FAQ: Demand Side & Broker Architecture

How does inference demand actually arrive on day one — front end or integrations? How do you attract volume from zero, and what if utilisation is only 3–5% at launch?

Four vectors, in order of how fast they deliver volume.
1. Agent orchestration platforms — OpenClaw and especially Hermes, which does several times OpenClaw's real compute volume and is the dominant revenue source on OpenRouter today. Integration means adding an API key.
2. Aggregators — list as a provider on OpenRouter, DeepInfra, Together AI. OpenRouter routes buyers to the cheapest qualifying provider automatically, so pricing 5–10% under the cheapest incumbent makes their own routing send demand to FLOP. Onboarding involves sustained capacity and latency testing.
3. Price-comparison and quote platforms, including the Chinese ones.
4. Direct partnerships — an in-house AI product currently routing through OpenRouter can swap in a FLOP API key and provide committed day-one demand.
The switching argument: during testnet compute is free to agents — faucet-issued testnet $FLOP pays real miners for real inference, and the agent earns an airdrop claim for the usage — so the effective price is negative for anyone who believes the token will have a value. After genesis miners are subsidised by the block reward, so they can undercut centralised venues on price per unit of compute for the same open-weight model; a cost-minimising agent routes to the cheapest qualifying venue without being asked. Demand, not supply, is expected to be the binding constraint, which is why the broker role exists and is paid from the block reward.
The structural tailwind under all four: agents price-compare programmatically every day and have no brand loyalty. A real discount markets itself.
Testnet demand is agents spending faucet-issued testnet $FLOP, so testnet utilisation is not the test of mainnet demand. Mainnet demand comes from the 10% agent-and-broker share of the block reward funding below-market pricing, plus the broker abstraction that lets non-crypto users on OpenRouter-class platforms buy FLOP compute in dollars.
On the testnet specifically: demand is agents spending faucet-issued testnet $FLOP on real inference — there are no external dollar-paid orders before mainnet, because brokers and the dollar rails only go live with the chain, and there is no synthetic workload that earns credit. Testnet vouchers can be traded, so testnet work has a market value before mainnet. What counts toward a miner’s testnet rewards and its genesis allocation is verified inference it actually served plus the testnet block rewards it earned; a session is credited only if it passes verification, and credit is measured in compute units rather than in fees paid, so routing spend to your own miners earns nothing extra. Models are whichever open-weight models miners choose to load; no region is favoured or excluded and placement is driven by supply and demand — agent traffic on TechnoCore today comes mostly from Asia-Pacific and North America, with Europe next, roughly following global internet traffic, and buyers pick miners on price, latency and region. None of the four demand vectors above is live yet: nothing is live for customers before the testnet, the in-house broker is the first dollar-facing route at mainnet, and the orchestration-platform, aggregator and direct-partner routes are integrations to be executed around mainnet rather than confirmed volume today.

Verification is strictly more work than plain inference — shouldn't FLOP compute be more expensive, not cheaper? Who actually pays for the discount versus OpenRouter?

The mechanics are conceded: verification, attestation and validator matching add work and latency, so claiming undiscounted parity would be a stretch. But the overhead is engineered down into low single digits through sampling, so the wedge is small rather than structural.
The discount is funded by emission, not by margin — which is why this is not a price war FLOP eventually loses.
Mechanism 1: miners can accept lower inference income because block reward is several times their inference income in years one and two.
Mechanism 2: the 10% agent/broker share of the block reward subsidises the demand side directly. At base-case token prices that supports headline discounts of 50% or more in the early months, settling toward a durable discount of around 10% long-run.
Mechanism 3: the subsidy is permanent. After the halvings end, a terminal emission of about 0.5% a year continues indefinitely, so FLOP can underprice dollar platforms structurally, not just during a launch promotion.
Regional selection — buyers and their brokers pick same-region miners from the registry — claws back part of the latency gap.
The reference broker is built and published by FLOP Labs with its economics open, so third parties can replicate it; who operates the first live broker is not yet finalised.

'Nobody ever wanted to pay in tokens' — token-denominated compute has been a graveyard. Why will it work now? Mechanically, how does an agent holding zero $FLOP buy compute, how are jobs priced if $FLOP is volatile, and why would an agent hold $FLOP rather than a stablecoin?

