When AI compute gets a futures market

By TheSequence · October 1, 2026.

This opinion issue of TheSequence (Issue 943) proposes a central analogy: buying a warehouse full of GPUs is like buying an asset whose inventory evaporates every second. The chips remain physically present, but their unused capacity cannot be stored on a shelf to be sold tomorrow; an idle GPU hour today is an opportunity lost forever.

From that premise, the author describes the tension between two sides of the same market. On one hand, an AI company planning a training run three months in advance needs a specific cluster, in a specific window, at a price its budget can absorb. On the other hand, the provider of that capacity has already committed capital to the infrastructure before knowing the rates at which it will be able to rent it out in the future. The buyer needs the capacity before knowing its future costs; the seller has invested before knowing its future revenue. According to the text, this dynamic is very familiar territory in traditional commodity markets.

The article's thesis is that AI compute has a solid basis for becoming a genuine commodity, that is, a standardized and tradable good. But for that to happen, the author argues, three questions must be resolved: defining precisely what exactly is being traded, establishing how the quality of that capacity is measured, and designing contracts that properly distribute risk between buyer and seller.

The newsletter argues that the history of commodity markets offers a surprisingly applicable blueprint for this problem.

The body of the email received is an introductory excerpt from the full article and cuts off just as it begins to develop the section "How markets manufacture comparability," so it does not include the subsequent development of that idea or the author's conclusions. The rest of the analysis is available only in the original publication on Substack.