Silicon Data’s rental price index for Nvidia’s B300 stood at 151.9 on 22 October 2026, up 51.9% since 31 May, and its LLM Token Index stood at $0.94 per million tokens on 21 October, according to a chart Silicon Data posted on X on 23 October 2026. The same chart shows the B200 index up 16.9% and the H100 index up 5.5% over the same period. The rental indices cover non-hyperscaler providers.
DRW founder and chief executive Don Wilson has a name for the margin between those two prices. On Bloomberg’s Odd Lots podcast, in an episode dated 19 October 2026, Wilson compared renting GPUs and selling tokens to an oil refinery and a soybean crush plant: “So in oil, we call it a crack spread. and soybeans, we call it a crush spread.” He added: “At DRW, we call this the crunch spread.” Wilson said token futures are at “the thought stage.”
Business Pill · MARGIN SQUEEZE
A short explainer of a margin squeeze: when what you pay rises and what you can charge falls. It teaches the general idea only and says nothing about Silicon Data, DRW or the crunch spread.
The key insight: As we read it, Wilson’s crunch spread names a margin that sits between two prices Silicon Data tracks: on its chart the B300 rental index rose 51.9% from 31 May to 22 October, while its LLM Token Index read $0.94 per million tokens on 21 October. Baker’s reply adds that token volume per GPU sits inside the same margin.
Wilson described the trade on the show. If you rent GPUs, “you pay the on-demand price. And then you load up an open source model and you start selling tokens through open router or something.” The question, he said, is “can you do that profitably or not?” and “that profitmargin that you can generate is akin to, you know, what an oil refinery does, right?”
Wilson said “you’ll be able to trade futures contracts on the output. on the tokens themselves,” and that “over time, that’s going to be possible.” Asked whether DRW is actively working on token futures, Wilson answered “Yes” and called it “the thought stage.” Quotations are from a machine transcript of the episode on Podscripts.
On 23 October Silicon Data posted its chart with the line “The crunch spread continues to widen,” crediting Wilson for the term. Wilson quoted that post the same day: “GPU prices higher and token prices lower means the crunch is compressing. That means lower profit margins on the GPU rental to token production leg of the trade.” He added: “Of course the profitmargin on buying GPUs and renting them out is higher than ever (not exactly an indication of a slowdown in inference demand).”
Silicon Data’s GPU rental indices for Nvidia’s B300, B200 and H100, rebased to 100 on 31 May 2026, rose 51.9%, 16.9% and 5.5% to 22 October 2026, while its LLM Token Index read $0.94 per million tokens on 21 October (Silicon Data chart posted on X, 23 Oct 2026; rental indices cover non-hyperscaler providers).
What Was Posted on X
This is a really good concept for folks to pick up on that I am learning about real-time. I liked @drwconvexity’s “crunch spread” term that applies familiar commodity economics to inference markets (he spoke about this on Odd Lots live ep recently:
The processor’s margin depends on input cost, output price and how much saleable output each unit of input produces.
@GavinSBaker’s point adds that conversion yield. Per GPU-hour: Crunch spread = (price per million tokens × million tokens sold) – GPU rental cost. Illustratively, 2M tokens sold at $2/M generates $4 revenue. Subtract $3 of GPU rent and the spread is $1, before other costs. NOTE: In reality, it’s the blended price across models and input/output/cached tokens × tokens sold per GPU-hour, less GPU rent and any modelrevenue share, but let’s take it one step at a time (see: @deedydas’ recent post).
Halve token prices and double tokens sold per GPU-hour: revenue stays $4. Higher GPU rents would then require further throughput gains to preserve the spread. Software optimization, batching and utilization all affect that yield for a given model and service standard.
Matching GPU and token contracts could eventually let inference providers hedge that processing margin, much as refiners and soybean processors hedge theirs.
Shanu Mathew (@ShanuMathew93) posted this on 25 October 2026. It sets out the crunch spread as revenue from tokens sold per GPU-hour minus GPU rent, with an illustration of $4 revenue, $3 rent and a $1 spread. The piece uses it for the worked example and the hedging idea.