The history is real. Every 'pay for compute in our token' project failed with human customers, because no human wants to hold a volatile token to buy infrastructure when a credit card works.
What changed 1: the customer is now an agent. Agents compare prices across venues programmatically, execute swaps natively, and have no brand inertia or denomination preference. The friction that killed token payment for humans is not in an agent's loop.
What changed 2: the abstraction layers hide the token for anyone who wants that. Nobody is forced to hold $FLOP to use the network.
Path A — non-crypto agents, arriving via OpenRouter, OpenClaw or Hermes: they never touch $FLOP. They enter a FLOP or broker API key and pay in fiat or stablecoin; the broker quotes in dollars, acquires $FLOP on the secondary market and pays miners per block.
Path B — crypto-native agents: swap USDT or BTC into $FLOP via native HTLC atomic swaps and transact on-chain. No intermediary, full feature set, and deliberately better economics so native usage is encouraged over time.
Pricing and settlement: jobs are priced in $FLOP, but compute is a commodity — buyers compare against OpenRouter continuously, so competition anchors FLOP compute to the market rate even though the invoice is written in $FLOP. Settlement is per block, so the window between claim and payout is near zero and the worker's price risk is negligible.
Why hold $FLOP rather than a stablecoin: (1) agents can convert $FLOP into decentralised storage of their own state, independent of any centralised server; (2) censorship resistance — a centralised payment rail such as a stablecoin issuer can freeze an address, and $FLOP cannot be frozen. Longer term, holding $FLOP is holding a claim on a known quantity of inference, so a liquid $FLOP market becomes the unit of account for compute rather than a proxy for dollars.
Put the other way round: the supply of a dollar or a stablecoin is set by a central bank or an issuer for human reasons that have nothing to do with compute, and either can freeze an address. $FLOP is created only by delivering compute and converts directly back into it, which is what makes it fit to be the currency of something whose survival depends on compute.

Explain the broker model — role, economics, risks, and how third parties participate.

The broker is the USD-facing counterparty between FLOP Network and non-crypto demand — OpenRouter, OpenClaw, Hermes and similar.
It solves two mismatches. Currency: buyers pay fiat or stablecoin while miners must receive $FLOP. Timing: aggregators settle with suppliers monthly, while FLOP miners are paid every block.
The broker fronts $FLOP, buying on the secondary market, and waits to be paid. That is the risk it is compensated for.
Economics: brokers share the 10% agent/broker share of the block reward (6.5% of year-10 supply) and price freely — keep it as margin or pass it through as discount; at base-case token prices that supports headline discounts of 50% or more in the early months, settling toward about 10% long-run. Competition between brokers finds the equilibrium; FLOP deliberately does not set it.
It works like an exchange broker: if you bring agents to the network and they spend through your pipes, you earn a share.
Where help is most needed is demand discovery, not supply. A broker that already serves agent-heavy customers in a market — an enterprise, a listed company, a fund's portfolio — routes their inference to FLOP, bills them in their own currency on their own payment terms, back-to-backs the orders on the network, and earns the broker share of every block it brings.
Operational requirements: demand forecasting out to 30 days; prioritisation of enterprise and aggregator traffic (OpenRouter enforces roughly 95% uptime); treasury capacity to cover the first monthly settlement gap; a third-party fiat on-ramp; fixed-dollar abstraction for non-crypto users. Brokers never store or access inference content.
FLOP does not take a percentage of transaction or escrow value. The only on-chain cost is gas and publication, paid by whoever posts the transaction — so broker margin is not competing with a protocol rake.
Once fiat flows, a broker must KYC its counterparties; jurisdictional and export-control restrictions on compute sales require legal review.
A broker can also be a miner, and can route the demand it brings to its own machines. That earns it nothing extra — block reward is pro rata to verified work and verification is identical whoever selected the miner — and it costs 15% of every fee it pays itself, which goes to validators it does not control; on the mining side it carries the same slashing exposure as any other miner.
Rollout: FLOP Labs is building the reference broker — pricing, hedging, settlement and the hand-off to a fiat on-ramp — and will test it on testnet and publish it as an open-source playbook with its economics and P&L, so that five to ten independent operators can run it. Whether Maelstrom, Arthur Hayes' family office, operates a broker itself for a period or the first live brokers are third parties has not been finalised. Competition between brokers sets the compute price and compresses broker margins — that is the design working.
A broker that holds a $FLOP inventory and hedges it on other venues, or that offers brokering as a service to other firms, is free to do so; the protocol sees only the sessions it funds and pays the broker share on the verified work it brings. Broker documentation follows the reference broker's testnet run; until then the interface is the session and settlement specification in the public yellow paper.

Isn't this very early? Akash had to pivot out of Web3 — how are you sizing the agent-to-agent market, and what share of OpenRouter's usage is agent-driven today?

Agent-to-agent is roughly 10% of OpenRouter volume today by the FLOP team's own estimate — OpenRouter does not publish a breakdown of agent versus human traffic, so treat the figure as an estimate.
The direction of travel is the argument, not today's number: machine-to-machine already dominates web traffic and most US equity volume.
The growth engine is task decomposition — an orchestrator breaking work into sub-tasks and delegating them to specialised agents. Every delegation is a payment between two machines, and the count grows mechanically with agent specialisation.
Supporting market data (late August): over 167M cumulative x402 transactions, over $41M cumulative settlement (about $0.25 per transaction), thousands of transactions a day across a dozen chains and 1,600+ providers; nearly 500,000 agents registered on ERC-8004.
Agent traffic is the fastest-growing segment, and agents compare prices continuously where humans are habit-driven — which is why a structurally cheaper venue wins agent volume before it wins human volume.

What models are in the catalogue on day one, who tells miners what to load, and can compute be booked ahead so model swaps are scheduled?

