Should AI Companies Be More Expensive Than Dow 30 Stocks... Really?

To understand how stocks will be impacted by AI, we should stop looking to the future. Look backwards instead, to the 1920s — another era when, in the words of John Kenneth Galbraith, 'men sought not to be persuaded of the reality of things, but to find excuses for escaping into the new world of fantasy.’

The parallels between the 1920s and the 2020s are uncanny. Commentators today blur the lines between sci-fi and reality. Where does it all end? Will optimism produce a great crash, as it did in 1929? Or will we arrive at what Ray Kurzweil calls the ‘Singularity,’ the point where AI surpasses human intelligence and technological change becomes exponential?

This week, I will present the bear case against AI, with a particular focus on LLMs (Large Language Models) and two companies leading this space, namely OpenAI and Anthropic.

In a subsequent article, Graham will offer a more bullish rebuttal.

Writing On The Cocktail Party Wall?

Before diving into financials, we should analyse the zeitgeist of our time.

There are two telltale signs of a bubble.

1. Skeptics who warn of danger are shunned;
2. Shares become the main topic of conversation at dinner parties.

Tell someone that AI is an overhyped fad, and watch their reaction. Many will dismiss you as a naysayer. Keep your ears open on the Tube, or in the supermarket. How often do you hear someone say ‘AI’ or ‘ChatGPT’? When everyone is talking about the same thing, herd mentality may be taking hold, the very psychology that fuels market bubbles.

Another telltale sign of a market bubble is people saying: ‘It’s different this time’. Just before the 1929 crash, one economist said that markets had reached ‘a permanently high plateau’, as if the laws of finance had changed and stocks would go up forever. Today, people say that AI will change everything. Maybe it will, but we’ve heard it all before.

The following illustrations are fictional, generated using DALL·E 3 (OpenAI). Nevertheless, they are plausible caricatures of the times. They highlight plausible parallels between the 1920s (left) and 2020s (right).

Do The Valuations Make Sense?

The clearest sign of a bubble is the high valuations that OpenAI and Anthropic have secured in recent funding rounds. Investors need to answer a simple question: does it make sense for these companies to command higher valuations than Coca Cola and Chevron?

The missing ingredient is —you guessed it— a profitable business model.

• Microsoft’s financial statements suggest that OpenAI made a $12bn loss in Q3 2025.

• Sam Altman (OpenAI’s CEO) has warned losses will persist.

• Similarly, Anthropic’s gross margins were negative -94% in 2024.

Profits could be limited, due to the high cost of building AI. OpenAI plans to spend $1.4tn on compute capacity over the next few years — that’s equivalent to the GDP of Spain! It would only pay off if LLMs improved when we trained them with more data.

In reality, returns may diminish...

LLMs are ultimately statistical models. Words are represented as numbers. Predictions are based on probabilities. LLMs don’t think. They assemble words which statistically make sense. At a basic level, they predict sentences like: ‘The [ ] sat on the [ ]’ — like auto-complete on your phone. LLMs are auto-complete ‘on steroids’, trained with more data.

More data improves predictions, up to a point... and then it doesn’t. Think of a coin toss. After 100 flips, you’ll generally see a 50/50 split between heads and tails. Another 200 flips won’t change the ratio.

Language works the same. After enough Christmas cards, the pattern becomes obvious: 'Dear so and so'… insert seasonal greeting… then sign-off with, ‘Best wishes.’ You won't improve the AI by showing it more Christmas cards.

Now consider the investment case. Companies pour money into data infrastructure. They hope to produce better models, but what if performance plateaus… while data costs keep rising? You see the problem. 

Don’t take it from me. Yann LeCun, the Chief AI Scientist at Meta, said:

If you think that we're going to get to human level AI by just training on more data and scaling up LLMs, you're making a mistake.

In a similar vein, AI-critic, Gary Marcus, has said:

There is an old saying, that the definition of insanity is doing the same thing over and over again and expecting different results. Basically, that's been the field of AI for the last few years… The scientific hypothesis is you pour more data, you'll eventually get AGI. And the thing is people keep trying that experiment. It doesn't really work.

Betting on Breakthroughs

The risk for investors is that inflated market valuations reflect a breakthrough that has not yet happened. It is as if investors were being asked to invest in the Wright brothers in 1903, on the basis that jet propulsion was ‘just around the corner.’

Sustained, economically viable innovation will require breakthroughs beyond the current LLM paradigm, potentially integrating LLMs with other approaches, like World Models.

• LLMs predict the next token or word in a body of text.

• World Models predict the next state in a given environment.

The idea is that World Models could facilitate progress in autonomous driving, robotics and other areas, especially if we create a sort of LLM-World Model hybrid. However, progress is by no means guaranteed, for several reasons.

