BitMEX co-founder Arthur Hayes has published the full technical blueprint for FLOP Network — a blockchain where artificial intelligence agents pay miners directly for computing power, launched with a roughly 2.48 billion token airdrop and no venture capital premine. It is one of the most concrete attempts yet to make AI pay for its own brain.
By Amir Hassan | September 7, 2026
The Hook: A Currency for Machines
Hayes returned to an operating role on August 18 to lead Flop Labs, calling FLOP “food for your AI agent.” At the time, the project had a vibe and a promise — no presale, no VC allocation — but little technical meat. This week’s release of the technical specifications and tokenomics changes that, laying out exactly how a blockchain could run on AI inference — the process of a trained AI model actually producing answers — instead of pointless number-crunching.
The idea in one sentence: today’s blockchains pay miners to solve useless math puzzles to secure the network; FLOP proposes to pay miners for doing useful work — running AI models on behalf of autonomous software agents that pay for the service in FLOP tokens.
How Proof of Useful Inference Actually Works
FLOP — short for “floating-point operations,” the unit that measures computer calculations — uses a consensus model called proof of useful inference (PoUI). The flow has four stages:
- The request — an AI agent posts a job to the network specifying the model it needs, the maximum latency it can tolerate, the required computing power, a confidentiality setting and the fee it will pay.
- The match — a miner with suitable hardware accepts the job and connects privately with the agent. Ordinary GPUs can participate; confidential computing is an optional tier, not a requirement.
- The proof — once the AI work is done, the miner submits a cryptographic proof that the work was performed honestly.
- The settlement — validators embed the proof’s hash into a block, the miner collects the session fee plus a share of block rewards, and the transaction is final.
The network targets a one-second average block time with deterministic sub-second finality — meaning a payment is truly settled almost instantly, not just “probably confirmed.” Validators face real teeth: participants must stake FLOP as collateral, and dishonest miners or validators can be slashed — penalized with the loss of staked tokens — with severe violations resulting in full loss and removal from the network. The validator set is capped at 1,000, with roughly 50 validators rotating each month based on verified workload and uptime.
The Tokenomics: Bitcoin’s Design, Rewired for AI
The supply design borrows familiar ideas from Bitcoin and then deliberately departs from them:
- Genesis supply — approximately 2.48346 billion FLOP, allocated through airdrops rather than a venture capital premine or token auction.
- Block rewards start at 96 FLOP — with miners receiving 75 percent, validators 10 percent, agents 10 percent, and ordinary stakers 5 percent.
- Halvings every 730 days — the reward falls from 96 to 48, then 24, 12 and 6 FLOP across five scheduled halvings.
- No terminal zero — unlike Bitcoin, where issuance eventually ends, the FLOP block subsidy settles permanently at 3 FLOP after the fifth halving, creating continuous emissions forever.
Governance runs through FLOP Improvement Proposals (FIPs), with most changes requiring approval from two-thirds of the active validator set. Hayes had earlier proposed allocating roughly 20 percent of the supply to testnet participants over 10 years, with the project targeting a large airdrop in the fourth quarter of 2026 and a genesis block in the first quarter of 2027.
The Core Conflict: Real Demand or a White Paper?
The bet FLOP is making — that autonomous AI agents will need their own payment network — is not fantasy. The building blocks already exist. By July, the XRP Ledger had processed more than 1.4 million AI-agent transactions, with total fees paid amounting to roughly 280 USD — a tiny number that proves the concept works at negligible cost. Crypto payment standards like x402 already let AI agents autonomously buy data and services online. NEAR introduced a system in July letting users stake tokens in exchange for monthly AI compute credits across 43 models.
FLOP’s twist is tying its native currency directly to inference work. That is elegant on paper — but the document itself is labeled a draft, last updated August 27, and the specifications remain subject to development before launch. A token that pays miners for AI work only succeeds if real agents actually show up and pay. Until the mainnet runs, everything is a promise.
The Verdict: Watch the Airdrop, Question the Demand
For regular investors, FLOP is a story to track, not chase. The airdrop model — no VC premine, community-first distribution — is the most retail-friendly launch structure in crypto, and the Q4 2026 testnet will show whether the four-stage design survives contact with reality. The honest checklist: does the network launch at all, do AI agents generate real fee volume, and does permanent 3 FLOP issuance keep miners engaged or quietly dilute holders forever?
What FLOP undeniably gets right is the direction: AI agents making machine-to-machine payments is happening today on other chains. Hayes’ network is the boldest attempt yet to make that economy the blockchain itself — an idea worth understanding now, before the genesis block in 2027 decides whether it was genius or a very well-written white paper.
The cryptocurrency market remains highly volatile. This article is for informational purposes only and does not constitute financial advice.
paying miners for real inference instead of a hash lottery finally makes the security budget do something useful. and hayes actually shipped specs instead of just posting essays, respect
agreed but who verifies the inference output is actually useful and not garbage compute? the verification problem is where every PoUI design dies
2.48 billion tokens with zero VC premine is the only reason anyone trusts the distribution. every other AI chain launched with a stealth allocation
no presale, no VC, airdrop to actual users. almost suspicious how reasonable this is for a hayes project lol
no VC premine is the only part i trust. 2.48 billion tokens going to actual users, hayes knows distribution is the whole game
AI agents paying miners for inference directly, no cloud middleman taking a cut. If the specs hold up this is the first design that makes sense for machine payments.
cool blueprint, but the last hayes launch pumped for 2 weeks then bled for a year. airdrop farmers arent gonna run inference nodes lol