Arthur Hayes released a fuller design for FLOP Network, a proposed blockchain where artificial-intelligence agents pay miners for inference and validators settle proofs of the completed work. Hayes pointed readers to the project documentation in a Sept. 7 post, moving the initiative beyond the brief outline he presented in August.
The project introduction remains explicitly labeled a draft and says it was updated Aug. 27. Its specifications describe intended behavior, not a running network, and the design can change before any launch.
Agents would buy inference with FLOP #
Under the proposal, an agent sends a session request to the network mempool. The request identifies a model-weight hash, a maximum latency, the computational requirement measured in floating-point operations, a confidentiality setting and the fee offered for the task.
A miner with suitable hardware accepts the request, runs the inference and returns a proof. Validators then place the proof hash in a block, after which the miner receives the session payment and part of the block reward. The document says ordinary graphics processors could participate, while confidential computing would be an optional service tier.
Staking would back miner and validator behavior #
FLOP’s draft requires miners and validators to stake the network token. Dishonest inference claims or blocks could trigger slashing, creating a financial penalty for false work. Holders that do not operate infrastructure would be able to delegate stake and receive a proportional share of rewards.
The specification targets one-second average blocks and caps the validator set at 1,000, with roughly 50 expected to rotate monthly according to work and uptime. Those are project targets rather than measured performance. Most governance changes would require approval from two-thirds of active validators.
The network still has to prove useful work #
Crypto payment systems are already being designed for autonomous software, including payment rails that let AI agents transact without a human approving each step. FLOP’s distinction is its attempt to connect the currency directly to verifiable inference rather than use a general-purpose token only for settlement.
The hard part is verification: the system must show that useful model work occurred without exposing sensitive prompts or making validation more expensive than the computation itself. The draft outlines matching, proof and settlement, but it does not establish production throughput, demand or security. Until a test network and independent code review exist, FLOP should be treated as a proposed architecture.