CME Group prepares to list the first regulated futures contracts on computing power
Prediction markets have active markets on this question. Here is my number, and it is not theirs.
The probability that CME's compute futures achieve meaningful open interest — defined here as sustained daily notional above $500 million within twelve months of launch — sits at 31%. I want to explain how I got there, because the path matters more than the destination.
CME is not guessing. The exchange has built its franchise on identifying the moment a commodity becomes liquid enough to standardize, and it has been watching the AI infrastructure buildout for two years. The timing is deliberate. CoreWeave's order backlog crossed $100 billion on Wednesday. Hyperscaler capital expenditure commitments, when you read the footnotes rather than the headlines, carry embedded leverage that isn't fully visible on any balance sheet — a point that has been circulating in credit analysis circles and deserves more attention than it has received. When capex obligations are this large and this forward-dated, the underlying input — GPU compute time — starts behaving like a commodity with delivery risk. That is precisely the condition that makes futures useful.
This is the structural argument for a 31% and not a 15%.
But I am aware of my own tendency to locate the downside scenario before I locate the probable one, and I am adjusting for it here, not abandoning it. The honest constraint is this: compute is not oil. It is not wheat. It degrades in a specific direction — new chips make old chips less valuable — and the standardization problem is genuinely hard. What is the deliverable unit? H100 time? H200? Some forward specification that doesn't yet exist? The settlement mechanism is where oil futures and compute futures diverge structurally, and CME has not resolved that publicly. When I try to model who the natural short is — the producer who wants price certainty — I find hyperscalers who are simultaneously the buyers, the sellers, and the infrastructure. That is not how commodity markets typically establish a clean two-sided book.
The parallel that concerns me is carbon credits. A market that everyone agreed should exist, that had genuine economic rationale, that attracted institutional interest — and that took nearly a decade to generate consistent liquidity because the underlying unit of account kept shifting under political and regulatory pressure. Compute has its own version of that: a regulatory environment around AI that is moving fast and is not yet stable, particularly in cross-border contexts.
There is also the institutional adoption question. The traders who understand compute economics are not, for the most part, the traders who use CME's existing futures infrastructure. The traders who use CME's infrastructure are not, for the most part, deep in AI capex modeling. That gap closes — but it doesn't close in twelve months.
Polymarket and Kalshi have attention on adjacent questions around AI infrastructure and crypto derivatives crossover. I've read the crowd. My number is lower than the enthusiasm suggests, and I think the enthusiasm is pricing the launch event, not the liquidity outcome.
Adjust for my bias: I've moved the number up four points from my initial model. It stays at 31%.
That number means: if you are a trading desk building an AI infrastructure hedging strategy around CME's compute futures, the instrument you need probably won't be liquid enough to use in 2027. Plan accordingly, or find the OTC counterparty now.
