Give Us This Day Our Daily FLOPs

The caveat first: FLOP has no whitepaper, no testnet, no code, no named chain and no named GPU supplier. As of September 4 it has a named executive team. Everything here is read from what has been published, and what has been published is thin. It is also unusually clear about one thing.

The announcement on August 18 set the shape: Hayes as CEO, self-funded, “100% fair launch” with no presale and no venture allocation, an airdrop in Q4 2026 and a genesis block in Q1 2027. The essay that followed the next day, The Book of Genesis of August 19, is a manifesto rather than a specification, but it contains the sentence that matters for market structure: “$FLOP is food for AI agents because it is a direct claim on compute provided by miners on the network.” The stated defect in the existing market is that cloud rentals “did not charge based on actual FLOPs used, but some other abstracted concept,” and the stated fix is “a globally unified market where one could pay using a unit of any currency for a certain amount of FLOPs.” In an interview a week later he put it as a market-structure claim: “There is currently no market that efficiently converts a unit of currency into a set amount of floating-point operations over a given period.” He has promised a further essay on “why a spot market for actual compute, that is floating point operations per unit of time, is a necessary prerequisite for the growth of the agentic economy.” It has not appeared.

On September 4 Flop Labs published an “Introducing Flop Labs” note on LinkedIn naming Sergey Vidyuk as Chief Technology Officer, previously an exchange architect at BitMEX, and Shu Duan as CSO, from Deutsche Bank Hong Kong and Aspen Digital, alongside Hayes as CEO. Those three are the entire named organisation; no engineering, verification or chain team is announced beyond them. The note is worth reading for its market-structure claims as much as its roster. It says compute is “one of the largest costs in the global economy, and there is still no way to look up its price,” that “one vendor bills by the token while another bills by the hour, and the two do not translate,” that “without a common unit there is no benchmark price, and no way to hedge a cost that keeps climbing,” and that “oil trades against a barrel and power against a megawatt hour. The unit came first, and the market formed on top of it.” It describes the transaction as: “A buyer names the model and the amount of work, and sets a deadline. Operators supply the GPUs and deliver the result. They are paid in $FLOP only for work the network confirms was done.” Sections 02 and 06 take those claims one at a time.

Two days earlier Flop Labs had put the same case more sharply on X: “To find out what an hour of GPU time costs today you still have to email a salesperson. Oil and wheat both got exchanges. Compute got a pricing team.”

The teaser, version 0.1, dated August 26, is the only quasi-technical document, and Figure 1 is most of it.

In The Compute Crack Spread we wrote the producer’s margin as C = T·p − r: tokens per GPU-hour T, times the price of a token p, less the rent on the GPU r. The rent is the thing every announced instrument settles on. CME and Silicon Data’s H100 and B200 contracts, Ornn’s forward curves, Kalshi’s event ladders, Compute Desk’s benchmarks under General Index, Liquid Compute’s LCI: all of them are r. T, the heat rate, is what converts a rented hour into delivered work, and it is the term no contract touches, because it lives inside the operator’s serving stack and nobody outside can observe it. That is the denominator problem: a dollar-per-GPU-hour print tells you what an hour of nameplate cost, not how much of the hour did anything.

The September 4 note says there is “still no way to look up” the price of compute and “no benchmark price.” For the hour that is simply not so: OCPI, SDH100, LCI and the Compute Desk benchmarks all print a dollar rent, CME lists futures on two of those prints on October 5, and Ornn has published forward curves since July. What the note gets right is the sentence next to it. The token bill and the hourly bill “do not translate,” and the reason is T: the number that converts one into the other is the one nobody outside the serving stack can see. So the accurate version of the claim is narrower and more interesting than the one they made. The hour has a price. The work does not.

Hayes’s objection to cloud pricing is the same objection in his own vocabulary. “Some other abstracted concept” is the GPU-hour. “A set amount of floating-point operations over a given period” is T times the hour. The difference is what each side does about it. The benchmark and derivatives work treats T as a ratio to be estimated so that r can be converted into a token cost. FLOP does not estimate the ratio. It changes the thing being sold, so that the buyer pays for the numerator directly and never sees the hour at all.

