Don Wilson on the DRW Crunch Spread and Token Futures

Key Takeaways

• DRW founder Don Wilson expects financial markets to introduce tradeable futures contracts on AI tokens alongside raw compute hours.
• Physical commodity processing models apply directly to AI inference, drawing lineage from oil refining and grain processing.
• Benchmarks like the Silicon Data Token Index are creating the price discovery required to settle cash or derivative contracts on compute output.
• Recent credit market scrutiny around neocloud providers like CoreWeave demonstrates why fixed multi-year GPU leases expose operators to severe market risk without financial hedges.
• Trading desks evaluate model monetization margins using the Crunch Spread.

The Crunch Spread

• Input Cost (Compute): Rent GPUs at current hourly on-demand compute market pricing. In Wilson's words: “you go and rent GPUs, you pay the on-demand price.”▶ 18:48
• Processing (Model Execution): Deploy an open-source model directly onto the rented GPU infrastructure: “load up an open-source model.”▶ 18:54
• Output Revenue (Tokens): Sell the generated output tokens via a marketplace: “start selling tokens through OpenRouter or something.”▶ 18:56
• Net Margin (The Spread): Calculate the difference between token sales revenue and raw compute rental costs, analogous to an oil refinery's crack spread or a soybean processor's crush spread. Wilson explains: “Well, can you do that profitably or not? Well, that profit margin that you can generate is akin to what an oil refinery does, and or what a soybean crush plant does... So in oil, we call it a crack spread, in soybeans, we call it a crush spread, so you heard it here first. At DRW, we call this the crunch spread.”▶ 18:59

When This Works (and When It Doesn't)

This framework applies when companies rent raw GPU compute capacity to host open-source language models and monetize the resulting inference tokens across third-party routing marketplaces. It treats compute as an industrial feedstock. If an operator can rent an H100 cluster at spot rates, serve an open-weight model, and route those tokens into aggregate exchange points, the operation functions exactly like an agricultural mill buying raw grain and selling processed meal.

It breaks down when software differentiation moves beyond raw token generation. Proprietary model owners with captive distribution do not sell raw commodity tokens; they sell enterprise workflows, proprietary data integration, and user interfaces where pricing disconnects from hourly compute costs. The spread also fails when latency requirements force dedicated hardware allocation that cannot clear spot volume, or when marketplace volume on aggregators dries up. As Wilson observes, trading output requires standardization: “The other thing is that eventually, and you know, we may be getting close, you'll be able to trade futures contracts on the output, on the tokens themselves.”▶ 18:12 Until those token indices mature into deep financial instruments, operators bear unhedged execution risk.

What to Do With This

If you run an AI product reliant on open-weight inference, stop viewing GPU bills as fixed overhead and start tracking them as cost of goods sold against spot token values. This week, audit your cluster utilization against external routing prices.

First, calculate your actual compute input cost: take the hourly rate you pay your cloud or neocloud provider per GPU hour, including networking overhead. Second, measure your model execution efficiency: tally how many generation tokens your cluster produces per dollar spent during peak serving windows. Third, check your real output revenue: compare your effective cost per million tokens against the prices posted on OpenRouter or benchmarked on the Silicon Data Token Index. If market aggregators are selling identical model output for less than your generation cost, shut down your dedicated instances and route through spot aggregators instead. If your margin is positive, you have established your operational crunch spread.