• Data availability: We train LLMs by ‘simply’ feeding them a corpus of digitized text. World Models train on richer and messier datasets, encompassing sensor data and video footage. This is more expensive to source.

• Closed-world vs open-world learning: LLMs improve by training in the relatively ‘closed’ world of text and language, where rules are more defined. Conversely, World Models operate in more dynamic, unpredictable environments, where it is unclear that learning scales.

• Resource intensity: Training World Models is computationally expensive. In order to generate a return on investment, we need to scale models over a long and uncertain road, towards products that people willingly pay for because they solve complex problems reliably (e.g. ironing your shirt without burning the house).

While the above challenges are being addressed, investors risk missing opportunities in other areas. So unless we want to speculate about the future, our investment thesis should be grounded in the technology that exists today. This brings us back to LLMs.

How Good Is the End Product, Really?

AI companies should only be worth as much as Coca-Cola or Disney if their products are going to be used by everyone, everywhere. Towards this end, LLMs are marketed as general-purpose ‘silver bullets’ — tools that can be used by everyone, everywhere… and for anything.

In reality, we still need specialists in the loop to ensure quality, safety, and correctness. You can’t fully remove experts without consequences. This potentially limits mass adoption.

• Companies are already reversing course: A UK study found 55% of firms that replaced employees with AI later rehired, possibly because of the following quality and safety concerns.

• Security failures are real: Moltbook —a social platform built with 'vibe coding'— was hacked and 35,000 email addresses were stolen. These breaches can potentially hit corporations hard. Think about the possible lawsuits, reputational damage and regulatory fines.

• Quality is inconsistent: A study by the Center for AI Safety tested whether AI could complete freelance software tasks, e.g. on Fiverr. The jobs were simple. The results were not impressive. In only around 2% of cases did the AI complete the job successfully.

• There is also a theory around productivity gains: AI in theory makes workers faster. But is this an illusion? Model Evaluation & Threat Research (METR) tried to find out. In early 2025, they ran an experiment with two groups of software engineers: One used AI. One didn’t. The non-AI group finished a set of tasks faster than the AI group.

Gary Marcus sums up the problem:

The quality that's been promised isn't really there. It doesn't really make sense.

The bottom line is that LLMs are powerful tools, but not magic. Over time, users will become more discerning about when, where, and how AI can be used.

One Final Question: Where is the Durable Moat?

Even if LLMs do become commonplace tools, it is not clear that OpenAI and/or Anthropic would be the long-term winners. It’s worth running Warren Buffett’s thought experiment: ‘If I had a tonne of money, could I compete with the business?’

Admittedly, Anthropic may have an edge over OpenAI, in terms of specialisation and switching costs. Anthropic has started to embed itself into daily workflows, by positioning itself specifically as a developer 'side-kick', rather than a generic LLM tool. This could make it harder for software developers to migrate away once habits are formed.

At the same time, both companies could see their moats pressured if LLM learning plateaus. In this scenario, competitors could catch up with today’s ‘good enough’ performance which, in turn, becomes a commodity.

• On the theoretical side, AI products are built on algorithms, but much of the theory is public, for instance, in academic papers. Even patented breakthroughs expire. At the end of the day, information wants to be free!

• On the practical side, data infrastructure helps create a moat, but there’s a ceiling. If the model no longer improves when we add 'more data', then scale isn’t protection. Any firm with enough money, infrastructure, and elite engineers can spend billions and compete.

Some analysts already warn that models like DeepSeek could trigger a price war. In this context, the real advantage would be ownership of proprietary domain data. In finance, firms like Bloomberg could outcompete general AI labs. Imagine a world where LLMs are 'commoditised' and Bloomberg decides to integrate financial insights with LLMs. Bloomberg would have an edge. More generally, it may indeed make sense for industry incumbents to absorb emerging AI firms.

On the flip side, if data infrastructure does create a moat, then we may be entering a new, less profitable software paradigm. Historically, software has been capital-light relative to hardware-heavy industries. But AI may change that: building and training advanced models requires enormous capital for compute, chips, and data centers. On this point, it’s also worth mentioning that data infrastructure carries a depreciation risk, because AI chips advance fast. Older systems must be upgraded or they lose value.

If the AI boom slows, there is the potential for contagion risk across the financial system, because various pension funds and banks have started to invest in AI. All of these factors lead me to believe that OpenAI and Anthropic most probably should not be priced above giants like Boeing or Chevron. The profits aren’t there. The math doesn’t work. The growth story is speculative. They may survive, but are they blue chips? Unlikely.

AI companies are selling a powerful story - i.e. we must adapt now or be left behind. Whether their claim is true is beyond scope. The key issue is that AI companies are incentivised to sell a compelling vision. And they’re entitled to promote that vision, but as investors, we have zero incentive to follow the herd…