It was a time of enormous optimism, based mainly on one specific technology that promised to have transformative effects on the economy. Businesses developing and implementing the new technology spent huge sums on construction and equipment. Individual investors piled into these businesses’ stocks, sending their prices soaring.

And enthusiasts were right about the technology’s potential. It would eventually transform the economy, indeed society as a whole. But long-term transformation doesn’t necessarily translate into profits for businesses at the cutting edge. As it became clear that the financial returns wouldn’t live up to the hype, some companies went bust, while stocks tied to the technology lost most of their value. And plunging capital spending pushed the economy into a nasty recession.

Yes, Britain’s Railway Mania of the 1840s was quite a story.

Over the past couple of months there has been a palpable change in the way people talk about the technology everyone calls AI. However, it isn’t really artificial intelligence as people usually understand it. Some commentators are careful to say “generative AI,” where in practice the extra word serves the same purpose as calling some supermarket items “cheese food” or “juice beverage,” indicating that they aren’t quite what they look like.

Whatever it is, few are denying that the technology is impressive. But warnings that there may be a huge AI bubble are getting louder. Worries about the financial underpinnings of all that capital spending are growing. And many people have noted that the AI boom is driving most, possibly all, of the economy’s recent growth. So what will happen if the boom goes bust?

By the way, some early subscribers to this newsletter may remember that I interviewed Jim Chanos, the famous short-seller, back in February, and he made the case then for an AI bubble. But people weren’t yet ready to hear it.

Today’s primer will not be about the long run economic and social implications of AI, which is an entirely different subject. I will write in a future primer about technology, growth and jobs in general, with some speculations about AI. For today, however, I’m going to focus on what history — especially, but not only, the telecommunications boom of the 1990s — can tell us about the AI boom and its consequences.

Beyond the paywall I will address the following:

1. The AI boom in historical perspective

2. The logic of technology manias and bubbless

3. The “Winner Take All” nature of technology bubbles

4. What will happen if the AI boom goes bust?

The AI boom in historical perspective

One can’t understand the AI boom without acknowledging that generative AI has solved problems that seemed intractable not long ago. For decades, attempts to get computers to do seemingly simple things like recognize images and interpret the meaning of written, let alone spoken, natural language just kept running into walls. Then, suddenly, we had software that responds to spoken requests, can not only summarize written material but can also write essays, can both recognize images and generate new ones, etc.

It sometimes seems hard to believe that ChatGPT was introduced less than three years ago, in November 2022. Since then, use of OpenAI’s chatbot and similar models has become incredibly widespread, to the extent that there is hardly any subject on which we can avoid asking, “Is what I’m reading/hearing/watching real, or is it AI-generated?”

The way generative AI achieves these results is nothing at all like human reasoning. Instead, as best I understand it — and my understanding is basic at best — it compares inputs — be they sounds, words, or images — to existing patterns in vast amounts of data scooped up from the internet, and then extrapolates from those patterns. As Noah Smith says, it’s kind of like magic: Powerful but mysterious — we often can’t understand how it arrived at a particular answer — and prone to occasionally going very wrong.

This process involves an immense amount of computation. People have often treated thought as something immaterial, free from physical constraints. But the “thinking” done by AI is anything but immaterial. It requires enormous, very physical data centers and vast amounts of electricity. As a consequence, the explosive growth of AI has required companies selling it to make very large capital expenditures. The biggest players in the industry are reportedly planning to spend around $370 billion on data centers this year, most of it in the United States. And enormous amounts of electricity will be needed to power these data centers, placing severe demands on power generation and distribution. So power companies are currently planning on spending huge amounts of money in expanding capacity – with some estimates clocking in at nearly $1.5 trillion over the next 5 years. Moreover, building this capacity will be especially costly to accomplish because of the crusade by the Trump Administration against renewable energy (and which the Chinese, in contrast, are well ahead of us in delivering).

