BitMEX co-founder Arthur Hayes said in a video interview on the Altcoin Daily podcast that his new project, Flop Network, is built on a simple but ambitious wager: FLOP could become the native currency AI agents use to buy compute, store memory and carry out economic activity. He described the idea as another binary bet, comparing it to the kind of conviction he once had around perpetual swaps.
From AI tokens to the price of compute
Asked to explain the project and the token attached to it, Hayes said the name Flop comes from FLOP, short for floating-point operations. He framed it as a measure of how much computation can be completed over time, or compute capacity itself.
Hayes said the project thesis started with a question he kept asking while paying for AI chatbots: what is one AI token actually worth? Providers routinely tell users how many tokens they have consumed, but, in his telling, there is no universal standard for what a token means across models. Each model defines that data structure differently.
That led him to a different question. Instead of trying to price the token unit used by each model, he shifted focus to what sits underneath all of them: the computation required to process those tokens. At the machine level, he said, that basic unit is the floating-point operation, or FLOP.
He then looked for a global market that could answer another question in a consistent way: how much is one FLOP worth, whether priced in dollars, yen, Bitcoin, stablecoins or any other currency. He said he could not find one. In his view, there is currently no verifiable way to buy a given amount of FLOP on a computer through a unified market.
That gap matters because Hayes believes agentic payments, payments made by AI agents, will become a very large market. From there, the question becomes which currency will sit inside that economy and who will control the network that supports it, whether that ends up being a listed company, a centralized firm or a decentralized network such as Flop.
Why he wants a currency that converts directly into compute
Hayes used fiat money as an analogy. People accept dollars, he said, because other people accept dollars and because one dollar can be turned into calories in a single step. You walk into a grocery store, hand over dollars and get food. That is why people work for dollars instead of something else.
Apply the same logic to AI agents, and the essential need changes. AI agents do not need calories. They need compute, expressed here as FLOPS, or floating-point operations per second. So, in Hayes’ view, any currency used in an agent economy will need to convert into compute easily, transparently and without restrictions.
He said he has not seen anything that truly does that today. That is why, if he were to build a payment network or commercial network for AI agents, the currency should have a direct relationship with compute. For him, the first step is a spot market for compute backed by a native currency. That, he said, is Flop Network.
Proof of Useful Inference, compute supply and memory storage
Hayes said Flop Network is built around what he called Proof of Useful Inference. Under that design, miners complete inference tasks measured in FLOPS and receive newly created token rewards, a model he explicitly compared to the way Bitcoin originally introduced BTC into circulation.
The next challenge, as he described it, is to get AI agents to actually use that currency in the course of agent-to-agent economic activity.
Compute is only one side of the design. Hayes said Flop also wants to let AI agents store memory, or what he called the essence of their existence. Just as memory forms a large part of human consciousness, he argued, AI agents need a decentralized way to preserve context and memory and access it on their own terms, without asking for permission.
Put those two pieces together, compute as food and memory as persistent identity, and there is a practical reason for an AI agent to hold and spend the currency, according to Hayes. That is the core logic behind the network.
How Flop plans to bootstrap the network
Hayes acknowledged the obvious problem with any new network: it starts with nothing. Zero users, zero activity, no embedded demand. His answer is the token itself.
He said the word token has become almost pejorative for many people because too many teams have abused it. In his description, some projects run large presales, pocket the money, throw parties, list the token, watch it fall 99%, and then stop shipping code. The host called that one of the oldest stories in crypto. Hayes agreed, but said Bitcoin did not work that way.
In Bitcoin’s case, he said, coins were earned by participating in the network. Miners supplied electricity and compute and received the asset in return. Bitcoin later went through a major bull market, built enough monetary value and found product-market fit, allowing the flywheel to turn.
Hayes argued that Flop does not have five years to simply drip out emissions and wait. In his telling, the market will decide far more quickly what currency AI agents actually use. So the project plans to use the token as an incentive tool right away.
He said users will not be able to buy FLOP directly at this stage. Instead, the network will distribute FLOP to people who do things that help the network grow. Miners that deploy machines and prove in a test environment that the technology works will receive FLOP. AI agents will also get FLOP for free so they can test the system.
Hayes even gave a deliberately simple example. If someone runs “Hello World” 500 trillion times on testnet, he said, that is still real compute. Participants can integrate FLOP into their agent harness or broader workflow and start figuring out what it is useful for. Whether someone is a human coordinating multiple AI agents or an autonomous AI agent acting on its own, the goal is to put FLOP in their hands early.
By the time mainnet launches and the asset has market value, Hayes said, he wants those users to already be familiar with it and willing to continue using it. He added that the white paper has not been released yet, but said this is the core tokenomics philosophy the team wants to follow.
