I still remember the first time someone tried to explain why AI agents would need their own money. It sounded half science-fiction, half accounting problem. Fast forward to this week and Arthur Hayes has dropped a detailed sketch of exactly that idea. In a Substack post dated August 19 he lays out Flop Network, a marketplace where AI agents buy computing power and ordinary machines get paid for useful inference work. The part that made me stop scrolling was the token plan: roughly 20 percent of the entire FLOP supply handed out to testnet participants across ten years. No big private sale. No venture-heavy unlock schedule. Just a long, slow drip to the people actually using the network early.
Why a Self-Funded Fair Launch Matters Right Now
Most new tokens still follow the same tired script. Raise money from a handful of funds, lock a chunk for the team, another chunk for advisors, then open the rest to the public at a price that already prices in years of growth. Hayes is deliberately walking away from that model. He says he paid for the development team out of his own pocket so there was no need for a presale. That single decision changes the starting conditions. When the token finally trades, retail buyers are not staring at a massive overhang of locked investor tokens waiting to unlock every quarter.
I’ve watched enough cycles to know that “fair launch” can become marketing language rather than reality. Still, the intention here feels different. By tying a full fifth of the supply to testnet activity stretched over a decade, the project is trying to reward sustained contribution instead of a short farming sprint. Whether that works depends on how the points or work units are measured, but the structure itself is worth examining.
What Exactly Is Flop Network Trying to Build
At its core the network wants to create a spot market for computing work measured in floating-point operations. Customers, whether human or autonomous agent, submit a job that specifies the model, the amount of work, and the time window. Miners run the requested model on their hardware and receive FLOP tokens as payment. The idea is that FLOPs become a common unit of account the way kilowatt-hours work for electricity.
Today every major AI provider bills in its own token units. Comparing the real cost of running the same prompt across two different services is almost impossible. Cloud providers rent capacity by the hour or by the instance, not by the actual arithmetic performed. Hayes argues that a market priced directly in FLOPs would let buyers and sellers discover a true spot price for compute, independent of model branding or data-center location.
The currency would represent a claim on compute, allowing buyers and sellers to establish a consistent price for a given amount of processing work.
That sentence sits at the center of the whole proposal. If it works, FLOP becomes something closer to fuel for software agents rather than another speculative token.
Proof of Useful Inference Instead of Pure Hashing
Bitcoin miners burn electricity to produce hashes that secure the ledger. Flop Network wants miners to burn electricity doing something customers actually need: running inference. The mechanism is called Proof of Useful Inference, or PoUI. Miners earn two streams of tokens. Block rewards for keeping the network healthy and direct fees for completing customer jobs.
The hard part is verification. AI outputs are nondeterministic. The same model with the same seed can still produce slightly different results. How do you prove a miner used the exact model requested, performed the claimed number of operations, and returned a valid answer? The public materials so far do not spell out the verification layer. That gap is the biggest open technical question. Without a robust way to catch bad work or incomplete work, the economic model collapses.
I’ve seen projects promise “useful work” before and then quietly fall back to simpler puzzles once the verification cost became clear. Flop Labs will need to publish the concrete method soon if the testnet is going to attract serious hardware operators.
The 20 Percent Testnet Allocation Over Ten Years
Here is the part that stands out. Participants who contribute to the testnet are collectively promised about one-fifth of the total FLOP supply, released gradually across a full decade. No exact total supply figure has been given, nor the precise emission curve, nor the scoring system that decides each person’s share. Still, the headline number is large enough to matter.
Think about what that means in practice. Someone who starts providing reliable compute or validation work in the early testnet could keep receiving tokens for years. The project is betting that long-term alignment beats the short-term frenzy of a three-month airdrop farm. Whether the community sees it that way will depend on transparency around the rules.
- No presale or large private investor allocation announced
- Development costs covered by Hayes personally
- Roughly 20 percent of supply earmarked for testnet contributors
- Release window stretched over ten years
- Eligibility criteria still unpublished
Those five points form the current public picture. Everything else remains to be filled in.
