The Compute Crack Spread

Our companion piece on the October listing went looking for the balance sheet that carries a floating exposure to a GPU-hour and found that every commercial participant with a real exposure prices it fixed: the neocloud sells two-to-five-year take-or-pay, the lender sizes term debt against that contracted revenue and excludes the merchant tail, the buyer signs the same contract from the other side. No lease resets to Silicon Data, no covenant references Ornn, no ABS trigger keys to an OCPI print. The listing therefore runs the benchmark film backwards: oil and gas got deep because the physical market floated off a screen first — PEMEX formula pricing in 1986, Henry Hub spot liquidity before the 1990 contract — and compute is being asked to list first and hope the physical market reprices to it.

There is, however, a part of the compute complex that already runs the film forwards. The token market is unbundled, transaction-based, continuously repriced and disclosed at the endpoint level. OpenRouter routes on the order of 25 trillion tokens a week across 400-plus models and roughly a hundred providers, publishes a per-endpoint order book for every model — price, context, quantization, uptime — routes stochastically with weights proportional to the inverse square of price, and reprices in hours (we mapped that microstructure in August). That is a cash market in exactly the sense the contract-success literature requires: Brorsen and Fofana found an active cash market “perfectly predicts” whether a commodity gets a futures market.

So the administrators are indexing the leg that does not float and ignoring the leg that does. This is not a complaint about SKU-specific on-demand rent as a settlement object — it is the cleanest thing to settle on, and we have said so. It is a question about what the index is for. A take-or-pay buyer has no exposure to the on-demand rate. It has a fixed cost and a floating output. Its renewal price is negotiated off its own token economics; its lender’s exposure is to whether those token economics cover the fixed rent — the contracted project converts technology deflation into tenant credit, and tenant credit for an inference tenant is its token margin. The variable the whole physical market actually faces is the spread between the fixed compute price it signed and the floating token price it sells into. The rent index is a proxy for that spread, and the numbers below say it is a poor one.

A refiner does not hedge crude and hedge gasoline; it hedges the crack — the margin of converting one into the other at reference yields, the 3:2:1 being the convention. A merchant generator does not hedge power and hedge gas; it hedges the spark spread, power minus gas times heat rate, and the market heat rate (power over gas) tells it which units run. The compute version was set out in our spark-spread work as a three-leg chain with two heat rates: power to GPU-hour through a rate fixed by physics, GPU-hour to token through a rate fixed by nothing but utilization. Power is 2.7% of the GPU-hour price. The interesting conversion is the second one.

Three inputs, three sources, and each is already published somewhere. The token price comes from the endpoint book — not an expenditure-weighted aggregate like Silicon Data’s SDLLMTK, which by its author’s own description is a usage-weighted average “irrespective of models” and falls when users switch models rather than when prices fall, but a grade-standardised price for a named model class. The reference throughput comes from a published benchmark sweep, held constant across venues so the series isolates price. The rent comes from the index the futures settle on. For the worked series we use Llama-3.3-70B-Instruct, because it is the one model class for which a deposited, reproducible throughput sweep exists on H100s (Llama-3.1-70B, vLLM, 1,024 runs: 2.88M total tokens per GPU-hour at 10 requests per second, 9.87M at 50, a median of 11.16M saturated), and because thirteen endpoints served it on the day of the pull.

Read the figure as a merit order. Two thirds of the active book prices tokens at a market heat rate of 3.5–3.6 million tokens per GPU-hour — SambaNova, Groq, CoreWeave, Google and their neighbours cluster within a few percent of one another at $0.71–0.72 per million blended — which a saturated H100 covers three times over. Cloudflare at $1.43 and Together at $1.04 sit at the top of the book with heat rates of 1.8 and 2.4 million, the kind of price that clears even at 10 requests per second and that survives, on OpenRouter’s routing rule, only because buyers direct orders to them. At the bottom, AkashML and Parasail need 6.6 million; Novita, the cheapest active endpoint, needs 8.8 million; and the two degraded endpoints price at 11.1 and 8.8 million — DeepInfra at $0.227 blended sits almost exactly on the saturated median throughput of the sweep. The bottom of the book is priced at the marginal cost of a saturated H100, and the venues there are the ones the router is currently flagging for uptime. That is what a market clearing at its own heat rate looks like.

Three features of the series are the findings. First, the crack is large and positive for almost everyone at saturation — a median +$5.37 per GPU-hour against $2.53 of rent — and negative for almost everyone at 10 requests per second. The sign of a host’s P&L is set by utilization, not by price, which is the conclusion our spark-spread work reached from the cost side and is now visible from the revenue side. Second, the dispersion is five-fold across active venues for an identical declared good on one screen, a spread Stigler’s search-cost model cannot explain; the residual is unverified quality (fp8 versus bf16 versus “unknown”), capacity, and deliberately non-best-execution routing. A crack-spread series has to carry a grade, or it carries that dispersion as noise. Third — and this is the point of the exercise — the series moves with the hedger. Rent has been nearly still: seven-day changes under 1% across Silicon Data’s SKUs, the hyperscaler series unchanged on 39 of 50 trading days. The token leg repriced 5× in a day when OpenAI cut to its Luna tier against a disclosed 1.44× efficiency gain, DeepSeek cut V4-Pro 75%, Tencent raised Hunyuan up to 400%. A hedger who locks rent on GPU1 has hedged the quiet leg.