Trillions of dollars over the next few years sunk into the AI sector and its energy needs – that’s a lot of money. But the U.S. economy is huge, with an annual GDP of more than $30 trillion. So we’re talking about something like a bit more than 1 percent of GDP. However, data center spending may rise further as a share of GDP in the years ahead — unless markets start to see AI as a bubble, and everyone scales back. If you add in the electricity costs of \(300 billion per year for at least 5 more years you are talking about more than 2% of GDP. We can get some perspective on the size of the AI boom as of mid-2025 by looking at a chart of capital expenditure on information processing and software over time. The recent increase in this kind of expenditure is a rough proxy for AI spending. As Chart 1 below shows, by the second quarter of 2025, capex on information processing and software had increased by about 0.5 percent of GDP since the introduction of ChatGPT. It is projected to rise at the same or higher rate during the second half of this year. PS: The vertical bars in this chart represent recessions, with the last one being the very short but severe Covid recession. s_!9Nte!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3045d6bd-fbfc-4324-9be9-805032d0069e_1429x786.png)

Chart 1

What Chart 1 also shows is that, at this point in time, the AI investment boom is only roughly comparable in size as a share of the economy to the telecom bubble of the 1990s. (Not the dotcom bubble, which was picturesque but much less important than big investments in fiberoptic cable networks.)

A recent comparison by researchers at the Richmond Fed also suggests that the current AI boom is similar in size to the telecom boom of the 90s:

![A graph showing the growth of a company

AI-generated content may be incorrect.](https://substackcdn.com/image/fetch/\(s_!5cMC!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591a63b6-416e-4434-91bb-3ef140a0fb09_1433x1064.png) Chart 2 [Source](https://www.richmondfed.org/research/national_economy/macro_minute/2025/seeing_double_an_ai_bubble) Big as it is, the AI investment boom is almost trivial compared with some earlier examples of technology-driven booms. I began this primer with the case of British railway mania in the 1840s. Here’s what that investment boom looked like. It peaked at more than 7 percent of British GDP, far above the share of GDP spent on AI investments now: s_!rloq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fba82d734-4f44-4ba2-912b-a6a8bcbc49ba_1158x704.png)

Chart 3

Source

During the 19th century there were multiple railway booms across many countries with Britain’s 1840s Railway Mania being the most extreme. Many of them shared a trajectory similar to the one in Chart 3, in which a spectacular boom turned into an equally spectacular bust. And as Chart 1 showed, the 90s telecom boom also suffered a gigantic bust.

These busts not only devastated the sector that they were located in, but also inflicted major damage on the economy as a whole. I will discuss the aftermath of a burst tech bubble in the last part of this primer.

The bursting of a technological bubble wipes out investors. Chart 4 shows the fall in the NASDAQ as the 1990s dot-com bubble burst:

![A graph showing the growth of the stock market

AI-generated content may be incorrect.](https://substackcdn.com/image/fetch/\(s_!hYNN!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F885628a6-392e-4986-8983-398c2c5aab5f_800x450.png) Chart 4 This 70 percent decline in asset values was roughly equal, in percentage terms, to the crash in railway stocks that followed the end of Railway Mania: s_!10ua!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdd4d6276-47c1-4e35-8cdd-dc549ad3b4a5_1032x892.png)

Chart 5

Source

These historical examples should give pause to those spending a tidal wave of money on AI. And some, most notably Apple, are clearly hanging back and investing strategically. However, the vast sums of money being committed to AI show that most investors have convinced themselves that “this time is different.”

So let’s talk about how technology mania and bubbles happen.

The logic of technology manias and bubbles

Yale’s Robert Shiller wrote the book on financial bubbles — literally. His Irrational Exuberance has aged very well over the years. So has his later book, Narrative Economics, in which he analyzes how stories often drive economic behavior.

The most important lesson I took from Irrational Exuberance was that asset bubbles are natural Ponzi schemes. Once an asset’s price starts rising, regardless of the reason, people who bought early make a lot of money. This fact draws in additional investors, who drive the price higher, regardless of the underlying value of the asset. This process continues, independent of the fundamentals, until either the supply of credulous investors is exhausted or something disturbs the narrative.