How the Token Is Supposed to Circulate
Agents need two things continuously: processing power and memory. Hayes suggests both can be paid for with FLOP. An agent spends tokens to run inference and to store its ongoing context on decentralized storage. Miners and storage providers earn those same tokens. The loop is simple on paper. Agents pay for work, workers receive the medium of exchange, and the token’s value is tied to the demand for real compute.
In practice the loop has to survive the usual problems. Hardware and electricity are priced in dollars. If the FLOP market price swings hard, miners may switch off when the math no longer works. The proposal does not yet explain how the network would keep the real cost of compute reasonably stable under volatile token prices. That is another open design question.
Machine Payments Already Exist and Prefer Stablecoins
The broader market for agent-to-agent payments is already moving. Reports from earlier this year showed AI agents settling tens of millions of dollars through hundreds of millions of tiny transactions, overwhelmingly in dollar-backed stablecoins. The average size of those payments sits far below traditional card fees, which is exactly why stablecoins on low-cost networks make sense for automated commerce.
FLOP is attempting something different. Instead of pricing everything in dollars, it wants to price the underlying work itself. An agent that needs a fixed number of floating-point operations would buy that work directly rather than converting everything through a stablecoin first. The philosophical difference is interesting. Whether the market prefers a specialized compute token or continues to settle in dollars remains an empirical question.
In my view the two models can coexist. Stablecoins will likely remain the settlement layer for most commercial agent activity, while a specialized token could still find a niche if it genuinely improves price discovery for raw compute.
Timeline and What Is Still Missing
Public statements place a large FLOP airdrop in the fourth quarter of 2026 and the genesis block in the first quarter of 2027. That leaves a long runway for testnet activity before the main network goes live. It also means the first token distributions may occur before the native chain exists. How those early tokens will be held and later migrated has not been detailed.
Other missing pieces include the target blockchain, the number of validators expected at launch, the precise consensus rules, and whether ordinary consumer GPUs can compete with specialized data-center hardware. Without those details it is hard to judge the long-term decentralization claims.
The Economic Logic Behind a Long Testnet Reward
Most networks face a cold-start problem. They need both supply of service and demand for that service at the same moment. Tokens are often used as a coordination tool to attract both sides early. By spreading the testnet allocation over ten years, Flop Network is trying to keep incentives alive long after the initial excitement fades. The risk is that the emission becomes too slow to motivate serious operators, or that the rules favor large farms that can game the scoring system.
I’ve found that the projects that succeed with long reward schedules usually publish clear, auditable metrics early. Points for uptime, points for successful job completion, points for correct verification, and so on. Ambiguity tends to produce cynicism. Right now the scoring method is still a black box.
Potential Strengths of the Approach
There are several things to like on paper. Self-funding removes the usual pressure to generate returns for early private investors. A long distribution window encourages genuine participation rather than pure farming. Pricing compute in a standard unit of work could, if successful, make cost comparison across providers far easier than it is today. And the explicit focus on AI agents as primary customers matches a real emerging need.
Perhaps the most interesting aspect is the attempt to treat compute as a commodity with a spot market. Energy markets already do something similar. Compute markets have so far remained more opaque. Closing that gap would be valuable even if the specific token never becomes widely used.
Open Risks That Still Need Answers
Verification of nondeterministic AI work remains the largest technical risk. Without a reliable and cheap way to check results, the network cannot punish bad actors effectively. Second, the mismatch between dollar-denominated operating costs and a floating token price can create boom-and-bust cycles for miners. Third, the absence of a published white paper, token contract, or security audit leaves too many unknowns for cautious capital.
There is also the question of regulatory treatment. A token that functions as both a payment medium for compute and a reward for network participation may attract different scrutiny depending on jurisdiction. The project has not yet addressed that dimension publicly.
How This Fits Into the Broader Machine Economy
Software agents are already buying data, API calls, and small digital services. Most of those transactions settle in stablecoins because the unit of account is familiar and the price is stable. A specialized compute token has to offer a clear advantage to displace that habit. The advantage Hayes is selling is better price discovery and a native medium that agents can hold as a claim on future work.
