Will Compute Futures Compute?
Time to scrape the rust off the ol’ blog! Been a while. Life comes at you fast sometimes.
The Rustoleum (a product originally from my home town, whose plant I watch burn in a spectacular fire while waiting to play baseball in James Park) is a post about the new Big Thing, which is actually something I wrote about last September–compute futures.
The CME, ICE, and tokenized exchange Architect Financial Technologies have announced the launch of futures contracts on compute. This land rush is a direct consequence of the AI boom. (I will write a post in the near future on AI issues more broadly).
The main attraction. The compute market is huge, and growing. As CME CEO Terry Duffy said in a statement touting the launch:
“As the backbone of the digital economy, compute is the new oil of the 21st century,” CME CEO Terry Duffy said in the statement. “Every AI model trained, every transaction cleared and every byte of data processed runs on compute, which is becoming a fast-emerging asset class in its own right.”
As I noted in my 2025 post, market size is a necessary but not sufficient condition for a successful futures market. Another important condition is volatility.
Most of the recent angst about compute cost is its direction, not its volatility: Compute costs have been increasing steadily as demand from AI has outstripped increases in chip supply, data centers, etc. That is, the market is largely directional:
A stead trend down through late-2025, then a steady upward trend through the present (with a weird blip in March, 2026).
Long-term trend dominated markets are not conducive to successful futures–you need substantial variation around the trend. Not much volatility evident there.
Ironically, Don Wilson of DRW–a big proponent of these markets (one of his firms creates the index graphed above)–pooh-poohed the relevance of the failure of DRAM futures as a precedent to the prospects for compute futures precisely because DRAM prices were dominated by a (downward) trend. The same logic would apply to compute futures.
As I also noted last year, another important contributor to futures market success is a fragmented, highly intermediated value chain. I further noted that the AI/data center/compute value chain appears to be anything but. The biggest players in this market are therefore not natural participants in this market.
You might argue that AI users could be, especially since AI cost has become a big issue of late. (More on this below). But in many products, big, low-leverage, public firms have little incentive to hedge price risk. Indeed, it’s long been something of a mystery why corporates hedge. Long ago Clifford Smith and Rene Stulz identified motives, including taxes, leverage, and agency costs. But I don’t really see those as driving widespread corporate demand to hedge compute costs.
This is not to say that compute costs are not an issue. Indeed, several stories in the last few days make it clear that initial AI euphoria is running up against economic reality. Namely the reality that AI ain’t cheap.
And:
He said that Uber CTO Praveen Neppalli Naga went viral after telling The Information in an April interview that Uber had already blown through its Claude Code budget for 2026.
The comment led to what he described as a “head-exploding moment,” sparking discussions about AI token consumption within the company and the trade-offs it creates, such as on head count.
What makes my head explode when reading these stories is that these supposed business geniuses didn’t realize that treating something as free would result in massive overconsumption. I mean seriously: they didn’t think this was going to happen with probability one?
Maybe they should have considered the fate of Red Lobster.
A couple of implications of this. The first is that it is likely that Anthropic et al have way underestimated the elasticity of demand when the market settles down into an equilibrium in which AI end users exercise more judgment and control over their use.
The second is that companies will definitely have to figure out organizational and management strategies to incorporate AI into their businesses. The proper use of AI will require the exercise of judgment, and the development of mechanisms that incentivize the balancing of benefits and costs. Since (a la Coase, 1937) whereas markets allocate resources by price, within firms price mechanisms are typically uneconomic so they rely on fiat instead, these mechanisms will focus on budgets and managerial decisions on when and how to employ AI.
These implications have implications of their own.
First, since AI is costly–not least because of compute cost–normal market mechanisms will prevent or at least strongly mitigate the doomsday scenarios of mass unemployment. Many of the doomsday scenarios are based on bad economics, specifically, focusing on absolute rather than comparative advantage. “AI can do everything better!” But when a resource is scarce, that doesn’t determine market outcomes. Market prices, for compute, and for labor and capital, will reflect comparative advantages. AI will hurt those occupations in which it has a comparative advantage. But you can’t have a comparative advantage in everything.