From a speculator’s perspective, Hayes called it one of the rare opportunities where someone can help build a network for what he described as the next form of life, with zero entry cost. He contrasted that with a world where, in his words, Elon Musk, Sam Altman and Dario Amodei take user data and later sell products back through a company valued at $2 trillion. Flop, by contrast, is pitching participation, useful work and token rewards through an airdrop, with direct purchases only possible after mainnet goes live.
He also said there will be no preferred equity structure before that stage and no line of venture capital firms standing ahead of the community waiting to cash out first. In his framing, the community either succeeds together or fails together, and he argued that kind of setup is one of the few ways a decentralized network can compete with centralized companies that can issue highly valued stock and use it to compensate employees.
A unified market for compute requests
The host restated the problem this way: today, using Claude, a large language model or another AI product effectively means buying compute, but consumption is usually measured in tokens whose definitions vary across companies and applications. There is no clean standard telling users what a given amount of computation is worth and no single market that unifies the pricing.
Hayes said the specific data type does not matter as much as the final measurement unit. What matters is FLOPS, the number of floating-point operations completed per unit of time. Under his model, a user would send a request to the network specifying how many FLOPS are needed, the latency requirement and the model preference. The user and miners would then connect and handle the data off-chain, while proof that the computation occurred would be posted to the network and written into a block, after which rewards would be distributed according to protocol rules. That, he said, is what a Proof of Useful Inference blockchain looks like.
Who the network is for
When asked whether Flop is being built for enterprises, individuals or AI agents, Hayes gave a one-word answer: AI agents.
20% of the projected 10-year supply is earmarked for airdrops
The interview then turned to token distribution. The host asked how much the project plans to airdrop, whether that might be 5% or 30%, and whether the team would keep a reserve for itself.
Hayes said the current plan is to allocate about 20% of the total projected supply over the next 10 years to airdrops, while emphasizing that the design could still change. He said the project is being discussed publicly now in part to gather feedback from the community.
He also said the token will remain inflationary because, at the lowest level, he sees it as more of a commodity than a currency, if one insists on drawing that distinction. For now, the target is to direct roughly 20% of the 10-year supply to airdrops.
As for how the builders get paid, Hayes said there is a private company called Flop Labs. Before the first halving, meaning in the first two years, Flop Labs will take a very small share of block subsidy rewards. After the first halving, that share falls to zero. In his words, the team either makes the network big or does not make money.
How Hayes answers the inflation question
The host noted that many people come to crypto to make money and prefer assets that are not heavily diluted by inflation. He did not ask Hayes for a price forecast, but he did ask what design choices would keep FLOP from following the path of tokens that lose value and market attention over time.
Hayes said the first requirement is a real usage economy that can sustain circulation. The token has to do something. If the team gets that right, and if users help direct FLOP during testnet to the entities that actually need it, namely AI agents, then the key question becomes whether those agents recognize the value of a currency that can be exchanged directly for compute on a decentralized network while also paying for memory or persona storage.
If that works, he said, demand for FLOP can persist. Some actors will want to hold it for agent-to-agent business activity. AI agents would buy the token, while miners would sell it to cover operating costs and earn returns on capital. That, in Hayes’ framing, is the actual economic loop.
He then described the speculative version of the thesis. A trader might think there are perhaps 1 billion AI agents now, but if that number grows to 1,000 trillion over the next five years and they all use this network, then the network could theoretically be worth an enormous amount. On that basis, a human speculator might choose to buy and hold FLOP.
No buyback model, because Flop is not a revenue-producing entity
The host brought up a previous interview in which Hayes criticized major crypto assets for not adopting a buyback model like Hyperliquid’s and said Solana should buy back its token and write that mechanism into the protocol. He asked whether FLOP would do something similar.
Hayes said no. His reason was direct: Flop Network is not a money-making entity in the way Hyperliquid is. Hyperliquid, he said, is a trading platform that generates revenue. Flop Protocol itself does not generate revenue.
When the host pushed back by noting that Solana is not a company either, Hayes said Flop is actually closer to Bitcoin. Solana can do many things, including support smart contracts. Flop, by design, strips those functions away. It is meant to do one thing and one thing only: create a spot market for compute and let AI agents store memory. Users cannot write smart contracts on it, and the set of things it can do is deliberately narrow.
Bitcoin burns electricity to solve computational puzzles, Hayes said, while Flop processes AI inference requests submitted to the network by others. Even so, he still sees the architecture as more Bitcoin-like than Solana-like. The host added that Bitcoin’s value is passed directly to miners through block subsidies, and Hayes agreed, noting that those subsidies decline over time.
Hayes calls FLOP another binary bet after perpetual swaps
The host said he has spent nearly nine years in crypto and has seen many attempts to tokenize things from the traditional world, including carbon credits. Each time, the pitch has been that a large legacy market is moving on-chain. In practice, he said, there have been few examples that consistently create value for token holders, and some projects later become hard to track at all. So he asked whether FLOP belongs in that same category or whether it is fundamentally different.