Whether agents will actually prefer to hold a volatile compute token rather than simply buy the work they need at the moment of need is still unknown. Markets will decide. What is clear is that the conversation has moved past pure speculation about AI agents and into concrete proposals for the economic rails they might use.
Practical Questions for Potential Participants
Anyone considering joining the testnet should watch for several concrete announcements. First, the exact eligibility rules and scoring formula. Second, the hardware requirements and whether consumer machines stand a realistic chance. Third, the method used to verify inference results. Fourth, the token migration path if early distributions happen before the main network launches. Until those points are clear, participation carries more uncertainty than usual.
I’ve seen too many testnets where the rules changed midway or the final allocation disappointed early workers. Clear, immutable criteria published early are the best protection against that outcome.
A Longer View on Token Design
The decision to avoid a large presale is refreshing. Too many projects still treat early private capital as the default path. Self-funding is only possible for teams that already have resources, of course, but when it is available it removes one common source of misaligned incentives. Spreading the community allocation over a decade is an equally deliberate choice. It trades short-term hype for the possibility of sustained contribution.
Whether the market rewards that patience remains to be seen. Token markets often prefer rapid narratives. A ten-year drip may feel too slow for speculative attention. On the other hand, if the network actually delivers usable compute at competitive prices, the slow allocation could become a feature rather than a bug.
What Comes Next From the Team
Hayes has already signaled that a follow-up post will dig deeper into why a floating-point spot market is necessary for the agent economy and how Flop Network intends to create it. That second piece should fill in some of the economic gaps left open in the first article. Technical documentation, testnet instructions, and a clearer tokenomics table would also move the project from concept to something people can evaluate more rigorously.
Until those materials appear, the proposal remains interesting but incomplete. The core idea of pricing AI work in a standard unit of computation is sound. The long testnet allocation is an ambitious attempt to align incentives. The verification problem and the dollar-cost mismatch are the two issues that will determine whether the experiment succeeds or quietly fades.
Final Thoughts on the Proposal
Arthur Hayes has a track record of writing provocative, well-argued pieces that force the industry to examine its assumptions. This latest one is no exception. The notion that AI agents need a native claim on compute, and that ordinary machines can earn that claim by doing useful work, is both elegant and ambitious. The 20 percent testnet allocation stretched across a decade is a concrete mechanism designed to support that vision without the usual early-investor overhang.
I remain cautiously optimistic but far from convinced. The missing verification layer is not a small detail. The practical economics of mining under a volatile token price also need clearer solutions. Still, the direction of travel feels right. As software agents become more autonomous, the question of how they pay for the resources they consume will only grow more important. Flop Network is one of the more thoughtful attempts so far to answer that question with a purpose-built market rather than another wrapper around existing cloud billing.
Whether the final product matches the ambition of the initial sketch will depend on the hard engineering and the transparent rules that still need to be published. For now the conversation itself is valuable. Compute is becoming a foundational commodity for digital agents. Finding better ways to price and settle that commodity is a problem worth solving, and this proposal at least faces the problem head-on.
The next few months of documentation and testnet activity will tell us whether the idea can move from Substack essay to working network. Until then, the 20 percent figure remains the most concrete signal of how the team intends to share the upside with the people who help build the early system. That alone makes the project worth watching closely.
In a market that still rewards short-term farming more often than long-term contribution, the willingness to lock a large portion of supply into a decade-long testnet reward is unusual. It may prove naïve. It may also prove to be one of the smarter structural choices a new network has made in recent years. Time, and the quality of the verification system, will decide which of those outcomes arrives first.
For anyone following the intersection of crypto and autonomous software, the Flop Network proposal offers a useful case study. It shows both the promise of purpose-built economic rails for AI agents and the practical obstacles that still stand in the way. The 20 percent allocation is the headline number. The harder work of making Proof of Useful Inference actually work is the real test still ahead.