Second, we are in the early adoption stage of AI. The shock companies are experiencing when looking at their compute costs will lead them to rethink and refine their usage (as discussed above). This will impact the demand for AI–both in the level and the elasticity. That will be a learning process, leading to demand fluctuations. AI suppliers will take some time to learn how demand is evolving, and their pricing policies will change accordingly. Which will feed back to firms, and also to upstream suppliers (GPU fabricators) will also have to adjust their pricing.
In the near term this will lead to price volatility. A sort of reflexive, endogenous volatility. But I would expect that to settle down in relatively short order, in which case pricing will be dominated by upstream costs, notably chips and power. Especially since there will likely be pretty intense competition between AI firms.
Meaning that this adoption learning process volatility is likely to be transient, and insufficient to support a big futures contract in the longer run. And the trending aspect implied by my posited post-settling down equilibrium will not generate the short term volatility needed to support a futures contract–even assuming the other issues (the value chain, weak corporate hedging demand) do turn out differently than I expected.
In other words, despite the hype, I’m skeptical about the long term viability of compute futures.
Reminds me of bandwidth trading and the associated futures in Enron days. There are a couple components of a market: volatility, a definable contract, natural physical positions. But like bandwidth, there’s an inherent and massive long-term downward trend in prices that will make it difficult to see happening.
Much more relevant is are the energy inputs into compute, which can be adequately hedged today.
Comment by Hal — May 28, 2026 @ 1:59 pm
Natural buyers are the users of AI that might want to lock in a price and turn a variable cost into a fixed cost….natural sellers are who? Data centers, AI providers, NVDIA/Micron? The point about demand elasticity is correct. Without massive data center expansion, supply elasticity could also be a challenge as people adopt AI.
Eventually, I think we will see different models of AI. Right now, everyone uses one, and it’s the Ferrari. But sometimes you just need an AI model that is as good as a Ford. You don’t need the Ferrari. So, models will segment the market as well. Freemium models will do some things, and pay-for models will do more. Firms will build their own models (see Kirkland and Ellis) to keep data inside their silo, and so they don’t unintentionally share trade secrets or information.
New markets are hard to start and build liquidity in, even when they are in the right place at the right time
Comment by Jeff — May 28, 2026 @ 11:29 pm
Compute isn’t particularly fungible. AI compute (training) is optimized for very particular hardware – code written for google’s tensor chips is not going to run on GPUs, etc.
Non-AI workloads (representing most of the purchased “compute”) are tied to very specific cloud environments, software licenses, etc. Moving systems between clouds is generally not possible.
These days, many workloads are memory/cache bound rather than CPU bound.
Nor can compute be stored.
“I’m skeptical about the long term viability of compute futures” would be a very mild way of expressing my reaction.
The graveyard of failed future contracts is large and overcrowded.
Comment by derriz — May 31, 2026 @ 8:50 am
I think another argument in support of the Prof’s position concerns scarcity: There’s long lead times for chips for sure, but they’re not any longer than the time taken to build a new data centre, so on the timescale of corporate decision making, additional AI compute is available on-demand. And if they DO need additional compute quickly, they can buy it at comfortable spot prices from the existing commodity suppliers, as the Prof notes above.
Comment by HibernoFrog — June 3, 2026 @ 4:28 am
1. I think can find a timeframe of the TY or SPX that looks similar. Likely a shorter time period I’ll grant you but still this chart shown has 2 trends not 1. I will say that there certainly is a lack of intra period volatility. Given how early we are in this buildout, are we to assume that prices will only go up from here?
2. Has any of the LLM’s opined on the their desire to lock in some of their compute costs? On the opposite side would hyperscalers hedge some their output coming on-line in the next few years. I understand it is very different from the oil and natural gas markets, but aren’t there some similarities?
3. If corporations hedge interest rate risk would they not want to hedge some of their AI costs once they get a handle on their needs?
I just started looking into this market so I am looking to hear both sides of the debate.
Comment by Mike Dillon — June 15, 2026 @ 12:16 am
I am starting to think that local MW capacity (deliverable via data center infrastructure) may be the best route. Chip design agnostic.
Comment by Mike Dillon — June 15, 2026 @ 1:41 am