The Refiner and the Merchant

In August the Federal Reserve Board published a working paper that treats compute as a non-storable commodity and asks what can be priced, hedged and securitized against it. Six weeks later the first listed chip-hour futures, due to start trading on October 5, were pushed back to November at the earliest, with the regulator describing the market beneath them as fragmented and opaque. Between the paper and the delay sits a practical question that neither answers: who in this market holds a position the contracts can hedge, and who only appears to.

This piece answers it by sorting the market into seats and following the two that pool. The refiner pools silicon from several makers behind one interface and sells tokens. The merchant pools commitments from several providers behind one price list and sells chip-hours. General Compute, which closed a $400 million debt facility in July secured on inference chips from SambaNova, is the clearest refiner on the record. The merchant is a position more than a company: the block trader, the forward originator, and, at the largest scale, any firm that rents a neocloud’s fleet to serve its own customers.

They are different businesses, and the sections take them in turn: what each is long, what each cannot hedge, how each sits against the listed contracts and the index under them, what the electricity retailers of 2021 teach about the merchant, and what a lender to either should write into the loan. The short answer is that the listed hedge fits one seat and not the other, and that for the seat it fits, the hedge and the squeeze arrive from the same event.

Every firm between a chip and a user answers two questions: does it own the silicon or rent it, and does it sell chip-hours or the tokens they produce. The answers sort the market into five seats.

The neocloud was the subject of The Customer Is the Collateral, and its credit ledger is not repeated here. The exchange matters to this piece because it is where the merchant’s residual can be sold, and because it is the one seat that transparency helps without qualification.

The refiner on the record. General Compute buys decode chips from SambaNova, lists Cerebras, Positron and d-Matrix as further options, and pairs them with graphics chips for the prompt-reading step. It runs them in rented colocation space, holds an option on 15 megawatts of air-cooled racks, and sells dedicated racks under long-term contracts, reserved decode capacity with burst, and a pay-per-token interface. Its $400 million facility from Upper90 starts at $100 million and draws as customer demand arrives; TechCrunch reports the collateral is the SambaNova chips, and that it may be the first loan secured on inference-specific silicon rather than GPUs. Citybiz reports $300 million of price-protected hardware supply agreements alongside it.

The exchange on the record. SF Compute does not own any of the GPUs it provides access to, manages access to more than $100 million of hardware, and raised $40 million in December 2025 at a $300 million valuation. Its order book takes buy and sell orders by chip, node count, duration and rate, and a buyer with unused time places a sell order and gets credits back, less a fee. Prime Intellect has integrated twelve clouds and quotes H100s at $1.50 to $4.00 an hour on demand. Compute Exchange runs auctions with one-to-thirty-six-month terms and a secondary market, and puts reserved terms at 40 to 70 percent below on-demand and spot at 60 to 90 percent below.

The merchant on the record is harder to name, because the merchant is a position rather than a company. Anyone who takes a block of capacity on a take-or-pay commitment and resells it in smaller pieces and shorter terms is one. Our block-trade arithmetic, from a wholesale H100 block near $2.50 an hour resold at retail near $3.40, puts breakeven fill near 74 percent: below that, the trade loses at any price. That number is an estimate, and it is the merchant’s whole business in one line.

Each pooling seat is long something it cannot hedge. The refiner is long the efficiency wedge, the rate at which purpose-built silicon turns a chip-hour into tokens, and short obsolescence on chips with no resale market. The merchant is long fill: its commitments run for months and its sales run by the hour, so every unsold hour is a loss taken in cash. The next two sections take each in turn.

A refiner’s margin is a crack spread. It buys a chip-hour, converts it into tokens, and sells the tokens. The spread is the token price times tokens per chip-hour, less the all-in cost of the chip-hour. Everything about the refiner’s risk follows from the three terms in that line.

Tokens per chip-hour is the edge, and it is the unhedgeable factor. General Compute’s case rests on splitting inference in two. Reading the prompt is compute-bound and stays on graphics chips. Writing the answer is memory-bound and moves to purpose-built decode silicon. The company claims 1,000 to 2,000 tokens a second per user on the decode chips against about 320 on graphics chips, and racks drawing 20 kilowatts against 120. Those are the company’s numbers, not audited ones. If they hold, the refiner produces more tokens per dollar of silicon than a graphics-chip cloud, and that gap is its margin. It is also, in the Federal Reserve paper’s terms, the efficiency wedge: the multiplier between the chip-hour price and the token price, driven by a steady software trend and by sudden algorithmic jumps. The paper proves that wedge cannot be hedged with chip-hour futures. The best hedge explains about 3 percent of its variation, because the jumps arrive through a channel the chip-hour market does not contain. A refiner is a long position in that wedge. So is its lender.