The crack is a difference of two deflating series, and their deflation rates are not the same. Rent deflates on the technology curve — ln 2 over a 2.5-year performance-per-dollar doubling, about 27.7% a year, the assumption we have used throughout. Token prices for a fixed model class deflate faster and more erratically, because they carry the technology curve and the host-layer margin compression that a 1/p² router enforces. Holding the reference throughput fixed — a constant-grade series — the surface below shows the saturated crack one year ahead.

Who is short this spread — who loses when token prices fall faster than rent? Any balance sheet that has bought compute forward at a fixed price and sells tokens at a floating one: the inference host with a take-or-pay, the lab that signed a five-year cluster lease and prices its API against competitors, and, one step removed, the lender to either. The companion piece ranked participants by who needs a hedge and found no natural two-sided flat-price demand. The crack is where the two-sided demand is. The host is short the crack by construction. The neocloud that sells take-or-pay is long it — it has sold rent forward to someone whose ability to pay depends on the crack — and its lender is long it through tenant credit. A crack-spread instrument has a buyer and a seller on day one. A rent instrument has sellers and speculators.

The August 19 request for comment is usually read as a threat to the October listing. It is more useful read as a specification. The Commission is not asking whether compute should have a derivatives market — Chairman Pham’s cover line is that “America cannot win the AI race without a robust derivatives market for compute.” It is asking what a cash market has to look like before a contract can settle on it, and its questions are the same ones the contract-design literature has asked since Silber. Set them against the two candidate legs.

Which leaves an awkward implication for the comment file. The Commission has asked whether any cash price series satisfies its Appendix C criteria for a settlement object. On the rent side the honest answer is not yet, and not observably — the cash-market exhibit supporting the CME filing was submitted under separate cover with confidential treatment requested, so no commenter outside the exchange can evaluate it. On the token side the answer is the data exists and is public, but nobody has constructed the series. Those are very different failures, and only one of them is closed by anybody writing a comment letter.

The series above is a demonstration, not a benchmark. For it to settle anything it needs the four things the rent indices have been criticised for lacking, and two more that are specific to tokens.

The adoption argument in the companion piece was that a benchmark gets used when either a dealer warehouses bilateral exposure and lays the residual off in futures (the WTI route) or the physical market reprices to the index so end users have something to hedge (the Henry Hub route). The second route looked closed for compute, because no operator will re-paper a five-year take-or-pay to float on Silicon Data — the contract exists precisely to remove that exposure, and the lender sized the loan against it.

But the physical market does not have to reprice on the rent side. It can reprice on the token side, and it already has: every inference host’s revenue floats on token prices today, involuntarily, through the router. The step from “my revenue floats on tokens” to “my compute cost has a component indexed to tokens” is small, and it is the step a toll takes.

From there the dealer route and the cash-market route converge. A merchant writing token-indexed tolls for hosts is long the crack; it lays the rent leg off on GPU1/GPU2 — the futures finally have a commercial short with a reason to be there — and carries the token leg until a token index is listable, warehousing it the way the gas marketers warehoused basis after Order 636. The rent futures get their dealer, the token index gets its physical reference, and the utilization that no administrator can see becomes the thing the toll’s conversion fee is paid on — which means someone has to meter it. Attested, independent metering, published under a methodology, is the other half of the product, and it is a business in its own right.

Three things. If the rent indices begin to move with the token market — if SDH100RT’s weekly changes start tracking the median endpoint price for the SKU’s reference model class — then the rent leg is carrying the information after all and the crack is redundant. If a major operator re-papers a term contract to float on a rent index before any host signs a token-indexed toll, adoption has arrived through the rent side and this piece has the order wrong. And if a disclosed-constituent, mix-held-constant token index with a published grade appears from an existing administrator, the “nobody has built it” claim is dead, and the remaining question is only who settles on it first. We would regard any of the three as good news for the market, and would say so.

• Token book. OpenRouter public endpoints API for meta-llama/llama-3.3-70b-instruct, retrieved August 21, 2026. Prices, quantization and 30-minute uptime as published per endpoint.
• Rent. Silicon Data H100 Rental Price Index product page (SDH100RT $2.53), retrieved August 21, 2026.
• Throughput. NLR vLLM online-inference sweep, dataset deposited with arXiv:2604.07345 (inference_online_finite_llama3_70b, 1,024 runs, node of four GPUs, divided through).
• Contract and regulation. NYMEX Submission 26-370 and CME SER-9785 (Aug 11, 2026); CFTC request for comment on compute derivatives (Aug 19, 2026), 91 FR 54259, RIN 3038-AF77, comments close Oct 20.
• Literature. Brorsen & Fofana (2001) on cash-market preconditions; Silber (1981) and Carlton (1984) on contract success; Stigler (1961) on search costs; Bandi & Su, arXiv 2607.12156.
• Kinetic Alpha. The Contract Has a Date. It Needs a Dealer. (Aug 21) — the hedger map and adoption history this piece builds on. The Long End Has Two Sellers Now (Aug 21) — tenant credit as the transmission channel. The Other Side of the Spark Spread (Aug 21) — the three-leg chain and the NLR throughput figures. The Token Leg Has a Price Screen (Aug 4) — OpenRouter microstructure, SDLLMTK treatment, SIT grade construction. The Contract Got a Date and Compute Became Collateral (Aug 14). Silicon Data’s Product Suite (Aug 16). How the Indices Calculate Price.