Going Vertical
In my recent post on Don Wilson and GPU trading, I expressed skepticism at the prospects for such trading based primarily on market structure issues. Specifically, a commodity futures market can thrive if the underlying product has a highly intermediated value chain, but that vertical integration or long-term contractual relationships between producers and ultimate buyers is inimical to their success. Further, I surmised that transactions costs considerations would lead to long-term contracts and/or vertical integration dominating the GPU sector.
And indeed, that outcome is becoming clearer by the day. OpenAI (it of ChatGPT) has agreed to a massive chip purchase from AMD and has also bought a 10 percent stake in the chipmaker. OpenAI has also entered into a five-year deal with Oracle to buy “computing power.”
There is also increasing verticality in power supply arrangements. (Apparently the new industry buzzword is “circularity” but we are really talking about vertical relationships here, so I’ll stick with traditional IO speak). Example: AI firms building gas-fired plants in Texas. Musk’s xAI building a gigawatt scale plant in Mississippi to power a data center outside Memphis.
Microsoft has entered a 20 year deal with Constellation Energy to restart–believe it or not–a unit at Three Mile Island. And on and on.
It’s somewhat amusing to me. Years ago Paul Joskow wrote a seminal article about coal mine mouth power plants in which he showed that this co-location led to long term contracts or vertical integration. We are now seeing data center mouth power plants.
The economics are basically the same. It makes sense to locate something that consumes as much (or more) power than a generator can produce to economize on transmission costs and enhance reliability. However, this creates bilateral monopoly driven transactions costs issues that lead to vertical integration and long term contracting.
Going back to GPUs, there is no reason for colocation, but economies of scale on both sides of the market–huge consumers operating data centers, and a small number of GPU producers–create similar bilateral monopoly opportunism risks, which in turn are leading to the implementation of traditional responses to these risks–tight vertical connections between producers and consumers.
This market structure is inimical to the development of a thriving derivatives market. So I wouldn’t be surprised to see the launching of GPU futures contracts, but I fully expect them to sink faster than a Nork destroyer after launch. But unlike the Nork ship, GPU futures would not be refloated.
That’s a convincing argument that the current big consumers of AI capacity aren’t likely to be interested in buying futures contracts for AI capacity, since they would naturally prefer to vertically integrate (thus becoming both consumer and provider, like we already see with cloud computing/hosting, such as Amazon Web Services and Microsoft Azure). But is there any merit to the argument that there are other demand sources who might be interested? Namely:
– New market entrants: Every company and his dog are adding AI features to their IT systems. If those are provided as turn-key solutions by the FAANGS/Open AI/Microsoft then sure, the Prof’s position holds. But we see in Cybersecurity, for example, that some big companies go in-house. Plus this is fertile ground for startups. So there could be a fairly large number of fairly small entities who would happily pay to secure the cost of their future AI processing capacity.
– Additional capacity for incumbents: Even if the FAANGS/Open AI/Microsoft are vertically integrating, would they cover just their baseload demand and buy-in capacity (possibly in advance – so, futures) to cover peaks? Or would they build capacity for even their peak load, and sell the surplus (so no futures)?
The market dynamics are interesting.
Comment by HibernoFrog — October 16, 2025 @ 3:04 am