The token price is the revenue, and it deflates. The refiner sells at whatever a token fetches, and the trend of that price is down: the paper models it as a deterministic decline plus downward jumps, and the market has delivered both. A refiner’s edge is relative. It earns the spread only while its silicon converts chip-hours to tokens more cheaply than the alternatives, and the alternatives are moving: graphics-chip software, the hyperscalers’ own inference chips, and the next decode vendor. The stress is not a token price fall as such, since the refiner’s own cost per token falls with its efficiency. The stress is the wedge closing from the other side.

The silicon is the collateral, and it has no resale market. This is where the refiner departs from every chip loan in the rated record. A used H100 has a price: American Compute’s bands put it at 45 to 74 percent of list in 2026. A used SambaNova SN50 has no published price at all, because the first units ship in the second half of 2026 and only one deployment partner is named. The lender’s backstop is silicon whose value depends on the vendor’s survival and on model architectures not moving away from what the chip was built to run. The vendors’ own standing tells the story. SambaNova raised $1 billion in July at an $11 billion valuation, seven months after acquisition talks with Intel that valued it near $1.6 billion. Cerebras completed the largest semiconductor listing on record, raising $6.4 billion, on first-quarter revenue of $193 million and a customer list it describes as concentrated in OpenAI, G42, a university and AWS. A sevenfold swing in a vendor’s value inside a year is not a criticism of the vendor. It is the width of the range within which a refiner’s collateral sits.

Pooling across makers is the refiner’s structural improvement. A graphics-chip pool cannot diversify Nvidia’s product cycle; every lender’s collateral falls on the same annual clock. A fleet spread across SambaNova, Cerebras, AMD and Nvidia breaks that clock. The cost is the paragraph above: the diversification is bought with silicon that cannot be sold.

What the refiner has going for it. A latency premium: the customers it names are coding agents and voice, where tokens per second is the product. Siting: a 20-kilowatt air-cooled rack fits in ordinary colocation, which is faster to power than a 120-kilowatt liquid-cooled one, and time to power is the binding constraint in this market. Supply protection: the $300 million of price-protected agreements is a hedge on the input cost that no graphics-chip buyer has. And a facility that draws with demand rather than ahead of it, which is the right shape for a business whose collateral cannot be resold.

Where the refiner is exposed. A fixed cost base of space, power and debt service against revenue that is partly per-token. Vendor concentration until the second decode vendor is live. Architecture risk on fixed-function silicon. And the one that applies to every seat: a demand slowdown, in which per-token revenue falls first and the dedicated-rack customers decide whether to renew.

A merchant’s margin is a spread between two prices for the same thing at two tenors. It commits to a block of capacity for months and sells it by the hour, the day or the week. The commitment is take-or-pay: the block is paid for whether or not it is resold. The sale is by use. Between the two sits fill, and fill is the merchant’s entire risk.

What the merchant is long. Compute, in the plainest sense. A merchant holding a three-month block at a fixed price gains when retail rates rise and loses when they fall. It is the one seat in the market whose exposure is a clean long in the rental rate, which is why it is the natural hedger of the listed contracts, and the subject of the next section.

What the merchant is short. Every hour it does not sell. The arithmetic is unforgiving because the inventory perishes. Our estimate for a block bought near $2.50 an hour and resold near $3.40 is a breakeven fill near 74 percent. At 70 percent fill the trade loses even if retail prices hold. At 50 percent it loses badly. No futures contract pays on an empty rack, and no insurer will write a floor on fill without an independent meter on the hours actually used. Fill risk dominates price risk, and it is unhedgeable by anyone but the merchant’s own sales desk. The ledger below is that arithmetic with the sliders exposed, so you can disagree with the base case.

Delivery is the merchant’s credit risk. The block is only as good as the operator that delivers it. Most commitments in this market are non-assignable, prepaid in part, and sit behind the operator’s own lenders. If the operator fails at month six, the merchant’s take-or-pay stops, its prepayment is gone, and its resale customers are still owed hours. That is the one scenario in which a merchant with a hedge can end up better off than one without, because the hedge pays on the full term while the commitment did not, and it is also the scenario in which a lender against the commitment recovers least.

The exchange is the merchant’s exit. What separates today’s merchant from the 2023 version is that unsold hours can be sold rather than eaten. SF Compute lets a buyer with unused time place a sell order and take credits back less a fee. Compute Exchange runs a secondary market. That converts a total loss into a resale-price risk, which a market can carry. It does not remove the residual; it prices it, and it prices it lowest at the moment the merchant most needs to sell.

