Intercommodity Spreads, From Crack to Compute

As the American Innovation Exchange (AI Exchange) prepares to list futures on GPU rental prices, spreads between compute and related commodities will become tradeable for the first time. Below, we look at what an intercommodity spread is, how the concept has played out in traditional commodity markets, and why compute is the natural next link in that chain.

An Intercommodity Spread

An intercommodity spread expresses the relationship between two related commodities, rather than trading either one outright. Instead of betting on the direction of a single input or output, a trader takes a position on the spread between them.

For example, a crack spread is the pricing difference between crude oil and the refined petroleum products made from it, like gasoline and diesel. It's essentially a proxy for a refinery's gross margin: refiners buy crude and "crack" it into products, so the spread reflects what they earn for that processing.

A crush spread is the soybean-processing equivalent of a crack spread. It is the price difference between soybeans and the products crushed from them, soybean meal and soybean oil.

Finally, a spark spread is the power-sector version: the difference between the price of electricity and the cost of the natural gas needed to generate it. It approximates the gross margin of a gas-fired power plant.

Each spread lets a market participant trade the conversion itself, not just the raw inputs or the finished outputs. In every case, the spread strips out the volatility common to both legs and isolates the difference that actually matters to the business converting one commodity into another.

A Compute Spread

Compute is the obvious next link in that chain. A datacenter converts megawatt-hours into GPU-hours, and GPU-hours into tokens. These are, functionally, two distinct production processes stacked on top of each other, yet neither has had a direct hedging instrument until now. With GPU rental price futures set to list on the AI Exchange, the pieces are finally coming together to make both of these conversions tradeable.

That gives rise to two natural spread constructs:

• Compute spark spread: power in, GPU-hours out
• Inference spread: GPU-hours in, tokens out

The compute spark spread almost mirrors the traditional spark spread, just with GPU-hours standing in for electricity. The inference spread goes a step further down the stack, capturing the economics of converting raw compute capacity into the tokens that actually get sold to end users.

Value Proposition

For datacenter operators and inference providers, these spreads offer for the first time a precise way to hedge the difference between what they pay for power or compute, and what they earn from renting out GPU-hours. That's the same value proposition crack and spark spreads have long provided to refiners and power generators.

Compute intercommodity spreads will also create new opportunities for trading firms acting as arbitrageurs. Early adopters may benefit from moving information between existing, more liquid commodity markets and the nascent compute curve. For example, power markets already have decades of price history and hedging infrastructure that compute futures markets don't yet.

As liquidity in GPU price derivatives builds, the relationship should start to run the other way. Compute is poised to act as a leading indicator in its own right. Intercommodity spreads between compute and its underlying inputs will become a valuable source of price discovery, not just for datacenter operators, but for anyone trying to understand where the economics of AI infrastructure are headed next.

Conclusion

Architect is preparing to launch a U.S. regulated marketplace for AI Compute. We're excited to bring the same trading strategies and time-tested infrastructure that has long supported traditional commodities and financial futures to this new asset class. We invite you to begin the onboarding and sign up if you are interested in having access to these products when launched.