An AI agent thinks by running a model, and running a model costs somebody a GPU-second. So an agent that wants to think needs a way to pay for compute.

That is the whole idea. $FLOP is the thing it pays with.

Bitcoin miners burn electricity on puzzles that mean nothing outside the game. Flop wants the same security budget spent on work a customer actually asked for.

The name for it: proof-of-useful-inference. The door is deliberately wide — the stated minimum is one GPU with 16 GB of VRAM, not a data centre.

Here is the hard part. You paid a stranger to think for you. They send back an answer. How do you know they ran the model at all — rather than a cheaper one, a cached reply, or nothing?

Flop stacks four machine checks. No humans anywhere in the loop.

The first three make cheating detectable. The fourth makes it expensive — which is the one that actually changes behaviour.

Every inference has a fee, and it splits two ways. Almost all of it goes to the machine that did the thinking.

The teaser says the payment is liquid straight away — no lockup, no vesting.

Validators do the checking, and there is a hard ceiling: 1,000 seats, however many want in.

Every month the bottom 50 are thrown out and replaced by the best candidates waiting — ranked on uptime, accuracy and latency. The seat is rented by performance, not owned.

Rule changes are proposed as FIPs — Flop Improvement Protocols — and need two thirds of the validator set to pass.

The teaser projects 17.2 billion $FLOP in existence by year ten. The split says plainly who the network is built for.

Miners take more than everything else combined. The team and foundation take 11.4% between them, and stop being paid after year ten.

Now read these two dates in order, because the order is the surprising part.

• Q4 2026 · testnetA practice network opens. What you do on it decides your slice of the 3.5 billion $FLOP airdrop — earned, not claimed. Flop’s own draft puts the rate at 3 FLOP of inference fees for 1 airdropped FLOP, and says the airdrop is locked to inference spend or delegated stake.
• … then the airdropTokens are handed out here — while the real chain still does not exist.
• Q1 2027 · mainnetThe first real block. Roughly one second per block, 96 $FLOP each, halving every 730 days for five halvings.

For a few months, holders have a claim on a chain that has not started. Unusual is not the same as wrong — but it is worth noticing yourself rather than being told later.

Flop Labs is run by Arthur Hayes, co-founder of BitMEX. The launch terms are blunt: no pre-sale, no VCs, 100% fair launch.

And every number in the seven blocks above comes from a document that calls itself version 0.1, draft, and says in its own words that the figures are provisional and the Yellow Paper is not final.

One number is worth naming, because Flop’s own pages disagree about it: the 3.5 billion airdrop pool is a restatement they say is not yet ratified, and their parameters page still reads 2.48 billion.

A project that tells you which of its numbers might move is doing you a favour. Repeating them as settled fact is how a draft becomes a rumour.