What the merchant has going for it. No hardware, so no obsolescence and no fixed cost base beyond its commitments. It scales with demand and shrinks with it. Its book sits on the neocloud and marketplace tier that the index samples, so its basis to the listed contracts is the smallest of any seat. And it is the seat that inference favors: thousands of small buyers on short, usage-based contracts is a retail book, and someone has to hold the wholesale side of it.

Where the merchant is exposed. Retail rates falling against fixed wholesale commitments. Usage falling with them, because the two move together in a slowdown. An operator failing mid-term. Scarcity spikes on the other side, if the merchant has sold fixed-price retail forward and must cover in a spot market that reached $8 to $12 an hour for an H100 in 2023 and 2024. And the paradox: the transparency that gives it a hedge takes away the spread the hedge was protecting.

The basis to the listed contracts runs the opposite way for the two seats. The merchant’s book is the closest thing in the market to the index, which makes it the natural seller and the likeliest source of early liquidity. The refiner cannot use the contracts except through three bases at once.

What the contracts are. CME’s filing with the commission specifies each contract as 730 GPU-hours, a month of one chip, priced in dollars per GPU-hour with a one-cent tick worth $7.30. They settle in cash on the arithmetic average of the Silicon Data on-demand settlement price for each business day of the contract month, and list 36 consecutive months. The H100 index behind them draws on 50 to 100 platforms across hyperscalers, neoclouds and marketplaces in 40 to 50 countries, from about 150,000 daily pricing records, normalized to a standard configuration. Since December 2025 the hyperscalers’ on-demand rates have been split into their own benchmark, so the headline index leans toward the neocloud and marketplace tier, which is the tier the merchant buys from.

What the delay is. The contracts were to take effect on October 5. On September 21 the commission extended its review to November 9, reported by GuruFocus, to study how the contracts operate and to seek industry input. The date is the statutory ceiling rather than a count from the announcement: under CFTC Regulation 40.3(a) a product filing gets a 45-day review, extendable once by up to 45 more where the contract raises novel or complex issues. NYMEX filed on August 11, so the initial period ran to September 25 and the maximum extension reaches November 9. PANews, summarizing The Information, reports the regulator’s description of the underlying market as fragmented and opaque, and CME’s expectation that approval will not arrive by early October. The commission’s earlier request for comment asked about the size and liquidity of the cash market, manipulation, and customer protection. The letter itself is not public as of this writing, and the figures above are as reported.

Who can use them. The answer sorts by seat.

The paradox. The merchant’s spread exists because the same chip trades across a wide range. The Federal Reserve paper documents a roughly fourfold spread on the H100 between the marketplace floor and hyperscaler on-demand, and Compute Exchange quotes reserved at 40 to 70 percent below on-demand and spot at 60 to 90 percent below. A merchant earns by buying at one point in that range and selling at another. A trusted daily index and a liquid futures curve exist to collapse exactly that range into one number. So the event the merchant needs, a curve it can hedge on, is the event that thins the spread it was hedging. The hedge and the squeeze arrive together.

The paradox has a regulatory face. The opacity the commission cited as its reason to extend the review is the same opacity that makes the merchant’s business: the regulator’s reason for delay is the merchant’s reason for existing. A market opaque enough to pay a merchant is a market whose index a regulator will question. As the index earns trust, the merchant’s margin migrates to the exchange, which takes a fee on volume, and to the refiner, whose spread was never in the chip-hour price at all.

Each scenario hits the two pooling seats through a different channel, and for each the question that matters to a lender is whether a hedge exists.

Two patterns stand out. The merchant’s largest risk, a falling rental rate, is the one risk the market is about to list a hedge for. The refiner’s largest risks, the wedge and the architecture, have no hedge anywhere and no prospect of one soon. And the one scenario that hits both through the same channel, a demand slowdown, is the one in which the exchange bid, the merchant’s only exit, is thinnest. That is the Federal Reserve paper’s wrong-way result restated for the two seats: recovery on what is held is lowest in the state that forces its sale.

The merchant is an old business with a new product. An electricity retailer buys power wholesale, on contracts of months or years, and sells it to households by the kilowatt-hour at a price it has promised. The product cannot be stored. The retailer holds the gap. When the two legs move apart, the retailer is the first-loss piece of the whole market, and twice in 2021 it was.

Britain. Wholesale gas rose nearly sixfold between February and December 2021 against a retail price cap that reset twice a year. The National Audit Office counts 28 suppliers failing from June 2021, with 2.4 million customers moved to other suppliers under the supplier-of-last-resort process and a further 1.6 million at Bulb, which went into special administration. The cost to consumers was put at £2.7 billion for the 28, about £94 on every household’s bill, plus £0.9 billion of government spending on Bulb in 2021–22 and £1.0 billion budgeted for the year after. The auditors’ finding was that the regulator’s approach to licensing and monitoring suppliers increased the risk and cost of their failing: a low bar to entry, financial-resilience rules that did not arrive until 2021 despite tightening that began in 2018, and no stress test of how the price cap and the last-resort process would interact. The suppliers that failed were, almost without exception, the ones that had not hedged their wholesale purchases against their retail promises.