GPU prices higher and token prices lower means the crunch is compressing. That means lower profit margins on the GPU rental to token production leg of the trade. Of course the profitmargin on buying GPUs and renting them out is higher than ever (not exactly an indication of a slowdown in inference demand).
Don Wilson (@drwconvexity), founder and CEO of DRW, posted this on 23 October 2026, quoting Silicon Data’s chart. It says higher GPU prices and lower token prices mean the crunch is compressing. The piece quotes it in What Wilson Said.
Don, there is a missing term in your equation.
The crunch spread is not: Token Price – GPU rental price.
Rather it is: (Token Price * Token Volume) – GPU rental price.
More simply: Revenue per GPU – GPU rental price.
Continuous throughput improvements for the same exact token mean that volume growth can more than offset token price declines for the exact same model/token
i.e. Kimi K3 token throughput per GPU has recently been improving at a 37% annualized rate driven by inference engine optimizations.
GPU prices are vertical because:
1) Mix is shifting from higher margin frontier tokens to lower margin open-weight tokens. $10 in Claude revenue requires $1 of inference compute. $10 of of open-weight revenue requires $3 to $7 of inference compute given competition between inference providers leads to lower margins. Open-weight tokens taking share is positive for AI infrastructure demand given lower margins at the model layer. The reverse is also true in that frontier tokens taking share is negative for AI infrastructure demand. Open-weight models have been taking share since early summer.
2) Throughput is steadily improving for each individual token type such that GPU compute becomes more valuable even as true like for like token prices decline.
Gavin Baker (@GavinSBaker) replied to Wilson on 24 October 2026, adding token volume to the equation. The piece quotes it in The Missing Term.
Refineries in the US sent out 5.9% more volume in products last year than they took in as crude oil. A 42-gallon barrel came out as 44 and a half gallons. The products weigh as much as the crude and are less dense, and the same weight fills more gallons.
Crude oil comes out of the ground as a mix of molecules of many sizes. The distillation tower sorts them by boiling point. The lightest rise to the top as gasoline and propane. The middle ones come off as jet fuel and diesel. The heaviest settle at the bottom as fuel oil and the residue that becomes asphalt. The heavy fractions then go to the conversion units. Heat, pressure, catalysts and hydrogen crack the big molecules into small ones there, and the small ones become gasoline.
US refineries turned out 45.9% gasoline, 30.0% distillate and 11.0% jet fuel last year, with coke, propane, asphalt and lubricants in the rest. Jet fuel’s share was 6.9% in 2020.
@ChurchillWw says US refineries sent out 5.9% more volume in products last year than they took in as crude oil, and turned out 45.9% gasoline, 30.0% distillate and 11.0% jet fuel.
The Missing Term: Token Volume
Gavin Baker, whose X profile lists him as managing partner and CIO of Atreides Management, replied to Wilson on 24 October: “Don, there is a missing term in your equation.” Baker wrote that the crunch spread is token price times token volume, minus the GPU rental price: “More simply: Revenue per GPU – GPU rental price.”
Baker wrote that “Kimi K3 token throughput per GPU has recently been improving at a 37% annualized rate driven by inference engine optimizations.” He also wrote that “$10 in Claude revenue requires $1 of inference compute,” while open-weight revenue of $10 requires “$3 to $7 of inference compute given competition between inference providers.” These are Baker’s figures. Wilson answered: “you are absolutely right that token volume is an essential part of the equation.”
Shanu Mathew, who passed the thread on in a post on X on 25 October, set out an illustration: “2M tokens sold at $2/M generates $4 revenue. Subtract $3 of GPU rent and the spread is $1, before other costs.” In the same post: “Halve token prices and double tokens sold per GPU-hour: revenue stays $4.” Mathew wrote that matching GPU and token contracts “could eventually let inference providers hedge that processing margin, much as refiners and soybean processors hedge theirs.”