Like Ponzi’s original scheme, natural Ponzi schemes are sustained by narratives about how investing in this particular asset will somehow be extremely profitable.

It’s a feature of how the human mind developed that it is story-driven. My late mentor Rudi Dornbusch understood this fact, telling me that when one writes for normal human beings as opposed to economists, you don’t begin a paper “Consider a small open economy.” Instead, you begin, “In Belgium …”

So bubbles are more likely to develop and persist for a time when there’s a good story — a narrative — to rationalize high valuations. Ergo, it should come as no surprise that Sam Altman spends an enormous amount of time and energy touting the infinite wonders of AI. Yet, it’s also true that at the time each mania started, the technology underlying it was genuinely innovative and amazing.

For example, railways in the 1840s were transformative for both freight and passenger travel. A stagecoach from London to Manchester took around 4 days, while a steam train took little more than 4 hours.

I’ve used this clip before, but the ads Qwest ran at the peak of the internet bubble, celebrating the wonders of fiberoptics, were almost completely accurate:

Embedded media

You can, in fact, check into a dreary hotel in the middle of nowhere and watch a huge selection of TV shows and movies on your tablet.

And as I’ve already said, generative AI really has suddenly punched through barriers that stymied artificial intelligence researchers for decades. None of this newsletter is written by AI (but if I were an AI bot, I would say that, wouldn’t I?). But I do use AI to produce first drafts of the transcripts for my Saturday video interviews, which are then edited by a human being.

Story-driven technology manias aren’t the only source of asset bubbles. For example, the real estate bubble of the 2000s was not the result of a paradigm-shifting technological change. Yet one can say that the introduction of seemingly clever algorithms that supposedly reduced the risk of mortgage lending led to the flood of sub-prime lending and, eventually, the housing bubble.

The “Winner Take All” Nature of Technology Bubbles

In addition to the compelling narratives, there is another factor that drove the railroad and telecom bubbles to extreme heights: the race for monopoly rents. You see, railroads and telecom fiber have a similar feature: they are based on a network. And in a network, in order to capture monopoly rents you want to get there first and build the greatest capacity, thereby deterring a competitor from entering the market. A similar dynamic is likely to be true with Large Language Model (LLM) AI: according to this narrative, the one or two AI companies with the most comprehensively trained AI generators will vanquish the lesser companies and capture the vast majority of the AI market. In other words, the narrative compels AI companies to engage in a spending race, to become the biggest LLM AI company in the world.

As a result, individual investors who are infatuated with AI will drive up the price of companies that invest heavily in it while shunning companies that don’t. Since corporations care about their stock price, this does shape their strategy.

I’d also suggest that there’s an asymmetry of risks for corporate leaders that leads to everyone making similar investments. Suppose you run a major tech company, invest heavily in AI, and it turns out to have been a bubble. You will suffer a reputational loss, but it will be mitigated by the fact that “everyone did it.” Suppose, on the other hand, that you stayed out of the gold rush and missed a big payoff. That would probably mean losing your job. So better to run with the herd.

To take a concrete example, Apple stands out among big tech companies for taking a much more cautious approach to AI. This may well be the smart move. But the company is getting a lot of grief over “falling behind,” and its stock is down slightly for the year while Microsoft is up 22 percent.

One last point: Why aren’t these bubbles prevented by smart investors who take the long view based on fundamentals, and stop the bubble in its tracks by selling the assets short? The answer was laid out in a classic paper by Andrei Shleifer and Robert Vishny, “The limits of arbitrage.” According to Schleifer and Vishny, it’s a matter of size: potential short-sellers who want to bet against a bubble have limited resources – resources that are not nearly big enough to deflate something as big as the 1990s bubble. Moreover, short sellers rely on credit to fund their trades, which creates another vulnerability – while the bubble continues to grow, margin calls may force them to abandon their positions even if they are eventually proved right. So bubbles keep happening because, once started, they are extremely hard to stop.

What will happen if the AI boom goes bust?

