The AI bubble

AI is a bubble; data is a major constraint (Source: here)

In November 2022, OpenAI launched ChatGPT, the world’s first free-to-use, publicly available Artificial Intelligence (AI) chatbot. Since then, a number of similar AI chatbots have been launched, including Google’s Gemini, Microsoft’s Bing, You.Com, etc. In addition, AI agents - artificial intelligence systems that can act autonomously to achieve specific goals - have also been adopted in many other areas of business, including self-driving cars, service evaluation, gaming, and industrial robots.

Anyone who has used an AI agent, such as a chatbot, will agree that AI technology is impressive, perhaps even revolutionary. Still, the same can be said for most new technologies. The key question from an investment perspective is not whether the technology can change they way we do things, but whether the future income from AI-related investments can support the current elevated equity valuations of AI companies?

To answer this question, let us examine the valuations of AI companies, which happen to be pretty eye-watering. Alphabet, Amazon, Apple, Meta Platforms, Microsoft, Nvidia, and Tesla, AI companies collectively known as the ‘Magnificent Seven’, saw their average share price rise by a whopping 65% in 2024 (see table below).

Company name	Symbol	YTD return (2024)
Alphabet	GOOGL	37.4%
Amazon	AMZN	50.7%
Apple	AAPL	28.8%
Meta Platforms	META	78.2%
Microsoft	MSFT	19.6%
Nvidia	NVDA	177.3%
Tesla	TSLA	68.3%

Year-to-date 2024 returns for AI companies (Source: here)

NVIDIA, the chipmaker and advanced AI software producer, currently trades at 53 times its actual earnings. This compares to the medium price to earnings ratio of 17 for S&P500 companies and a typical range of 20-30. The dramatic rise in the valuations of the ‘Magnificent Seven’ stocks means that these seven companies now account for 34.6% of the total value of the S&P500 index, which, it is worth remembering, comprises America’s 500 largest and most important listed companies.

It is not a sign of good financial health when just seven companies account for more than one third of the total valuation of an index that comprises hundreds of mostly well-run American companies.

Another factor that points to a bubble is that the surge in AI company share prices has been accompanied by extreme cheer-leading by the AI companies themselves. The cheerleading appears to have been highly successful, contributing to precisely the kind of investor herd-dynamic one usually sees ahead of bursting bubbles.

AI industry proponents have been falling over themselves trying to convince investors of AI's massive future potential, even as the companies still struggle to break even on their existing AI investments. Undeterred, AI proponents claim that AI will soon perform the same intellectual tasks as any human, including learning, reasoning, adapting, and making collaborative decisions. Just over the horizon, supposedly, lies the holy grail, when 'Artificial Super-intelligence' will increase machine intelligence exponentially. At this stage, it is alleged, machine brains will surpass human intelligence in all domains, operating beyond the limits of all human comprehension. This milestone, many say, will be reached within a decade (Source: here)

There are good reasons to discount much of this type of hype, because AI insiders have strong incentives to exaggerate the potential of AI, especially at this very early stage of development of the technology. No one can yet tell for sure what the future for AI holds, but all the big AI players fear they could miss out on some future game-changing moment that will sort the winners from the losers, exactly as happened in the Smartphone revolution, where Apple and Samsung emerged winners, while Nokia and Microsoft lost out.

In order to remain in the game to have a shot at the hoped-for-but-yet-still-elusive future payday, today’s big AI players are ploughing enormous amounts of money into AI. Privately, they will be aware that much of this investment is highly speculative and may only pay off in the far distant future, if at all. But they won't tell you.

Now, if only one company invested recklessly and lost then it would not be a major concern, but if an entire industry engages in over-investment then we are almost certainly in a classic bubble. After all, this is exactly what led to the DotCom and Sub-prime bubbles earlier this century and indeed in every other bubble in the past.

Already now, the extraordinary claims of the AI proponents are being challenged in some quarters. Some label them as misleading, exaggerated, or even outright lies (Source: here). More importantly, the claims made on behalf of AI appear to run counter to well-established long-term empirical evidence, which shows that research ideas are becoming less productive and less disruptive (See here). The charts below illustrate these points.