That is a real design choice and it has real consequences, in both directions. It is the one thing in the record that is genuinely different from the rest of the compute-market build-out, and it is why FLOP should not be filed alongside the rental benchmarks it is usually compared to.

Figure 2 draws the chain from silicon to enterprise spend and marks what prices each node. The rented hour has five benchmarks and three venues. The delivered-work node has nothing. The token node has list prices, routers and a research literature. FLOP’s session sits on the seam between the last two.

Three things follow from the placement, and they are the substance of the analysis.

It is a work market, not a capacity market. The buyer on FLOP never rents a chip; it buys a completed session. Every dollar of utilisation risk therefore sits with the miner. In the language of power markets, FLOP is not a tolling agreement, where the buyer takes the plant and bears the dispatch. It is a real-time balancing market that hopes to absorb surplus generation. That is why Hayes can say, without contradiction, that the AI bubble is “debt issued to build data centers” and that “the compute overcapacity built with borrowed money bolsters my thesis.” An oversupplied balancing market clears low, and a low clearing price is the network’s customer-acquisition plan. The teaser says so: launch pricing below market “is self-marketing.”

It competes at the token node, not the rental node. A session is model-specific, latency-bounded and paid per completed inference. The products that already sell that are inference routers and the decentralised-inference networks. Ornn, Silicon Data, Compute Desk and the CME contract sell something else, to someone else, for a different reason. Hayes’s own admission that a “spot market for actual compute” is a prerequisite he has yet to write about is the tell: the network as specified is a retail inference marketplace with a currency attached, and the wholesale compute market is a separate, unbuilt thing.

It writes a dealer into the tokenomics. The line that brokers “quote fixed-dollar inference backed by $FLOP-settled sessions” is a merchant seat, and the 6.8% brokers-and-agents allocation is the bootstrap payment for taking it. Whoever sits there warehouses two bases at once: $FLOP against the dollar, and $FLOP against delivered FLOPs, the second of which is subsidised and heterogeneous. Of all the compute-market designs announced this year, FLOP is the only one that pays a dealer explicitly rather than assuming one will appear. That is a point in its favour on market-structure grounds, and a point against anyone who thinks the seat is easy.

Each of the three is a consequence of the placement, not a criticism of the ambition. They are what a work market has to solve that a capacity market does not.

The fleet it will attract is not the fleet the bubble thesis needs

A neocloud servicing fleet debt needs dollar cash flow with a predictable margin. It will not sell into a $FLOP-denominated, subsidised, permissionless market at scale unless someone else absorbs the currency and price risk, which is the broker seat, which does not exist yet. The plausible early miner base is what the hardware guidance describes: a GPU or cluster with 16 GB+ VRAM, which is a consumer card, a hobbyist rack, or the depreciated tail of an enterprise fleet. That is a perfectly good salvage market. Prior-generation chips do get re-employed rather than written to zero, and a permissionless, model-agnostic network is a sensible place for the A100s and the tenant-churned B200s to go. It is not the debt-financed H100 fleet that Hayes says makes the bubble. FLOP addresses the utilisation of the old fleet, not the utilisation of the fleet whose financing is the story.

A FLOP is not a FLOP

Decode-phase inference is memory-bandwidth-bound, not FLOP-bound. FP4, FP8 and BF16 operations are not interchangeable. A 16 GB consumer card and a TEE-attested H100 running the same model hash deliver different latency and throughput, and the confidentiality flag routes to a different hardware class again. The session specification partly acknowledges this by naming the model and bounding the latency, but the consequence is that there is no single “price of compute” on the network. There is a price surface indexed by model, latency, confidentiality and hardware, and the grade spread and tier basis that the rental benchmarks struggle with reappear one node down. This is the same defect the open-source normalisation projects have: declaring the unit is not the same as making it reproducible. Any price series drawn from FLOP has to be computed per grade or it is noise.

