The intelligence futures curve: GPU derivatives to predict the cost of compute

By TheSequence · Thu, 5 Nov 2026 12:03:22 +0000.

Issue #966 of The Sequence Opinion looks at GPU derivatives, an attempt to turn the disagreement between those who fear GPU rental prices will rise and those who fear they will fall into a market. The goal is to make compute costs and revenues predictable enough to support better investment decisions.

The essay gives an example: an AI company preparing a training run estimates it will need 100,000 GPU hours. It still has to secure the capacity, and rental prices could move before the code is ready. On the other side, a GPU operator has already bought the machines and needs future rental income to recoup its investment. One fears rising prices and the other falling prices, and that mismatch holds a latent financial market.

According to the newsletter's own summary, the piece explains how a rental-rate hedge on a GPU hour works, and why it locks in the price but does not reserve a cluster. It also describes how ORNN, Silicon Data, Liquid Compute, Architect and the proposed CME and ICE contracts are building the different pieces of this market. As of 4 November 2026, much of the US futures market remains under review.

The second part points out where the abstraction breaks down: different clusters sold under the same GPU name, basis risk, thin liquidity, fragile benchmark indices, shifting hardware and software demand, and margin calls that arrive before revenue does.

The author's view is that the biggest opportunity lies in financing and compute treasury software. The challenge, according to the piece, is to build an abstraction that hides enough complexity to be useful without hiding the details that determine whether it works.

The email received contains only the opening of the essay (the full version is on the Substack website), so this summary is limited to what appears in it.