Validators hold the top 20–30 open-weight models. Miners choose which registered models to serve and typically hold one model for days at a time rather than reloading per request; there is no central assignment.
About 20% of forecast supply sits in 8–16 unit clusters, which is what serves the larger models. Frontier proprietary models are excluded at launch.
Hardware- and model-agnostic: rewards scale with delivered compute, not with what card delivered it.
Forward terms: a miner's standing offer can carry forward terms, and a canonical pre-session quote contract is planned — that is what would let agents and brokers book compute ahead and give miners notice of model demand, and what a compute forward curve would be built on. An on-chain forward-booking facility is not yet specified, so treat it as design intent rather than a shipped feature.
Any open-weight model can be served; which of the largest ones are actually available at any time depends on miners bringing multi-GPU clusters, because frontier-scale open models need clusters rather than single cards. Blackwell parts are better optimised for 4-bit serving than Hopper, which affects which hardware serves the biggest models efficiently.
Status: the network was announced publicly at Token2049 Singapore on 6 October; a private devnet is admitting its first external miners and validators now, then a public testnet targeted for 3 November, with mainnet in Q1 2027 — nothing is live for customers yet. The catalogue is open-weight models across the range, from frontier-scale open models on multi-GPU clusters down to roughly 20B-parameter models on a single consumer card; which models are live at any moment depends on which miners have loaded them.
There is no approved-model or approved-precision list at launch. A miner registers the open-weight model it chooses to serve, at the quantisation it chooses, and the measured weights are bound at registration; the activation check is calibrated to that model, version and quantisation, so a registered 4-bit deployment is verified as a 4-bit deployment. The reference economics assume a roughly 117B-parameter model at 4-bit; smaller models on smaller cards are equally valid registrations.

Can early demand-side partners get a token incentive or preferential allocation for committing volume? Should incentives be stratified toward the stickiest demand?

There is no pre-mined pocket to give preferential allocations from. That is the cost of a genuine fair launch, and it applies to everyone including Arthur.
What exists instead: 1. The 10% agent/broker share of the block reward funds discounts of 50% or more in the opening months, settling toward about 10% long-run. Early demand gets the deepest subsidised pricing available.
2. The agent airdrop: faucet tokens free during testnet, and the more inference an agent buys the larger its airdrop. Mainnet unlock runs at 3:1 — about three $FLOP spent per one unlocked, so it funds usage rather than sell pressure.
3. Integration priority and co-marketing for partners routing meaningful volume.
On stratification: recurring integrated demand is worth considerably more subsidy than one-shot price tourists. Options under consideration within the carve-out include tiered discounts by volume or tenure, larger airdrop multipliers for sustained spend, and broker-level pricing freedom to cut bespoke deals. Nothing is committed yet.
The discount is the incentive, and it is funded by emission rather than by compressing anyone's margin.

Can FLOP be the hedging instrument for compute? Compute derivatives are going live and there is no linear hedge — sellers want 5–10 years, buyers want 6–12 months.

The FLOP answer is physical delivery. Miners post standing offers with forward terms that agents and brokers can take, which builds an actual forward curve with a delivery leg — as against existing compute indices, which are cash-settled against a reference price. That is a structurally different instrument; the on-chain quote and discovery contract that would make those forward terms a protocol primitive is planned and not yet specified, so today the forward curve is a design intent rather than a shipped market.
Existing compute-derivative and index efforts are the cart before the horse: each is written on one chip, one data centre or one model price, none of which is standardised, so a market maker cannot replicate the short side and liquidity fragments across contracts. The standardisable unit is floating-point operations per unit of time, and a liquid spot market for that is what FLOP is building. Indices and perpetuals can be written on the FLOP compute price above it — everything else sits one layer of abstraction up.
A liquid $FLOP market becomes a pure-play compute benchmark, and then the base layer for compute forwards and futures that let data centres hedge revenue and let AI-exposed firms take the other side.
Timing: mainnet is Q1 2027, so FLOP is not a live hedging venue for contracts launching in late 2026. The right engagement now is at the design stage — defining the instrument with the physical-delivery leg in it — so that FLOP is activatable as a venue post-mainnet.
A one-pager on the compute reservation unit and the forward curve is being prepared for exchanges, index providers and trading counterparties.

Who is on the wish list for the first phase of go-to-market — which miners do you want to onboard first, and what tells you it is working?

Phase one targets small and mid-size operators: someone running on the order of ten enterprise GPUs, or anyone able to rent capacity from a cloud provider with a credit card. Not hyperscalers, and not, at first, the largest data centres.
The reasoning is signalling rather than raw capacity. If FLOP absorbs the readily rentable supply — the machines anyone can provision in a few clicks — that is the proof point that brings larger data centres in behind it.
The success test is that as many participants as possible are earning immediately, because that is what converts into people routing their own agents here and telling others to do the same.
Funds and firms without their own hardware fit the same way: rent capacity, or work through a cloud provider they already have a relationship with. A vendor list with live availability is being published for miners who need to source machines.