Texas. Griddy sold households the wholesale price itself, passed through in real time, for a flat monthly fee. In the February 2021 storm the grid operator set the price at $9,000 per megawatt-hour and, in the company’s account, held it there after load shedding had stopped. Customers paid about 300 times the normal rate. Griddy filed for bankruptcy on March 15, 2021. The customers who had signed up for the cheapest possible price discovered they had signed up to be the merchant.

The translation to compute. The sign flips. The compute merchant’s danger is retail falling against fixed wholesale, a slow bleed rather than a spike, and the paper’s spike regime supplies the other direction for any merchant who has sold fixed-price retail forward. The mechanism is the same: a committed leg and a floating leg, a perishable product between them, and a firm whose capital is small relative to the gap it holds. The remedy that Britain reached after the fact is the remedy a compute lender can write before it: a hedging requirement as a condition of the licence, or here, of the loan. Oil and gas lenders have done this for decades. A reserve-based loan requires the borrower to hedge half or more of its production for two or three years, and values the hedged volumes at the hedge price in the borrowing base. The hedge is what lets the bank lend more. A compute warehouse line built the same way would advance more against hedged blocks, and it would create the two-sided flow the listed contracts need to exist.

A lender to either seat is lending against a pool, and the first question is which part of the pool is really collateral. The answer is the tier mix. In a securitization, seniority decides who is paid first when cash falls short. In a compute pool, the contract tier decides who is served first when capacity falls short, and who pays whether or not they use it. The dedicated-rack customer on a multi-year term is senior. The reserved customer with burst rights is in the middle. The pay-per-token or spot user is the equity, and is rationed first. The share of revenue in the bottom tier is the share of the pool that absorbs the first loss.

A stylized borrowing base follows. The figures are our illustration of the shape, not a market quotation.

The two seats differ in what sits under the base. The refiner’s backstop is its chips, and for purpose-built inference silicon that backstop has no price. A lender to a refiner is lending against contracts and against the supply protection, with the silicon as a claim on a vendor’s roadmap. The merchant’s backstop is its commitments, which are worth their in-the-money amount and nothing more: a commitment at $2.50 when retail is $2.00 is a liability. A lender to a merchant is lending against resale contracts and against the hedge, which is why the reserve-based structure fits: hedged hours valued at the hedge price, unhedged hours haircut to a cautious price deck, and a redetermination on a schedule.

November 9, and what changes in the contract or the index. If the commission’s questions were about the cash market’s depth and the index’s resistance to manipulation, the answer may come as a revised index rather than a revised contract. Either way, the reference price the merchant hedges on is the thing under review.

Open interest beyond three months, and who is short. The merchant is the natural seller. If the first open interest is short-dated and sits with market makers rather than with holders of wholesale blocks, the contract is not yet doing the job the merchant needs.

The first merchant that hedges in public, or the first loan that requires it. A warehouse line against hedged blocks, with the hedge valued in the base, is the compute version of the reserve-based loan and the point at which the aggregator paradox becomes a business model rather than a trap.

Exchange volumes. SF Compute and Compute Exchange do not publish them. The first published figure for resale of unused commitments is the first measure of how deep the merchant’s exit is, and how deep it stays when rates fall.

General Compute’s utilization and draws. The facility draws as demand arrives. The draw schedule, if disclosed, is the best public reading of whether the refiner’s dedicated and reserved tiers are filling, or its per-token tier is.

SambaNova’s SN50 shipments in the second half of 2026, and its next round. The refiner’s collateral is this vendor’s silicon. Delivery on schedule and a round at or above July’s valuation would say the collateral has a future. Slippage or a down round would say the opposite, and there is no resale market to say otherwise.

Cerebras’s customer concentration after its listing. The second decode vendor’s public filings will show whether a purpose-built inference chip has a market beyond one or two buyers, which is the question a refiner’s lender is really asking about every decode vendor.

The first token-price curve. The refiner’s hedge does not exist until a token price by model grade is published and traded. The paper’s prediction is that it will decouple from the chip-hour curve when it does.

The first resale print on a purpose-built inference chip. One public sale of used purpose-built inference silicon, at any price, would turn the refiner’s collateral from an assumption into a number.

The sign the paradox is biting. Merchant margins narrowing while exchange volumes rise. That is the spread migrating to the fee, and it is what a working index looks like from the merchant’s side.