Capacity Futures

Capacity futures sit one step back from inference tokens. The unit is not the output — it is the right to use a defined amount of compute hardware in a defined window. Buyers are AI labs, enterprise procurement teams, and increasingly anyone who needs to know their capex line for the year before they have committed it. Sellers are the people sitting on the data centres.

This framing is interesting precisely because it is boring. It does not require any innovation in grading or settlement that the world has not already done for natural gas, aluminium, or shipping capacity. The playbook is well understood: a forward curve emerges, brokers intermediate, exchanges eventually standardise the contract, hedging products appear for both sides. The only new thing is the underlying.

That underlying is harder than it looks. A “GPU-hour” of an H100 in Q3 2026 is not the same product as a GPU-hour of a B200 in Q3 2027, and the generational turnover is fast enough that a five-year forward curve is essentially fictional. The market either has to define contracts at the level of capability tiers (which collapses back into the grading problem) or accept that contracts past 18 months are exotic instruments rather than core liquidity.

The most likely path is that the existing commodity exchanges look at this and decide that it does not yet hit their volume thresholds, and the market grows up inside the broker layer first — much as electricity spot markets grew up inside ISOs before becoming exchange-traded. The research question is when, and under what regulatory pressure, capacity futures graduate from bilateral broker deals into a real cleared market. The signals to watch are (a) the first hyperscaler offering a genuine forward contract rather than a discount on rate-card pricing, and (b) the first non-AI commodity exchange listing a compute product. Neither has happened yet.