Is AI really a bubble? A practical framework for answering the biggest question in the industry
God Translation Bureau is a compilation team under 36Kr, focusing on fields such as technology, business, workplace, and life, and mainly introducing new technologies, new ideas, and new trends abroad.
Editor's note: Everyone is talking about the AI bubble, but no one can explain it clearly. This article uses a framework with five indicators to tell you: It's not a bubble now, but what are the real danger signals? This article is from a compilation.
A month ago, I started to answer a seemingly simple question: Is artificial intelligence a bubble?
Since 2024, when I give speeches at events around the world, people have been asking me this question.
Although Wall Street bankers generally believe it is an investment boom, more and more people in meeting rooms and convention halls in Europe and the United States are starting to raise this question.
Some people have made up their minds. Gary Marcus calls it "the peak of the bubble." The Atlantic warns that "we may currently be experiencing an artificial intelligence bubble, and investors' excitement has far exceeded the technology's near - term productivity benefits. If this bubble bursts, the consequences could dwarf the collapse of the Internet bubble - and it won't just be the tech giants and their Silicon Valley supporters who suffer losses." The Economist says, "The potential cost is already alarmingly high."
The best way to understand such questions is to build a framework, a framework that can be continuously updated as new evidence emerges. It took me dozens of hours to build this framework, including data analysis, modeling, and numerous conversations with investors and executives.
This article is that framework: Five indicators to measure generative AI against historical bubbles.
It's a bubble if two indicators enter the red zone.
I have studied and personally experienced the Internet bubble. As an investor, I felt its impact firsthand. Like many of you, I was also active during the global financial crisis. Therefore, I have invested a lot of thought and analysis to build a robust framework to understand what is currently happening. Today, I will share it with you.
My thinking is based on the research of giants in the field such as Carlota Perez and Bill Janeway, as well as the financial analysis of major banks and analysts. But I haven't seen a framework elsewhere that can translate these theories into a practical dashboard for today's artificial intelligence: a set of parameters that can be read, compared with past bubbles, and used to guide decision - making.
What a bubble means
A bubble is one of the oldest stories in capitalism. Bubbles are fables about excess, belief, and collapse. But bubbles are not just financial phenomena; they are also cultural products. As moral stories about greed and stupidity, they always reappear again and again. The tulip mania is often wrongly described as a madness that bankrupted weavers and ruined merchants, but its disaster was far less than the legend. The damage it caused was limited to wealthy merchants and had little impact on the Dutch economy. But the myth has survived, and that's the key: Bubbles have become stories we use to warn ourselves of the danger of being overly optimistic.
Jan Brueghel the Younger, "Allegory of the Tulip Mania", 1640
Some bubbles are financial: the South Sea mania in the 1720s, the roaring stock market in the 1920s, the real - estate boom in Japan in the 1980s, and the housing market crash in 2008. Some are technological. In the 1840s, railways were hailed as the blood vessels of the new industrial body. And they were. But a body only needs so many blood vessels, and soon the railway tracks were laid in places where business couldn't support them. In the 1990s, the telecommunications industry promised to create a network utopia, but in the end, 70 million miles of redundant fiber optic cables were left idle underground. The Internet boom gave us a vision of a new economy, and most of it did eventually come true, but only after the valuation evaporated in 2000.
Interestingly, the academic community doesn't seem to have a consensus on what an investment bubble is. Nobel laureate Eugene Fama even says that bubbles don't exist at all.
I hope to go beyond the view that "you only know a bubble when you see it." There are two interrelated systems at work here. The first, and most obvious, is when the stock market is absurdly overvalued and then collapses. The second is whether the deployment of productive capital (invested in capital expenditure or venture capital) collapses. Of course, the two are related. A sharp drop in stock prices will make the cost of investment flow higher. And a long - term decline in productive capital investment may be interpreted by the stock market as a sign of an economic slowdown.
But to frame this question: A bubble can be defined as a situation where the stock value drops by 50% from its peak and lasts for at least five years. In the cases of the US real - estate bubble and the Internet bubble, the trough period lasted about five years. It took 10 years for the US real - estate market to fully recover to its pre - bubble peak, while the Internet took 15 years. Meanwhile, we expect the deployment rate of productive capital to also decline significantly, also by 50% from its peak.
It took 10 years for the US real - estate market to recover after the bubble burst.
In this analysis, I will use the term "bubble" to cover these two related aspects. Ultimately, it refers to a stage where prices and investment soar rapidly, during which valuations deviate severely from the potential and actual profitability of related assets. Bubbles thrive in an environment of abundant capital and attractive narratives and often end with a sharp and sustained reversal, erasing most of the paper wealth created during the upward phase.
In contrast, a "boom" may look very similar to a bubble in the early stage, with rising valuations and accelerating investment. But the key difference is that during a boom, the fundamentals will eventually catch up. The potential cash flow, productivity improvement, or real demand growth will rise to match the optimistic sentiment. There may also be over - investment during a boom, but it will eventually be integrated into a lasting industry and long - term economic value.
However, there is a gray area between the two: During these enthusiastic periods, it's really hard to tell whether the capital is laying the foundation for a new economy or just inflating unsustainable prices. It's like being in the eye of a storm: You can feel the wind, rain, and pressure, but you don't yet know whether the storm will clear the haze or flatten the houses.
