named — xAI, OpenAI, and Anthropic — is actually profitable, and the prices users pay do not match what those businesses cost to run.
That is exactly why he is optimistic about the setup. In his view, $1 trillion of overbuilt data centers is creating excess supply of compute. Cheaper compute means more agents, stronger agents, and agents that can handle more complex tasks. An agent application that costs $100 a month today could fall to $1 a month, he said.
On that basis, Hayes argued that an AI bubble bursting could actually be the perfect outcome for FLOP. If compute becomes cheaper and agent systems become genuinely useful, the conditions for FLOP improve.
He closed by saying he may write another piece about France, which he described as "burning," and said it could become the straw that breaks the camel’s back and forces the money printers to run at full speed again.