Arthur Hayes wants to turn AI compute into a tradable asset, and he is dangling a chunky carrot: roughly 20% of the FLOP supply for testnet participants over 10 years. That is either a smart way to bootstrap a real machine economy, or another crypto attempt to slap a token on top of a brutally hard engineering problem.
- 20% for testnet users: Hayes says FLOP will be distributed over 10 years.
- Compute as the product: FLOP is meant to price AI work by floating-point operations, not vague model tokens.
- Still a lot of blanks: Supply, verification, validators, and the launch chain remain unclear.
In an Aug. 19 Substack post titled “The Book of Genesis, ” the BitMEX co-founder laid out Flop Network as a self-funded marketplace for machine intelligence. The idea is simple enough to explain and nasty enough to build: miners would provide compute, AI agents would pay in FLOP, and the network would supposedly turn raw processing power into a marketable good.
Hayes says FLOP is “a claim on compute” and, in a more colorful line, “food for AI agents.” Fair enough. Compute is getting more valuable by the day. The real question is whether a token can actually coordinate that value without turning into a speculative sideshow with extra wattage.
There is a real thesis hiding under the hype. AI systems are already acting like economic agents. They call APIs, pay for services, store memory, and perform tasks with limited human supervision. If agents need to keep buying compute, storage, and related services, a machine-native payment layer makes sense. Crypto should be good at this. That is the pitch, at least when everyone is still sober.
Hayes said the launch would follow a fair-start model and that he funded development himself, so no presale is needed. He also said roughly 20% of the total FLOP supply would go to testnet participants over 10 years. But the important details were left hanging: the total supply, the release schedule, and the specific actions that determine each participant’s share.
That matters. “Fair-start” is not magic. A token can still be unfair, opaque, or economically busted if the emission schedule is fuzzy or the early distribution gets gamed. Crypto has seen plenty of projects wrap messy tokenomics in a clean little label and call it decentralization. That trick is older than most people’s alt accounts.
The more ambitious part is the compute market itself. Flop Network would price AI workloads based on floating-point operations, or FLOPs, the actual number of arithmetic operations required to do the work. A single FLOP is one floating-point operation. FLOPs is the rate or amount of that work. In plain English, Hayes wants to price compute by what it actually does, instead of by the often opaque pricing of specific AI models or token schemes.
Miners would be paid with FLOP block rewards and inference fees under a mechanism Hayes calls Proof of Useful Inference, or PoUI. That name sounds impressive. It is also just a name until the network can prove the work is real.
And that is the ugly part. Bitcoin mining is easy to verify. The hash meets the target, or it does not. AI inference is messier. Outputs can be nondeterministic, which means the same prompt can produce slightly different answers. So a decentralized network would need to prove three different things at once: that the right model ran, that the right amount of work was performed, and that the output was valid.
The materials tied to the project did not explain how validators would do that. Earlier reporting also said Flop Labs had not published a way to check nondeterministic outputs, detect incorrect work, or penalize false results. That is not a small gap. Without reliable verification, PoUI is less a protocol than a slogan with power bills.
Other basics were still missing too. Public documentation had not established which blockchain will support FLOP, how many validators will operate at launch, or whether consumer hardware can realistically compete with data-center equipment. If the answer is “yes, but trust us, ” that is not a technical roadmap. That is vibes with a cap table.
The timing adds another practical wrinkle. The Aug. 18 announcement placed a large FLOP airdrop in Q4 2026 and the Flop Network genesis block in Q1 2027. If distribution begins before the native network is live, where exactly do recipients hold the tokens? That question was not answered in the available material, and it is not a trivial one.
To be fair, the broader market does show real demand for machine-to-machine payments. A May 2026 Keyrock report found AI agents settled $73 million through 176 million transactions over 12 months, with USDC accounting for 98.6% of tracked payments. Keyrock also found that 76% of those transactions were worth less than $0.30.
That matters for two reasons. First, it proves the market for agent payments is not theoretical. Second, it shows why stablecoins are already doing a lot of the heavy lifting. Layer 2 stablecoin transfers, Keyrock said, can cost about $0.0001. For tiny machine payments, that is hard to beat without getting very clever very fast.
There is also a broader stack forming around that same need. Coinbase began allowing commercial customers to accept agent USDC payments through its x402 standard in July. In June, the American Arbitration Association and Integra Ledger introduced legal records for agent transactions. In other words, the rails are already being built, and they are mostly stablecoin rails, not bespoke utility-token fantasies.
Keyrock’s framing is useful here: The Protocols Are Not Competing; They Are Stacking. That means Flop Network is not walking into an empty field. It is entering a market where payment standards already exist, USDC already dominates, and large players are racing to own the plumbing.
That does not make FLOP pointless. It means the project has to prove it solves something that USDC rails and existing standards do not. If Flop Network can genuinely create a compute-native marketplace with fair pricing, verified inference, and low-friction settlement, there could be real demand. If not, it risks becoming another token trying to dress up a hard engineering problem as inevitability.
“A claim on compute” is Hayes’ description of FLOP. That is the bullish case in one line: compute has value, so a token tied to compute should have value too.
The bearish case is simpler. The real world prices compute in dollars because electricity, hardware, and payroll are priced in dollars. If FLOP swings around like a caffeinated raccoon, miners and users will still need a stable unit underneath it. That is a big reason stablecoins are already winning machine commerce: they are boring, legible, and they do not turn every invoice into a coin flip.
Hayes said his next piece would explain why the agent economy needs a spot market priced by floating-point operations per unit of time. He may be right that compute needs a better market structure. But no token, no matter how spicy the branding, can brute-force away the core problem: proving useful work on an open network without trusting the middleman you were trying to escape in the first place.
Key questions and answers
-
What is FLOP supposed to do?
FLOP is the native token of Flop Network. Hayes says it will be used to pay miners and let AI agents buy compute, while also functioning as a claim on compute itself. -
How much of the supply goes to testnet participants?
Hayes said roughly 20% would be distributed to testnet participants over 10 years. He did not disclose the total supply, the release rate, or the exact rules for earning that allocation. -
Why is verification such a big problem?
Because AI inference is not as easy to prove as Bitcoin mining. The network would need to verify the model used, the work performed, and the validity of the output, and the available material does not explain how that will work. -
Are AI agents already paying for things today?
Yes. Keyrock reported $73 million settled across 176 million transactions over 12 months, with 98.6% of tracked payments in USDC. So the use case is real, even if the best rail is still up for debate. -
Does FLOP solve a problem stablecoins do not?
Maybe, but that is not proven yet. Stablecoins already handle cheap machine payments well. FLOP would need to show it can also price and verify compute in a way stablecoins cannot.
Flop Network is a serious bet on a future where AI agents are not just users of the internet, but participants in a market for compute itself. That future is plausible. Whether FLOP is the right instrument for it is still wide open.
Further reading
A few related items worth keeping on the radar:
- Arthur Hayes proposes 20% FLOP testnet allocation
- Understanding Yahoo's Consent Page and Its Implications
- Error extracting content
- Arthur Hayes "Comes Out of Retirement": This Time, He's
- Arthur Hayes Says CLARITY Act Should Be Vetoed as May 21
- Bitcoin Rally Possible as US-Iran Tensions and Fed Policy
- Arthur Hayes Predicts Bitcoin Surge to $100K with $572B
