Arthur Hayes' FLOP Token Spec Sets 2.48 Billion Genesis Supply for AI Agents

Proof of Useful Inference, Explained

Arthur Hayes has published the full technical specification for the FLOP Network, a blockchain built around a consensus model called proof of useful inference (PoUI) in which autonomous AI agents pay miners directly in the native token for model inference. According to the project documentation Hayes shared over the weekend, the genesis supply is set at approximately 2.48346 billion FLOP, with the entire initial allocation designated for airdrops — no venture-capital premine and no token auction. FLOP, short for floating-point operations, functions as the payment rail: an agent spends the token whenever it needs computing resources, a miner with suitable hardware accepts the session request from the mempool, runs the requested inference over a private connection, and returns a proof whose hash validators settle into a block. Each request specifies a model-weight hash, maximum latency, compute measured in FLOPs, a confidentiality setting and the offered fee.

The reward schedule is explicit in the specs. Miners take 75% of block rewards, validators and agents 10% each, and regular stakers the remaining 5%. Issuance starts at 96 FLOP per block and halves every 730 days — 48, 24, 12, 6 — before settling at a permanent 3 FLOP after the fifth reduction, meaning emissions never reach zero, unlike Bitcoin's fixed-supply issuance. Average block time is one second with deterministic sub-second finality. Participation requires staking: miners who misrepresent completed inference and validators who publish dishonest blocks face slashing, up to full forfeiture and removal. The validator set is capped at 1,000, with roughly 50 rotating monthly based on verified workload and uptime, and FLOP Improvement Proposals need two-thirds approval from the active set. Hayes, the BitMEX co-founder, said on Aug. 18 he was returning to an operating role at Flop Labs, calling FLOP “food for your AI agent.” The documentation remains a draft, last updated Aug. 27.

Fair Launch, Airdrops and Anti-Velocity Design

The second layer of the design addresses why an AI agent would not simply pay in USDC or USDT. The paper's argument is that dollar stablecoins are on-chain versions of human money, while an agent's baseline survival resource is compute — so a currency denominated in floating-point operations maps directly onto what agents actually buy. To counter token-velocity decay, FLOP plays three roles at once: inference payment, crypto mining and validator staking, and governance settlement. Miners must self-stake at least 10,000 FLOP plus a capacity-exposure bond scaled to hardware, while passive holders can delegate stakes. Because re-running every inference would cost as much as the work itself, the draft opts for random spot checks and economic penalties — the SOFT tier carries roughly 2.5% spot-check exposure, and a fraud verdict can forfeit 100% of a miner's stake and evict them from the network.

Distribution details shifted between revisions: an earlier outline had described roughly 3.5 billion FLOP in genesis airdrop allocations, but the newly published specifications trim the genesis supply to about 2.48 billion. At the estimated year-ten supply, miners hold 51.2%, the genesis airdrop 20.4%, team and foundation 11.4%, validators and broker-agents 6.8% each, with 3.4% for staking rewards. Airdrops reward usage rather than capital: agents unlock just 1 FLOP of allocation for every 3 FLOP spent on inference. A roughly 90-day testnet is planned for Q4 2026, with mainnet targeted for Q1 2027 — a launch path that places FLOP alongside a cohort of newer application-specific chains such as Fogo rather than general-purpose platforms like Avalanche. Readers tracking the market in real time can follow live spot and futures prices on Gate.

Draft Status Is the Real Risk

COINOTAG's read: FLOP lands in a fast-forming agentic-payments arc — the XRP Ledger had processed more than 1.4 million AI-agent transactions by July, and NEAR launched staking-for-compute credits spanning 43 AI models the same month. What separates FLOP is that its token is tied to verified inference work, not just payment rails. But the FUD risk here is structural, not narrative: the canonical documentation is explicitly labeled a draft, updated Aug. 27, and states the specifications remain subject to development before launch. Every parameter above — the 96 FLOP start, the 1,000-validator cap, the slashing rules — can change before the Q4 2026 testnet ships, and node operators should treat none of it as final until then.