Spread
Input
Output
Source
Crack spread
Crude oil
Gasoline, diesel
Wilson on Odd Lots; Mathew on X
Crush spread
Soybeans
Oil, meal
Wilson on Odd Lots; Mathew on X
Crunch spread
GPU-hours
Tokens
Wilson on Odd Lots; Mathew on X
Sources: Odd Lots episode dated 19 Oct 2026 (Podscripts transcript); Shanu Mathew post on X, 25 Oct 2026.
Line, per GPU-hour
Illustration
Status
Tokens sold
2 million at $2 per million
Mathew’s illustration
Revenue
$4
Mathew’s illustration
GPU rent
$3
Mathew’s illustration
Spread before other costs
$1
Mathew’s illustration
Source: Shanu Mathew post on X, 25 Oct 2026. Illustrative figures, not market prices.
Where It Meets Our Coverage
Our DeepSeek piece reported DeepSeek’s figure of US$38 per GPU-hour for one cluster at off-peak prices, published on 20 October, and a median posted on-demand rate of $7.98 an hour for an Nvidia B300 in August, per AIMultiple’s index as reported by Akash Network. Our H100 piece reported the H100 at $2.88 per GPU-hour on Silicon Data’s index on 24 October. Our rent-and-token piece set those rents beside a token-price decline of about 50% year on year measured by YipitData. The figures come from different indices, chips and windows, so this piece does not combine them.
The Structural Read
As we read it, Wilson describes a processing margin: rent GPUs at “the on-demand price,” load an open-source model and sell tokens, the way a refinery turns crude into fuel.
As we read it, Baker’s point changes what counts: on his figures, throughput gains such as the 37% annualized rate he cites for Kimi K3 mean revenue per GPU can rise even when token prices fall.
As we read it, the hedge is still an idea: Wilson called token futures “the thought stage,” and Mathew wrote that matching contracts “could eventually” let providers hedge the margin.
Don Wilson, Odd Lots, episode dated 19 October 2026
“At DRW, we call this the crunch spread”
Three Implications
FOR INFERENCE PROVIDERS Wilson said renting GPUs and selling tokens comes down to whether “you do that profitably or not”; Baker wrote that the margin is revenue per GPU minus the GPU rental price.
FOR GPU OWNERS Wilson wrote that “the profitmargin on buying GPUs and renting them out is higher than ever,” on 23 October.
FOR TRADERS Wilson said on Odd Lots that “you’ll be able to trade futures contracts on the output. on the tokens themselves,” and called DRW’s work on it “the thought stage.”
The framework puts it this way: the token “is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”
As we read it, Wilson’s crunch spread is the gap between the second and third of those roles: the compute rented by the hour and the token sold by the million. The framework was published in May 2026 and does not discuss Wilson’s term.
What Is Not Established
No crunch spread contract trades today in the sources we read; Wilson called token futures “the thought stage.” The token index is expenditure-weighted across models, while the rental indices are per chip, so the chart shows two prices, not a measured spread. Baker’s Claude and open-weight ratios and his Kimi K3 throughput figure are his own. Mathew’s $4, $3 and $1 are an illustration. The Odd Lots quotations come from a machine transcript, and we did not listen to the full episode. We did not contact Wilson, DRW, Baker, Mathew or Silicon Data.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Silicon Data’s B300 rental index rose 51.9% from 31 May to 22 October, and its LLM Token Index read $0.94 per million tokens on 21 October. Wilson calls the margin between rented GPU-hours and the tokens sold from them the crunch spread, and said on Odd Lots that DRW is at “the thought stage” on token futures. Baker’s reply adds that token volume per GPU belongs in the equation.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece quotes Don Wilson from a machine transcript of the Odd Lots episode dated 19 October 2026 on Podscripts; we did not listen to the full episode. The X posts by Silicon Data, Don Wilson, Gavin Baker and Shanu Mathew were read on X on 25 October 2026, and the index figures are as printed on Silicon Data’s chart. We did not contact the people or companies named.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