This leads to our current question: Is artificial intelligence another bubble? What makes many observers uneasy are the numbers. Since the release of ChatGPT at the end of 2022, the annual data - center capital expenditure of hyperscale cloud service providers has more than doubled. They are heavily betting on the infrastructure needed to train and run increasingly large models. Last year, when Sam Altman proposed a $7 - trillion investment demand, it sounded like hubris. Now, investors are no longer laughing it off; they are starting to think about whether such a scale of expenditure is sustainable.
Since the launch of ChatGPT, the capital expenditure of hyperscale infrastructure providers has doubled.
Bubbles can't be diagnosed in real - time. We can only be wise after the event to tell whether the previous mania was reasonable or delusional. Instead of making a definite judgment - which is almost a useless guess - we should compare today's boom with history to find out where signs of a similar bubble may quietly appear.
We can imagine this as flying a plane. Pilots don't rely on just one instrument; they monitor a set of instruments to understand the flight situation. We will use five such indicators:
Indicator 1 - Economic pressure: Is the current investment scale large enough to distort the economy?
Indicator 2 - Industry pressure: Is the industry revenue commensurate with the deployed capital expenditure?
Indicator 3 - Revenue growth: Is the speed of revenue growth/expansion fast enough to catch up?
Indicator 4 - Valuation heat: How high are the valuations? Are stocks overvalued compared to history?
Indicator 5 - Capital quality: What kind of capital is supporting all this? Is it from a robust balance sheet or fragile and volatile capital?
Next, I will examine each indicator one by one, explain why it may be green, yellow, or red, and finally integrate them into a complete dashboard view.
Indicator 1: Economic pressure
The scale of the ongoing investment is huge. Morgan Stanley expects the expenditure on artificial - intelligence infrastructure to reach $3 trillion by 2029. But it hasn't reached the out - of - control extremes before the bursting of those major historical bubbles. What makes this dimension tricky is its dependence. In the United States, more than one - third of GDP growth can be traced back to the construction of data centers.
This is not necessarily a bad thing in itself, but it could become dangerous if the momentum weakens. An economy so heavily dependent on the growth of a single industry may find its foundation collapsing faster than expected.
The surge in capital expenditure (capex), that is, the investment in the physical infrastructure needed for AI, is an act of optimism. That's the nature of capital expenditure. Spending money today is based on the belief that it will become a source of income tomorrow. If today's investment is wise, it will eventually bring about productivity improvement and economic expansion.
If you want to sell hula - hoops by promoting health and happiness, you first have to buy hula - hoops. And at some point in the supply chain, there needs to be a factory to manufacture them. Financing that factory is capital expenditure. You deploy capital expenditure to build something useful that can be sold. Capital - expenditure investment is usually, but not always, a precursor to business growth.
AI data centers have a similar nature, and even more so. They are not just factories that produce a single product; they are infrastructure. Microsoft, OpenAI, and the US government all hold this view. They regard computing power as the infrastructure of the 21st century, as important as the early highways, railways, power grids, or telecommunications networks. I also made the same argument in my previous book, so you can guess my stance. In addition to the United States, sovereign governments around the world have promised to invest more than $1 trillion in AI as infrastructure by 2030.
Building such infrastructure inevitably requires a historically huge amount of funds, comparable to the scale of past railway and power - grid construction. McKinsey predicts that by 2030, about 156 gigawatts of new power capacity will be needed to meet AI - driven demand - which is equivalent to the combined power grids of Spain and Portugal. The company estimates that this will cost $5 trillion to $7.9 trillion in capital expenditure. For reference, this is about twenty times the cost of the Apollo program, making AI data centers one of the largest infrastructure - construction projects in modern history.
However, although the infrastructure is useful, things may deviate from reality when private capital gets involved. The financing structure is as important as the technology itself. The railways in the United States were mainly privately funded and experienced multiple investment bubbles. In contrast, the power and highway systems benefited from more public investment and coordination and had less excessive speculation. When the resources demanded by a boom start to distort the entire economy, it becomes dangerous. Wages are drawn into one industry, the supply chain is re - adjusted, and the capital market becomes dependent on it. When expectations are dashed, the rebound will be very strong.
One way to measure economic pressure is to look at the proportion of investment in GDP, that is, the part of national output invested in a single technological frontier. This is a relatively rough but telling ratio. It shows the extent to which the economy depends on a single technological bet.
In the case of AI infrastructure, most of the expenditure comes in the form of capital expenditure: servers, cooling systems, network equipment, reinforced - concrete enclosures, and the power infrastructure needed to keep them running.
By this standard, the railway bubble was the most burdensome. In the United States, railway expenditure reached its peak in 1872, accounting for about 4% of GDP, just before the first major crash. In contrast, the peak of the telecommunications boom in the late 1990s was close to 1% of GDP, a level that seems familiar today.
AI construction is in this middle ground. It is estimated that about $37 billion will flow into data centers globally in 2025, with about 70% earmarked for the United States, which is about 0.9% of US GDP. Goldman Sachs expects the expenditure to increase by another 17% in 2026. My own prediction is in line with this view: By 2030, the annual capital expenditure will reach $80 billion, with about 60% occurring in the United States, which will bring the US share to 1.6% of the 2025 GDP.
The economic - pressure indicator is divided into three ranges: green (below 1%), yellow (1% to 2%), and red (above 2%). So for now, generative AI is just in the green zone. Of course, considering the announced investment commitments, it seems likely to enter the yellow zone soon.
But there is a variable that historical bubbles didn't face: rapid depreciation. Once a railway track is laid, it can be used for decades. The United States was still using railway tracks built in the 19th century to transport goods in the 20th century. The telecommunications fiber optic cables laid in the 1990s were still transmitting data 35 years later.