Research productivity is declining (Source: here)

Research findings are becoming less disruptive (Source: here)

There are also indications that AI fundamentals are running into trouble even as AI investment piles up and AI stock prices continue to soar. One area of particular concern is data, which happens to be critical to AI growth. AI engines acquire their 'intelligence' by analysing data. Lots of data. All else even, the greater the amount of data, the higher its quality, and the greater its diversity the greater the learning potential of AI engines. However, in order to sustain momentum in performance, AI engines need access to exponentially more data. Unfortunately, the evidence is mounting that both access to data and data quality are beginning to constrain AI performance.

For one, there is not enough diversity in the available data. The most important source of data for AI engines is human activity on the internet, such as what we take pictures of, what we chat about, what we watch, what we buy, what we upload. Unfortunately, while this type of data is abundant it is only a thin slice of makes the world go around. For example, it only relates to our own species and it is mostly confined to retail commerce and entertainment of various kinds. This type of data is valuable if you aim to use AI for marketing purposes, but it will not get you very far if you are, say, trying to find a cure for cancer.

Second, the available data contains all kinds of biases as certain types of information are over-represented, while others are under-represented. Take, for example, what people look at on the YouTube. More people look at cute kittens than rat pups. Cute animal stuff is therefore heavily over-represented in the data, while disgusting animal stuff, which may be equally if not more important in shaping a correct and balanced understanding of the world can barely be found.

Finally, there is a strong skew in the data towards stuff that we readily understand, hear, touch, smell, taste, and see, in other words, stuff we can relate to. However, the vast majority of the things that drive the world – from micro-organisms and chemical processes to the forces that move the planets in our solar system – occur on a scale that makes it very hard for us to interact directly and generate data, since these things are way out of our immediate cognitive range.

As if the data limitations are not a serious enough concern, there are also major but poorly recognised issues with data quality, because so much online data is now distorted by monopolistic business practices. Specifically, what we search for, watch, buy, and interact with on the internet does not really reflects our free unconstrained will. Instead, our consumption and social choices are expressed within narrowly-defined and deliberately manipulated platforms upon which our choices are anything but free. If AI engines tap into this type of constrained behavioural data, what do these data actually reveal about us? Do they really tell us anything about ourselves, reality, truth, and nature itself? Probably not. Garbage in, garbage out.

AI enthusiasts are quick to point out we may be able to generate the data we need, but this often fails to take economy into account. It costs money to generate data and each successive venture into progressively less data-rich areas requires progressively greater outlays, rapidly undermining the AI investment case. Data limitations can be offset to some extent by the development of better chips as well as enhancements in transformer architecture (neural networks that convert input sequences into an output sequences). Yet, in my opinion, these technological enhancements only get you so far. If the data – the raw material of AI – is in short supply, biased, bad, or even missing entirely then AI will perform accordingly.

­­­­­–––––

AI is likely to have a bright future in the context of operations closely associated with human activity about which we have abundant and cheap data. When it comes to fundamental breakthroughs across the knowledge frontiers in the realms of science, however, the outlook is far more uncertain. Investors are likely to have to scale back expectations about what AI can deliver, particularly with respect to some of the more outlandish predictions for AI.

Due to the importance of access to good, unbiased data on human activity to feed into AI engines, it is essential that human economic activity can unfold in an inclusive and competitive environment with the widest possible participation. This means policy should aim to counter the current trend towards ever-greater monopoly power in the online economy. One way to do this is to promote Open Transaction Networks and other technologies that introduce competition by design (see here).  

Government funding of fundamental research has declined as a share of total research funding (Source: here)

Finally, it is critical to reverse the decline in the government's share of research & development (R&D) spending, which has declining since the 1960s (see chart above). Privately funded R&D has delivered great advances in many areas and must continue, but the focus of private R&D is mainly to develop commercially viable products rather than to gain fundamental insights. If AI is to find application in areas other than just satisfying humanity’s immediate consumption needs then data needs to be generated in areas where it are not commercially viable. And that requires government leadership.

The End