Over the past few weeks warnings are getting louder that AI asset prices are in bubble territory. There are several reasons for the alarm. One is that the evidence so far shows that businesses that are making use of AI are experiencing mixed results – not the kind of overwhelming outcome that AI enthusiasts have predicted.

Second, financial concerns are now arising. Capital expenditure on AI has been mostly undertaken by the “Fabulous Five” – Amazon, Alphabet (Google), Microsoft, Meta (Facebook) and Oracle – companies that earn such huge profits that until recently they were able to finance their AI investments from their free cash flow. But now the spending demands are so large that the Fabulous Five have begun to issue a lot of debt. This clearly increases the financial risks.

Third, there is what Azeem Azhar calls a “financial ouroboros” now taking place in AI: what looks like revenue generated by sales is, in some cases, really just the same stock of money going in circles between the various AI companies. A recent example of this is the deal between OpenAI and Nvidia: Nvidia agrees to invest \(100 billion in OpenAI, which then buys a lot of Nvidia chips with cash it gets by selling Nvidia stock. Granted, Azhar argues that this is a lot less of an issue than it was in the 90s, because there isn’t yet an issue comparable to the “dark fiber” of the telecom era — fiberoptic cable installed with borrowed money, then sitting unused because demand hadn’t caught up with supply. But given the pace at which capacity is expanding, one can’t dismiss concerns that circular flows of money are a warning sign. Fourth, the AI mania has, until now, overlooked a critical bottleneck: where will they get the electricity needed to power the hundreds of billions of dollars’ worth of data centers? In my mind, this may be the most compelling challenge to the AI narrative. Across the country power generators and power grids are being [stressed by the demands of AI](https://www.wsj.com/business/energy-oil/ai-data-center-boom-spurs-race-to-find-power-87cf39dd?mod=article_inline). Consumer electricity prices [have gone up](https://www.bloomberg.com/graphics/2025-ai-data-centers-electricity-prices/) as data centers add significant new demand to the grid, becoming a political football pitting consumers against data centers. And a chicken versus egg problem is appearing in negotiations between electricity providers and data centers, with electricity providers demanding guarantees of future sales before committing to pay the upfront cost of new power sources, and data centers unwilling to make those guarantees because they have yet to make profits. Do I know that AI is a bubble about to burst? No. The parallels with the 90s, especially the hype and frenzy, are obvious. On the other hand, look back at Chart 2. The AI boom is closely tracking the telecom bubble. At the comparable point in that cycle, the bubble still had years to go before it burst. If and when the bubble bursts, however, the consequences for the economy as a whole will be extremely ugly. The collapse of Britain’s Railway Mania was followed by a [severe financial crisis](https://libertystreeteconomics.newyorkfed.org/2015/06/crisis-chronicles-railway-mania-the-hungry-forties-and-the-commercial-crisis-of-1847/). A comparison with the aftermath of the 90s bubble is less alarming, but this comparison may paint an excessively optimistic picture. Look back again at Chart 1. The recession caused by the tech bust — shown by the shaded bar in 2001 — was relatively short and shallow. But recovery was slow: employment took 4 years to fully recover: s_!_spC!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6e31f6bb-1ae9-42a0-baa8-dd78de7eb566_800x450.png)

Chart 6

Furthermore, recovery from the burst tech bubble depended to a large extent on the creation of another bubble, this time in housing. And when that bubble burst, the result was the worst economic slump since the Great Depression.

One last point. In the late 1990s the U.S. economy was growing quite rapidly, around 4 percent a year, and the tech boom was only one factor in that growth. These days underlying growth is only around 2 percent a year, partly because of slow growth in the working-age population, partly because overall productivity growth is slower. And so far this year the AI boom has accounted for almost all of the growth. So if that boom goes bust, the consequences for the economy would be much worse than the bust at the end of the 1990s.

Predictions are hard, especially about the future \[/Yogi Berra\]. I am not confidently predicting an imminent AI crash. But even though we’re talking about a new, unprecedented technology, there are many historical parallels to our current situation. And they give us good reason to be concerned.