AI from an investor's perspective.

Part I: Is AI a bubble? Are AI valuations crazy?

AI is everywhere these days, and is in a way a hype. Even the number of LinkedIn posts about AI is currently in bubble territory.

However, unlike the SPAC, Weed or Meme bubble, the AI hype is underpinned by real investments: Meta is currently building a data center the size of Manhattan. And although there is some circularity in the commitments of OpenAI, Nvidia and Oracle, the majority of the investments are indeed supported by real cash flows at most of the major tech players.

Only Oracle's investments raise eyebrows, with the planned investments looking very high compared to the current level of income...

As far as valuations are concerned: they are, of course, considerable. For example, Nvidia has a forward price/earnings of a substantial 38.4x. And that while the revenue model of the major AI players still remains somewhat of an issue. But more about that later this week, in a next section. Yet we don't follow the most dramatic headlines: Insane would we have the current valuations (still?) don't mention: At the height of the dotcom bubble in 1999, Microsoft recorded a P/E of 71x, and Cisco even recorded 171x profit.

So AI certainly has a hype content. But we can't call it a real bubble for now...

Part II: Divergence or winner takes all?

The current build frenzy and the battle for ever larger data centers may give the impression that the AI market is becoming a “winner takes all” model. Kind of like Google has a 90% market share about search queries. Where network effects ensure that every subsequent search becomes better.

However, this is not the case with AI for now: as Deepseek showed last year, it is indeed possible to create a model with a relatively limited budget that is not much behind the industry leaders. In addition, there is also a strong push towards open source models, especially from China. This means that more than 2 million different AI models already exist.

It's still early in the AI battle. And time will tell. But it is a remarkable observation that, on the one hand, there is considerable investment in data centers — as you would expect in a winner-takes-all market — while, on the other hand, the number of models and their use suggests the opposite. In addition, first-mover OpenAI is currently losing strong market share...

Beyond ChatGPT

When it comes to AI, the popular press continues to be mainly about OpenAI. In reality, OpenAI remains popular among ordinary consumers, but Sam Altman's company is losing ground in the professional market.

While, as a consumer, you have almost unlimited amounts of your personal AI assistant in your pocket for a monthly subscription, a professional party pays based on your usage, at a price per “token” used. Platforms such as OpenRouter offer the possibility to shift very flexibly between the different models, and are therefore a useful proxy for seeing which models are most used. Below is the table with the tokens used across the different models.

Through promotional offers and free versions, Elon Musk's Grok was able to take 2 stage places. After that, Gemini and Anthropic's Claude are doing very well. ChatGPT no longer has any model in the top 10... Do AI models also have a 'winner's curse'?

Part III: About Capacity

In a previous post, we already talked about the huge investments in AI. In the coming years, the “big 5” tech players will spend up to 500 billion a year on capex to build data centers. These are amounts that we can barely imagine. For reference, that is about the same as the entire European agricultural sector. Or about 20 times the global annual turnover of McDonalds, should that make it concrete...

It raises questions about whether this will not lead to enormous overcapacity in the long run. Especially since many “smaller models” with more limited computing power still manage to reach 80% of the level of the best models. And can be “good enough” for many tasks.

Personally, we get a bit of “submarine cable boom” vibes from the early 2000s, where the Internet took over the world and a huge overbuild of overseas internet cables was built. In the end, most of the investments ended up somewhere near the cables: deep underwater. But the infrastructure that was built at that time did ensure that the Internet could develop into what it is today in the coming decades.

Are the current AI players the new submarine cable companies? Will the construction boom soon cause overcapacity & plummeting prices? In that case, as an investor, it's better to stay on the sidelines. But as a user, you can really enjoy such a situation. What do you think?

Part IV: AI as StockPicker

There are over 50,000 listed companies worldwide. As an analyst, it is therefore impossible to get to know them all in detail. This is one of the reasons why computer models have been well established in making stock selections for decades: ratios and valuation measures are calculated based on financial figures, and initial filtering takes place.

The still wide list of companies that survive this screening is flagged as “potentially interesting”. After that, the analyst starts doing his homework. As Warren Buffett puts it: you start with companies with the letter A, then go to those that start with B, and so on...

The disadvantage of this approach is, on the one hand, that the most interesting information is often not in the numbers themselves. But rather in the notes, the outlook, the guidance that company management formulates, and information that provides the hard figures with the necessary context. Until recently, this was only possible by carefully reading all business documents with the necessary discipline. And the most common decision in stock analysis is to ultimately not invest in it: for each company that is ultimately considered to be investable, an analyst often looks at 10 and concludes “not”.

Enter Large Language Models (LLM): these are just cut to go through large amounts of text and summarize it. Or looking for inconsistencies, quarter-to-quarter changes, and more. By making smart use of (well-trained) AI analysts, the filter aspect can therefore be expanded and, in addition to the numbers, also allow screening on the text. In this way, we can already flag the “potentially most interesting companies”.

Of course, it remains necessary that the human analyst then does his homework with the necessary knowledge and discipline. Simply asking ChatGPT what would be an interesting investment and blindly following that advice is obviously not a good idea. But using well-trained AI to screen documents, to make the list of companies you delve deeper into as relevant and interesting as possible as an analyst, helps stock analysts increase the hit rate.

Part V: How West-Flemish money laid the cornerstone of the AI revolution.

One Ray Kurzweil already predicted in 1990 that if the computing power of computers continues to improve at the same rate, we will be at the point where Artificial General Intelligence (AGI) will be reached by 2029. The remarkable thing is that in 2025, we are still on track to achieve that goal. So just hold on for a few more years! In that regard, it's not surprising that the big tech players with big budgets are throwing at AI: we can almost touch the holy grail.

The name Kurzweil may still ring a bell for West-Flemish investors with a good memory. In 1998 and in the dot-com bubble, his firm, Kurzweil Applied Intelligence, was purchased for 53 million USD in cash by... Lernout & Houspie.

When L&H went down, the man was full of cash and a passion for AI-avant-la-lettre. He spent his time wisely and founded the “Singularity Summit” in 2006, together with Peter Thiel, among others. It was at that Singularity Summit that Demis Hassabis, the founder of Deep Mind, with Peter Thiel. And so he managed to get the necessary funding for his AI company from Peter Thiel and... Elon Musk, among others.

Deep Mind was purchased by Google in 2014 for a whopping 500 million USD. Although he passed the box office, Elon Musk was here not amused. He tried to make another counteroffer, but it was rejected by Deep Mind. As a countermove, he started a new company, with the aim of developing AI as open source software. Named... OpenAI. The rest is (another turbulent) history.

But with the necessary nuance and editorial freedom, we can therefore (loosely) say that the West-Flemish L&H investor was one of the founders of what ultimately became the current AI wave...

‍

Authors: Jens Verbrugge, Tibo Dewispelaere, Wouter Verlinden