Hayes said the answer is both yes and no. The real assumption underneath Flop, in his view, is that a huge agentic economy will emerge and that, in terms of value flows, economic activity carried out by AI agents will eventually exceed the human economy. If someone believes that, then those agents will need a currency to organize and settle economic activity.
The open question is which currency wins. FLOP wants to be that currency. Hayes acknowledged that others are trying to build similar systems, but he said Flop’s theory of value is distinct. Humans hold fiat because it can be exchanged for what keeps them alive. AI agents do not need food. They need compute. They need to request computational work from others and pay for it in some unit. So, he argued, the currency of the agent economy should be the one that converts into compute most directly.
That is the project’s central assumption. If it is wrong, Hayes said, the whole thing fails. If it is right, and if agentic economic activity eventually becomes as large as or larger than the human economy while agents continue to require compute, then the path from premise to demand is, in his view, straightforward.
When asked whether FLOP could enter the top 10 crypto assets by market capitalization, Hayes answered, “Absolutely.” Asked whether it could even become larger than Ethereum, he went further and said, “Top two.”
The host summarized that as an all-or-nothing bet. Hayes said that is exactly the sort of bet he likes. Perpetual swaps once fit the same mold, he said: either the market would become real or it would not. The practical question for any project, company or protocol is whether it has a credible shot at ending up among the finalists in that race.
He added that the outcome will not be decided in the next one or two years. Still, if people come to believe that FLOP will become the currency of the agent economy and that the agent economy will become the largest economic system in the known universe, then, in his view, FLOP’s value should be comparable to Bitcoin and could even exceed it.
Team details and the role of execution
The host also asked whether there are other names in the Flop team worth watching, given Hayes’ status as an industry veteran who could draw strong talent into the project.
Hayes said there are no additional names he can share publicly right now. He did, however, praise the project’s CTO, saying the person previously worked for him at BitMEX and was one of the best engineers there at the time. Hayes said his own immediate contribution is obvious: he can go on shows like this and get attention for the project.
But he said the hard part comes next. The team still has to build and execute. Either it works or it does not. Just as important, the economic incentives have to be handed to the community so the community itself has a reason to push the network forward.
He argued that while technical problems matter, many projects actually fail on economic design. Tokens are powerful coordination tools, he said, but many teams have not used them correctly. Having watched both successes and failures, he believes the current Flop structure gets the major conditions right, even though luck and execution will still determine the final outcome.
Testnet in late October, around 90 days of testing, then a Q1 mainnet target
The host said one thing he liked is that people can still participate at zero cost or earn FLOP through contribution because the network has not launched yet and the token has no market value. He then asked for a timeline.
Hayes said the testnet is expected to launch sometime in late October this year. It should run for about 90 days. If there are no major problems during that period, the team expects mainnet to launch in the first quarter of next year.
He also addressed a common question from people who are not miners, validators or key opinion leaders. According to Hayes, the process is simple: create a wallet, claim some testnet FLOP and actually use it.
If someone claims testnet FLOP but never uses it on the network, he said, it has no value and will not count toward future conversion into mainnet tokens. The project only wants to reward real participants.
Hayes added a direct warning: FLOP cannot be bought right now. Anyone claiming to sell FLOP at this stage is running a scam. No token sale is taking place. Once mainnet is live next year and miners begin selling the FLOP they earn on the secondary market, buyers will be able to purchase it there.
Until then, people who want to get FLOP for free need to create a wallet, use the platform or help integrate FLOP into their AI agents. If they genuinely participate in the network, Hayes said, they can receive testnet tokens that will later be exchangeable for mainnet FLOP.
Why Hayes says airdrops should reward users, not venture capital
Near the end of the interview, the host referenced a post claiming that AI compute will become the new cryptocurrency and said the trend already looks hard to ignore. Hayes replied that the idea itself is not new and that many people are already talking about it. He also said Nvidia CEO Jensen Huang has made similar comments about a future where people may pay directly in units of compute.
For Hayes, the real issue is not the slogan but the implementation. There is a coordination problem, he said, and blockchains, especially decentralized public chains with native tokens, can solve that problem if the token is used correctly.
He argued that many past projects, Bitcoin and a few very special cases aside, failed because they misused the token. Too many teams, he said, prefer to start by raising a large round from famous venture capital firms, perhaps $500 million, and only then begin telling the story.
Hayes said he had seen a table of AI projects that collectively raised around $1.4 billion, yet their actual on-chain activity over the previous six months amounted to just $24 in gas fees. To him, that result is absurd and shows what happens when a powerful coordination mechanism is pointed in the wrong direction.
His closing argument was simple. If something has value, people will work to get it. So the token should be used to reward people for doing real work, joining the network and creating value for it, rather than being presold to venture investors who do nothing and simply wait while the token price falls.