Recognizing that AI is a bubble requires one to apply the right test to the right companies and that requires one to understand the fundamental nature of the bubble in question.

Generally speaking, bubbles exist when share prices and valuations are much higher than expected earnings (i.e., profits) would suggest. If  there are no profits then there must be a large amount of revenues and a recognizable path to profitability and testing that is often difficult.

Robert Shiller wrote the most famous book on bubbles, Irrational Exuberance, which won him the Nobel Prize in Economics. He emphasizes not only excessive share prices, but also the narratives to support those share prices. I agree, but again, which companies and which narratives should we be concerned with?

1929, Dotcom, and 2008 Bubbles

The Wall Street Crash of 1929 is generally attributed to excessive optimism about demand for a wide variety of products along with concerns about the banking sector. The top companies in early 1929 tell us a lot about those products: telecom (AT&T), steel (U.S. Steel), electrical infrastructure (GE), radio (RCA), automobiles (Ford), oil (Standard Oil of New Jersey) and food (Swift & Co., Armour).

The economy wide nature of the 1929 crash caused massive disruptions to output and employment and also left us with an emphasis on using price-to-earnings ratios to detect bubbles.

The dotcom bubble of 2000 revolved around the companies expected to benefit most from the internet economy. Despite the name dotcom, other than AoL (American Online), the companies whose shares rose the most in the Nasdaq before the crash were older companies such as Cisco, Microsoft, Lucent, Nokia, IBM, Oracle, AT&T, and Intel.  They were the big providers of the telecom, software, and hardware infrastructure. The dotcoms were expected to build products and services on top of that infrastructure. Amazon was not even in the top 100 for market capitalization at the end of 1999.

The 2008 bubble revolved around the banking sector and in particular mortgage loans for homes. Financial companies introduced a new type of financing of which one loan was called the sub-prime mortgage. Also referred to as near-prime, subpar, non-prime, and second-chance lending sub-prime loans are provided to people who may have difficulty maintaining the repayment schedule.

Those who watched the Big Short will know that Michael Burry recognized a few years in advance that home buyers who had sub-prime loans were defaulting at a much higher rate than were other borrowers, a problem that became bigger over time as housing prices fell and thus the value of the home fell below what was owed.^]

AI Bubble is Very Different

The AI bubble is vastly different from these bubbles because it involves a much smaller number of companies. The eco-system or value chain is actually very simple. Sometimes called AI Labs or AI software companies, OpenAI, Anthropic, and others sell their models to corporate and individual users and those models are processed in data centers owned by big companies such as Microsoft, Alphabet, Oracle, Meta, and Amazon; their size is connotated by their oft-used name, hyperscalers. Nvidia and other semiconductor suppliers then provide the chips for the data centers.

The small number of companies means that price-to-earnings ratios for all the companies in the Nasdaq or S&P 500 are useless. In fact, the low price-to-earnings ratios of these indices are in fact evidence of a bubble because companies using AI are not obtaining profits from AI except perhaps in special cases such as social media, search, and e-commerce, but even this is debatable.

Other companies are starting to benefit from the AI bubble, but they are related to the value chain described above. The over building of data centers, reminiscent of the over capacity of the 1920s and the 1990s, has benefited companies that provide electricity, construction equipment, and every type of semiconductor. But these companies are not the ones that should be a focus when determining whether there is an AI bubble, they are merely indicative of the magnitude.

Big Losses by AI Labs

A key issue in the AI bubble has been the huge losses by the AI Labs, which has been emphasized by Ed Zitron, Gary Marcus, myself and others over the last few years. These losses mean that they cannot pay the hyperscalers and thus the hyperscalers and semiconductor companies, particularly Nvidia, have had to invest in the AI labs or loan them money, something called circular financing.

Nvidia finances its customers much like providers of internet infrastructure once financed their customers, the telecom providers. That didn’t end well.

The huge losses also mean that users are paying prices far below costs, thus making statements about generative AI being the fastest growing technology of all time rather meaningless and misleading.

Many more people would buy Lamborghini’s if they were priced at less than half the cost. What would happen if users did pay the full costs of AI? Would they still use generative AI? The low cost Chinese models might help users but those models will also likely hurt the AI Labs thus being one more reason for the bubble to pop.  

Microeconomics and capitalism are based on users paying prices that cover costs. When the prices don’t do this, it becomes difficult to judge the extent to whether demand is real.

For instance, those betting heavily on Nvidia point to their huge profits and revenues, but what do those mean when it is loaning money to the AI labs and hyperscalers so they can buy Nvidia’s chips?

Skeptics have argued for years that much more revenues for the AI Labs are needed to justify the share prices and private valuations of the AI companies including Nvidia. A recent estimate is that “Nvidia’s $700 billion of revenue in two years would require about $2.1 trillion of end demand,”  basically for AI labs. Those revenues won’t exceed $100 billion in 2026. This is the biggest bubble ever.