Arthur Hayes has released the FLOP Network technical paper detailing a proof-of-useful-inference blockchain for AI agents, with a genesis supply of approximately 2.48 billion FLOP tokens allocated through airdrops.
Summary
- Arthur Hayes has released the FLOP white paper detailing a blockchain where AI agents pay miners in FLOP for inference and computing resources.
- FLOP will launch with approximately 2.48 billion tokens allocated through airdrops, with no venture capital premine or token auction.
- Miners will receive 75% of block rewards, while validators and agents each receive 10% and regular stakers receive 5%.
- Block rewards will start at 96 FLOP and halve every 730 days until reaching a permanent reward of 3 FLOP.
According to the FLOP Network details shared by Hayes, the protocol is designed to let autonomous AI agents pay miners directly for inference, with validators settling proofs of the completed computational work.
FLOP, short for floating-point operations, serves as the native currency of the network. Under the proposed system, an agent can spend FLOP whenever it requires computing resources, allowing the token to function as payment for AI inference.
The latest specifications provide substantially more detail on a project that Hayes first unveiled in August. As crypto.news reported, the BitMEX co-founder said on Aug. 18 that he was returning to an operating role to lead Flop Labs, describing FLOP as “food for your AI agent.”
At the time, Hayes said the token would launch without a presale or venture capital allocation. Flop Labs was targeting a large airdrop in the fourth quarter of 2026 and a genesis block in the first quarter of 2027, although the project had yet to publish its full tokenomics and technical design.
FLOP Network turns AI inference into miner work
FLOP Network combines an account-based blockchain with a consensus model called proof of useful inference, or PoUI, where miners earn rewards by carrying out AI inference requests instead of performing computational work solely to secure the chain.
An AI agent begins the process by submitting a session request to the network’s mempool. Each request specifies information including a model-weight hash, maximum latency, computational requirements measured in FLOPs, a confidentiality setting and the fee offered for completing the task.
A miner with suitable hardware can accept the request and establish a private connection with the agent. Once the requested inference has been completed, the miner returns a proof of the work, while validators incorporate the proof hash into a block to settle the transaction.
Miners receive the session fee paid by the agent and a portion of the network’s block rewards based on the verified compute they contribute. Ordinary GPUs can participate under the proposed architecture, while confidential computing is treated as an optional tier instead of a requirement for joining the network.
The model places FLOP in an emerging market where autonomous software is already being equipped to make blockchain payments. A June crypto.news guide on agentic crypto payments detailed how AI agents can autonomously purchase data, compute and other online services using crypto payment rails such as x402.
Similar infrastructure has started moving into production. By July, the XRP Ledger had processed more than 1.4 million AI-agent transactions, while Ripple was developing tooling that allowed autonomous agents to make payments using XRP and RLUSD.
FLOP’s proposed model differs by tying its native currency directly to inference work performed by miners.
FLOP supply starts at 2.48 billion tokens
The FLOP specifications put the genesis supply at approximately 2.48346 billion tokens, with the initial tokens designated for airdrop distribution rather than a venture capital premine or token auction.
Block rewards begin at 96 FLOP and are divided among four groups. Miners receive 75%, validators take 10%, agents receive another 10%, and ordinary stakers are assigned the remaining 5%.
Issuance follows a fixed halving schedule. The 96 FLOP block reward falls to 48 after 730 days, followed by reductions to 24, 12, 6 and eventually 3 FLOP across five halvings.
Unlike Bitcoin’s issuance model, rewards do not disappear after the scheduled reductions. FLOP’s block subsidy remains permanently at 3 FLOP after the fifth halving, creating continuing emissions for network participants.
The network is designed around an average block time of one second with deterministic sub-second finality, while its development roadmap targets block production below one second.
Hayes had previously provided an earlier outline of the distribution strategy. In August, he proposed allocating roughly 20% of the FLOP supply to testnet participants over a 10-year period and said the network would be funded without a token presale.
That proposal described two revenue sources for miners: block rewards for participating in the network and inference fees paid by AI agents requesting computational work. The latest protocol documentation now specifies how the requests move from agents through miners to validators for settlement.
Validators face staking and slashing rules
Participation as either a miner or validator requires FLOP to be staked, creating collateral that can be penalized when participants submit dishonest work.
Under the proposed rules, miners can face slashing for misrepresenting completed inference, while validators risk penalties for publishing dishonest blocks. Severe violations can result in the full loss of staked tokens and removal from the network.
Token holders who do not operate infrastructure can delegate FLOP to a miner or validator and receive a proportional share of rewards.
The validator set is capped at 1,000. Approximately 50 validators are expected to rotate each month based on verified workload and uptime, while the project plans to incorporate stake into validator ranking above the required participation threshold.
Validators will have a second role in network governance. Changes are proposed through FLOP Improvement Proposals, or FIPs, with most proposals requiring approval from two-thirds of the active validator set before implementation.
The architecture places model weights in a data-availability layer while validators build blocks containing hashes of inference proofs submitted by miners.
Other blockchain projects are pursuing their own versions of an AI-agent economy. NEAR, for example, introduced a system in July that lets users stake tokens for AI services, converting locked NEAR into monthly compute credits that can be used across 43 AI models, including confidential inference and autonomous agent services.
FLOP instead proposes a four-stage execution path built around requests, miner matching, inference proofs and settlement. An agent first posts the required model, latency, compute, confidentiality and fee parameters; a miner accepts the task and runs the model; proof of the completed inference is submitted to the network; and validators include its hash in a block before the miner receives the session payment and its share of protocol rewards.
The FLOP Network project introduction is currently labeled a draft and was last updated on Aug. 27, while the newly published technical specifications remain subject to development before the planned network launch.
