AI Companies: Cereals or Airlines?

In the post The Munger Games, inspired by my first-ever attendance at the Berkshire annual meeting and purchase and reading of Poor Charlie’s Almanack, I promised more commentary on Charlie Munger’s book once I’d reflected on it. One insight that stuck with me has come to the fore lately as I’ve tried to get my head around AI – an effort that isn’t theoretical, as I own an AI infrastructure stock and vacillate between holding and taking profits. This is a subject that is being intensely studied by analysts with expertise, but I am unburdened by that so will just apply a simple mental model (as Charlie might have done).

The specific Munger insight I’m referring to appears in the Almanack’s Talk Two, a speech to the USC Business School on April 14, 1994. He said, “Many markets get down to two or three competitors, or five or six. And in some of those markets, nobody makes any money to speak of. But in others, everybody does very well.” Munger cites airlines as an example of the former, putting one in mind of Warren Buffett’s joke that investors would’ve been better off if the Wright Brothers had crashed at Kitty Hawk. Cereal companies are an example of the latter, where competition is intense, yet everyone maintains a good-to-excellent return on capital.

Until two weeks ago, the market seemed confident that the leading AI developers were in a cereal-type competition, with robust profits to come for all. The emergence of Chinese open-source company DeepSeek shook them, leading to a rout in publicly traded AI-related stocks. DeepSeek produced an LLM (large language model) that achieved similar results to industry leader Open AI’s, reportedly using old Nvidia chips and at a small fraction of the cost.¹ If AI were to be commoditized, then how to justify the huge valuations being given to these companies? For example, Softbank is reportedly preparing to invest $25b in Open AI at a $300b valuation. (This decision shouldn’t comfort an AI company bull, given Softbank’s uneven track record.)

And if these companies are unable to generate the profits investors forecast, then how can they afford to keep buying so many of the chips that have made Nvidia worth $3T? The hardware, memory, and infrastructure companies, such as the AI data center company I own, are arguably a derivative of the developers and the chipmakers and will be worth less if the “bubble” deflates. Note, I put bubble in quotes because otherwise I would be begging the question addressed in this post: whether AI investing is a bubble.² If it is and deflates, the expected massive growth in power demand driving utility and uranium stocks will also never come (I’m pretty sure it never will anyway: see note 3).

Munger in his USC speech observed that cereal makers benefit from strong brand identity and a rational approach to competition. Airlines try to brand themselves, but at the end of the day the consumer is more likely to switch from United to Delta to save money on a flight than from Life to Cheerios to save $.50. As most reading this probably have, I’ve tried all the free AI models and don’t have a preference. As an experiment, my friend and fellow Steely Dan fanatic Pete Jakab asked AI to compare the songs Your Gold Teeth and Your Gold Teeth II and no model gave the “best” answer, but each had different strengths. ChatGPT’s answer came closest to our opinion, but DeepSeek’s answer was more informative and nuanced.³

On CNBC last week, Open AI’s Chief Product Officer Kevin Weil said that, while the company is mad that DeepSeek may have stolen some of its IP to train their model,⁴ they weren’t overly concerned. Weil explained that Open AI is better than DeepSeek because it “operates according to Democratic values,” moreover it will soon release the new version of ChatGPT that is “head and shoulders” above the competition. As to his first point, unless they’re researching Tienanmen I doubt people care much about the politics of their AI, and even if so — even if DeepSeek were banned here — the genie is out of the bottle on developing AI models cheaper. As to Weil’s second point, that is beyond my limited competence, but experts have addressed it and one good place to start might be this Substack by

Gary Marcus

:

The race for "AI Supremacy" is over — at least for now.

If Marcus is right that none of the AI companies can achieve technological supremacy, and if they have no brand identities, then they start to look less like Kellogg’s and more like airlines or mobile phone companies, another sector in which return on capital has long been anemic. The AI giants may be hoping to benefit from a lock-in effect, the way Amazon Web Services has many customers captive because of the difficulty of switching. The idea is that if a business has trained its model on ChatGPT, and the more its platform is integrated with it, the harder it is to change. However, AI is in the early stages of adoption, much of it is open and interoperable, and customers can and are already developing model-agnostic AI to avoid capture.

Later in the interview, as if somehow to justify the current valuation of AI companies, Open’s Weil cites the rapid advancement of AI’s capabilities and the benefits it will provide to humanity. To this I offer a sincere thank you but again quote Munger’s USC speech: “The great lesson in microeconomics is to discriminate between when technology is going to help you and when it’s going to kill you.” He recalls that Warren Buffett’s reaction to the invention of a new and doubly efficient loom was that if it worked, he’d close his textile company. Buffett refused to invest more capital into a commoditized “lousy business.”

Airlines and mobile phones have been wonderful inventions for humanity, but only the first movers reaped extraordinary profits. In the 1990s, companies like Global Crossing spent billions building the “information superhighway,” but mostly went bankrupt. Who did wind up benefitting were the Googles and Amazons, which grew on the back of the highway, and we people of the world, who use the internet. Similarly, it is possible that Nvidia’s enormous profits reflect its first mover advantage but are not predictive of its own long-term earnings or that of any other AI company. If the hyperscalers do continue spending lavishly on AI – and we’ll know soon whether DeepSeek has impacted their capex plans – this may ultimately redound not to their own benefit but to that of Main Street businesses that will have affordable AI to improve efficiency, and to humanity. Unless AI decides to kill us all.

1. The DeepSeek news stories’ ubiquitous use of the phrase “fraction of the cost” has irritated me because they often omit the modifier “small.” Without that, the fraction could be 99%. Okay.

2. My second bugaboo is the use of “begs the question” as a fancier-sounding form of “raise the question,” when the phrase has a very specific meaning, which is to express a tautology. Okay.

3. If this is the kind of question AI is used for, then there is no world in which its development justifies the power consumption that is being forecast. Inverting the problem, as Munger would advise, we can see that more efficient models like DeepSeek had to emerge (and will yet improve) or the industry would’ve collapsed under its own weight.

4. An obvious case of the pot – Open AI being one of the greatest IP stealers in history – calling the kettle black.