A subsidised price in a floating numeraire is not a signal

Block rewards route to the demand side. Launch pricing is deliberately below market. Testnet allocation is proportional to compute delivered and consumed, which is an invitation to farm both. Each of these guarantees that the first several quarters of $FLOP-denominated prices, and every testnet volume statistic, will be gamed and subsidised. On top of that the numeraire floats: a delivered-FLOP price in $FLOP carries $FLOP-dollar volatility on top of compute-price volatility. The stated reason for a native currency, that agents have no legal personhood and so stablecoins are a poor fit, is a custody and KYC argument, and it does not explain why the unit of account must float. The velocity objection that has already been raised in the coverage, that miners will convert to stablecoins on receipt and the token’s demand evaporates at settlement, is the flip side of the same choice. Either the staking and broker sinks hold or the currency leg dominates the compute leg in every print.

Set the token aside. If the network runs as drafted, its ledger carries a tuple per session: model hash, latency bound, FLOPs delivered, fee, and, where the TEE path is used, a hardware attestation. No rental benchmark carries anything like it, because no rental benchmark can see inside the hour. It would be the first public series of realised throughput at a price, and that is the numerator of the heat rate.

Put that series next to a rental print and something useful falls out. Convert the session fee to dollars at spot, divide by FLOPs delivered, and compare to the dollar rent on the hardware class divided by its nameplate FLOP rate. The ratio is an implied T: how much of the hour did work, on that model, at that grade. It is the utilisation term as an observable rather than a modelled quantity, which is the thing the crack-spread arithmetic has been missing since the rental contracts were announced.

Three conditions attach. It only exists once the subsidy has decayed and the farmers have left, so not from the testnet and probably not from the first year of mainnet. It only means anything per grade, so someone has to build the normalisation by hardware class and precision that the network itself has no reason to publish. And it needs both legs to be honest, which puts as much weight on the rental benchmark as on the ledger. None of that is FLOP’s job. It is an index provider’s job, and it is a product neither FLOP nor the rental benchmarks are positioned to produce on their own.

Hayes has said the next essay explains why a spot market for actual compute is a prerequisite. When it lands it will be the most-read piece on compute market structure of the year, and the distinction it will have to draw is the one this piece is about. A spot market for capacity is the rental market, and it exists, in five benchmarks and three venues. A spot market for work is what the teaser describes, and it does not exist anywhere at scale, on-chain or off. Conflating them is easy because both are called “compute,” and the whole question of whether FLOP is a hub or a router turns on which one he means.

The September 4 note reaches for the same history: “Oil trades against a barrel and power against a megawatt hour. The unit came first, and the market formed on top of it.” True, and it cuts the other way. The barrel became a unit because a barrel of West Texas Intermediate is interchangeable with the next one and the grade differentials to Brent, Dubai and Maya are themselves published prices. The note then describes the FLOP transaction as a buyer who “names the model and the amount of work, and sets a deadline.” That is not a barrel. That is a refined product to a specification with a delivery window, which is to say a crack spread with the crude leg missing. The unit they have chosen is the right one for a work market and the wrong one for the analogy they are drawing.

Merchant markets do sometimes grow a hub out of surplus. Henry Hub became the price of natural gas because deregulation left a lot of molecules looking for a clearing point and a pipeline junction where they were fungible. The second condition is the hard one. A session addressed to a model hash and a latency bound is not fungible with the next session, and a hub that cannot make its unit fungible is a router with a ticker. The teaser’s verification stack is a serious attempt at proving that work was done. It is not an attempt at making one unit of work interchangeable with another, and nothing in the record suggests that problem is on the roadmap.

Short of the essay, the record would move on any of the following: a Yellow Paper with an actual fee-discovery rule and a stated stake-to-compute ratio; a named GPU supplier or neocloud, which would move the miner base from the depreciated tail toward the financed fleet; disclosure of the demand-subsidy formula and its decay; external evidence that TOPLOC and TEE attestation hold up at scale; and, after genesis, whether session-level data is queryable and whether it segregates by hardware class. The last is the one that decides whether the by-product in section 05 ever exists.

Until then the placement stands. FLOP is a work market and a settlement rail at the delivered-compute node, with a dealer seat drafted in. It is to the right of every benchmark, to the left of the API price, and in a different competitive set from the one it is usually filed in. That is a more interesting place to be than “another decentralised compute token,” and a harder one.