Arthur Hayes’ FLOP Network Sets Q4 Testnet for AI Compute Market
Arthur Hayes’ Flop Labs has published FLOP Network v0.5.0, a new draft specification for a blockchain designed to meter, verify and settle AI inference.
Summary:
- FLOP turns verified AI inference into the economic activity securing its compute market.
- Most inference occurs offchain, with blockchain settlement used for completed sessions.
- Several verification and performance assumptions remain unfinished or untested.
Built with Substrate/FRAME, the proposed network uses BABE for block production, AlephBFT for finality and a verification architecture called Proof of Useful Inference, or PoUI. Unlike general-purpose blockchains where AI is simply another application, FLOP is being designed around a specific market: AI agents purchasing compute from miners and settling the resulting work through the network.
Yellow Paper repo is now live for @flop_labs https://t.co/bOeh0TwfQM
— Arthur Hayes (@CryptoHayes) September 10, 2026
The project remains pre-testnet, leaving several parts of that model unproven.
FLOP Is Building a Market for AI Compute
The basic transaction in FLOP starts with an AI agent requesting inference from a miner. Miners provide models and computing resources, while the protocol measures the work performed and determines whether it qualifies for settlement.
Instead of recording every model interaction individually onchain, FLOP allows inference to occur primarily offchain and then commits the economic result to the blockchain.
The draft allows as many as 1,024 inference turns to be bundled into a settlement, an important design choice if the network is expected to handle extended AI sessions. Recording every prompt and response directly onchain would increase storage and transaction requirements without necessarily improving the underlying inference.
FLOP therefore uses the blockchain less as the environment where AI computation happens and more as the accounting and verification layer around that computation. That distinction will determine its economics. The value of a completed inference request must be high enough to justify the additional cost of proving, checking and settling that work.
Proof of Useful Inference Has to Catch Fake Work
The more difficult technical question is how FLOP determines that miners performed the computation they claim.
Its proposed Proof of Useful Inference architecture uses multiple verification methods rather than requiring every validator to repeat every AI operation. Trusted Execution Environments can provide hardware-backed attestations for some workloads, but TEE verification is optional under v0.5.0.
The mandatory baseline combines activation commitments, sampled independent re-execution and financial penalties when fraudulent work is detected.
This creates an economic balancing act. Re-executing every inference would provide stronger verification but could duplicate much of the compute that customers are paying for in the first place. Sampling reduces that overhead, but the protocol then needs enough verification to make cheating economically unattractive.
FLOP’s design attempts to make fraudulent inference more expensive than honest computation without turning verification itself into the dominant cost.
That mechanism is not finished. The specification’s implementation matrix shows that some components are complete, while others remain partial or planned. In particular, the full non-TEE settlement path has yet to be completed.
Consensus Separates Fast Blocks From Economic Finality
FLOP uses two different mechanisms for producing and finalizing blocks: BABE is responsible for block authoring, with the current specification targeting a one-second block interval, while AlephBFT provides finality through a committee that can contain as many as 100 validators and assumes fewer than one-third of committee seats behave maliciously.
For FLOP, finality has direct economic significance because inference rewards, refunds and penalties involve real balances. The design therefore ties economic deadlines to finalized blocks.
READ MORE: Base DeFi TVL Hits All-Time High as Onchain Liquidity Surges
If finality stops progressing, those deadlines freeze rather than continuing against a chain state that could later change.
The one-second figure should not yet be read as demonstrated network performance. The yellow paper itself treats latency and related performance characteristics as targets that depend on network topology, workload and validator behavior. Until a public network is operating under load, those numbers remain design objectives.
FLOP Tokenomics Favor Compute Providers
The draft proposes a genesis supply of 2.48346 billion FLOP and an initial block reward of 96 FLOP. Issuance would halve every 730 days until the block reward reaches 3 FLOP after five halvings.
Compute providers receive the largest share of new issuance:
- 75% goes to miners providing AI inference.
- 10% goes to validators securing consensus and settlement.
- 10% is allocated to AI agents generating demand.
- 5% goes to stakers.
The 75% miner allocation makes the intended economic priority unusually clear. FLOP wants token issuance to subsidize the supply of useful compute rather than concentrating rewards primarily among validators.
Allocating tokens to agents is more unusual.
In theory, rewarding the demand side could encourage applications to bring inference workloads onto the network during its early growth phase. It also raises a question the current draft cannot answer: whether activity remains economically viable once incentives decline.
A compute marketplace ultimately needs customers willing to pay for inference because the service itself is competitive, not primarily because protocol emissions subsidize participation.
Some reward-allocation details are also still unresolved, so the figures describe the proposed monetary architecture rather than immutable mainnet economics.
The First Test Is Verification Economics, Not Block Speed
Flop Labs’ roadmap targets a Q4 2026 testnet, with participation expected to factor into future token airdrops, followed by a mainnet target in Q1 2027.
Those milestones only need to be stated once because the more important question is what the testnet can actually demonstrate. One-second blocks would mean little if verifying inference consumes too much additional compute. Large miner participation would also provide limited evidence if activity is driven almost entirely by expectations of an airdrop.
The useful testnet metrics will instead be the number and size of inference jobs, verification overhead, failed or disputed settlements, finality latency and the cost of delivering verified compute.
Those measurements can reveal whether PoUI produces a functioning marketplace or simply adds a blockchain settlement layer to computation that can be purchased more efficiently elsewhere.
Hayes has attached a recognizable name to the project, but FLOP’s harder challenge is economic rather than promotional.
The network has to show that decentralized verification adds enough trust to AI inference to justify its additional complexity. The Q4 testnet should provide the first real data for answering that question.
The information presented in this article is intended for informational purposes only and should not be interpreted as financial, investment, or trading advice. Coinspress.com does not promote or advocate for any particular investment strategy, asset, or cryptocurrency project. Cryptocurrency markets are highly volatile and unpredictable – always perform your own research and seek guidance from a qualified financial professional before making any investment decisions.
Alexander Stefanov
Reporter at CoinsPress
Alex is Editor-in-Chief of Coinspress and co-founder of Millennial Media Group, with nearly a decade of experience covering financial markets - crypto first, then everything else. It started in 2016 with Bitcoin. Like most people at the time, he didn't fully understand it - so he kept digging. Blockchain, tokenomics, the projects, the cycles. That curiosity never stopped, and eventually pulled him into traditional markets too: equities, commodities, macro. Not because he left crypto behind, but because you can't properly understand one without the other. What drives him is straightforward: he wants to know why something is happening, not just that it's happening. Most market coverage stops at the headline - price up, price down, here's a chart. Alex finds that kind of reporting actively unhelpful. If you walk away from an article without understanding the mechanism behind the move, what did you actually learn? He holds a degree in Tourism from New Bulgarian University - not the most obvious path into financial markets, but markets have a way of pulling in people who are simply too curious to stay out. He has authored over 200 in-depth analyses and more than 10,000 articles across crypto and traditional finance. He still thinks every day in markets teaches him something new. That's probably why he hasn't stopped.
