Compute Futures are Here. Sort Of.
We now have compute futures, something I’ve been wanting for year. But they suck.
I have been talking about compute futures for a long time. Since before it was trendy. Before AI drove compute demand through the roof. World famous financial exchanges and back street prediction markets alike are now offering compute future. And I am… underwhelmed.
Why? Surely, I’ve been wanting a way to trade compute and now I finally have it, right? Wrong.
What we have today is the ability to speculate about the price of compute. I understand this is one side of what financial markets offer. It’s not the side that I care about. What I am, and always have been, looking for is the side of financial markets that are used by people that actually care about using the underlying commodity. Not just betting on derivative.
Details are rather scant on Kalshi’s new compute futures and given that half the information I’ve read about it doesn’t even understand the difference between a futures contract and a forward I’m not going to hold my breath that this becomes something useful for anything other than speculation.
If what you really care about is actually buying and consuming compute and want financial markets for all the benefits they bring to that, we still have several unsolved problems that everyone is trying to sweep under the rug. Who care about physical delivery or settlement if all you want to do is provide a cool new replacement to meme coins.
I don’t need to provide the world with another thing to bet on. I want to provide something that has real utility for those us that need to use that compute.
This means solving problems like delivery. No one that is serious about consuming large amounts of compute today is buying a futures contract on a prediction market because their use of that compute will be tied to some very specific APIs and network connectivity. We have nothing even approaching a standardised way to consume compute let alone arranging actual delivery.
It also means solving the fungibility problem. The performance of the same cloud VM type across regions can vary by over 20%, the difference in performance across vendors is even greater. A compute future for a H100 is a completely nonsensical product if you actually care about using that compute and not just speculating on the price of it. And no, the solution here isn’t to make everything homogenous. It’s to know how to have the right exchange rate. An exchange rate that varies by workload.
I’m not just ranting about this either. These are problems that not that hard to solve. We are doing our little bit (see ORB and FLOPx). Doesn’t make it any less frustrating.
