Stock prices in the AI sector have gone down a lot, but expectations for how much money AI will eventually make are still higher than ever. That difference, maybe more than any other piece of information, is what makes this moment so risky.

A Story Under Stress

The AI boom, which has been the main story in the market for the past two years, is starting to show real cracks. And a question that would have sounded crazy in mid-2024 is now the most important one in financial circles: has the AI "bubble" already burst? John Higgins, the chief economist at Capital Economics, says that the answer is a qualified "yes," at least from one important point of view.

If we use tech sector valuations to define the "AI bubble," the picture is very clear. Since its peak in late October, the sector's premium over the broader market has dropped a lot. The price-to-earnings (P/E) ratios for the tech sector are at their lowest levels since the pandemic. The decline is not just happening to software companies, which were already having trouble because people were worried that AI would take over their main business functions. It has spread to semiconductor and hardware companies, which is the very "picks and shovels" level that many investors thought was safe from these kinds of worries. At the height of the dot-com bubble, the four biggest tech companies traded at about 70 times their two-year forward earnings. Today, the average two-year forward P/E for the four biggest AI data center spenders—Microsoft, Alphabet, Amazon, and Meta—is about 26 times. Yes, lower, but still higher than usual for the time.

In the meantime, the Magnificent Seven stocks: Apple, Nvidia, Microsoft, Amazon, Tesla, Alphabet, and Meta Platforms make up about 35% of the S&P 500's total market cap. This is a level of concentration that has few historical precedents and increases systemic risk if sentiment changes.

The DeepSeek Quake

The arrival of China's DeepSeek in January 2025 was the best example of how fragile AI-sector valuations are. On January 27, 2025, US stocks fell sharply, and chipmaker Nvidia lost almost $600 billion in market value after DeepSeek showed off a ChatGPT-like AI model that works for a lot less money than OpenAI's, Google's, or Meta's popular AI models. This was the biggest one-day market value loss for any stock in US financial history. Dan Ives, an analyst at Wedbush Securities, said that DeepSeek's development cost was only $6 million. This shocked Wall Street because American hyperscalers usually spend hundreds of millions or billions to build similar systems.

The effects were felt far beyond Nvidia. Micron Technology and Arm Holdings fell 10%, ASML fell 6%, Microsoft fell 2%, and Alphabet fell 4%. Energy companies, which had seen their stock prices rise because they thought they would be able to provide power to AI data centers, were hit hard. Constellation Energy, which had made a deal with Microsoft to restart the Three Mile Island nuclear plant for AI power, saw its stock price drop 21%. Vistra saw its stock price drop 28%, and GE Vernova saw its stock price drop 21%. The message was harsh and clear: AI infrastructure spending that was based on the idea that computational needs would always grow was suddenly open to serious challenge.

Later, Nvidia CEO Jensen Huang said that the market had completely misunderstood what this meant. Huang said, "I think the market reacted to R1 as if it were the end of AI." He said that the next step in scaling would be to teach AI models to reason better, which would still require a lot of computing power. His argument was valid, but it didn't change the fact that a single Chinese startup, which was limited by US semiconductor export rules, had shown how weak the story of "infinite, ever-escalating compute demand" really was.

The Big Investment Bet and the Math That Goes With It

Valuations are under pressure, and the amount of money being spent on infrastructure is now much higher than it has ever been in previous technology cycles. Morgan Stanley analysts think that by 2025, Amazon, Microsoft, Google, and Meta will spend 34% of their sales on capital expenditures, and by 2028, this number will rise to 37%. At the height of the dot-com bubble, the number was about 32%. Morgan Stanley says that Big Tech companies will spend about $3 trillion on AI infrastructure by 2028, but their own cash flows will only cover about half of that.

The individual promises are huge. Amazon has promised to spend $100 billion on data centers in 2025 alone. Meta has promised to spend more than $600 billion over the next three years. Microsoft planned to spend $80 billion in 2025, Google promised $75 billion, and Apple plans to spend $500 billion over the next four years, including on AI. OpenAI, a company that is still losing a lot of money, has promised to spend more than $1 trillion on AI infrastructure, but it only expects to make about $13 billion in 2025.

In the first half of 2025, AI-related capital expenditures were the main driver of economic growth, accounting for 1.1% of GDP growth. That is an unusual and historically significant event, and it raises an equally unusual question: what happens if the money doesn't come in to support it?

The fact that these investments are becoming more and more circular is also a cause for concern. Nvidia said it would invest $100 billion in OpenAI. Microsoft owns a large stake in OpenAI and is also a major customer of CoreWeave, an AI cloud computing company in which Nvidia also owns a large stake. Microsoft, on the other hand, made up almost 20% of Nvidia's annual revenue. When investors, customers, and suppliers are basically the same people, the chain of financial logic gets very short.

Some businesses have used off-balance-sheet structures to pay for the buildout. Blue Owl Capital and Meta, two Wall Street firms, recently gave money to a special purpose vehicle for a data center in Louisiana. Blue Owl borrowed $27 billion, which was backed by Meta's lease payments. The debt never showed up on Meta's balance sheet. If the AI bubble pops and the data center goes dark, Meta would have to pay Blue Owl billions of dollars. This is a lot like the off-balance-sheet financing structures that made things worse during the dot-com boom and the 2008 financial crisis.

No Returns and the Problem of Making Money

A more basic problem lies behind the headlines about capital spending: showing that AI really does bring back what is put into it. A well-known study from MIT said that 95% of companies that put money into generative AI were not seeing any returns. The study itself was more complex than the headline made it sound, but for many analysts, it made a worry that had been growing for months clear: that the adoption of enterprise AI is real, but it has been much slower and less consistent than valuations thought.

Hundreds of millions of people use big language models like ChatGPT, but only a small number of them have paid for an advanced version each month. This is a problem that many generative AI companies face: how to turn broad, casual use into long-term profit. OpenAI is worth $500 billion, but it has been spending $3.76 billion a year just on inference costs. This number rose to $5.02 billion in the first half of 2025. Former Fidelity manager George Noble said that OpenAI is losing $15 million a day on its Sora video generation product alone.

According to CB Insights, more than 1,300 AI startups are worth more than $100 million, and 498 of them are "unicorns," which means they are worth $1 billion or more. But Bernstein analyst Richard Bernstein says that the difference between investment and expected earnings that are believable is what makes a bubble. He pointed to OpenAI, which has already made about $1 trillion in AI deals, including a $500 billion data center buildout, even though it is only expected to make $13 billion in revenue. This is a difference that many people see as "certainly looks bubbly."

"Yes, there is a bubble, but it isn't over yet."

But the picture isn't one-dimensional. There are serious and credible voices saying that this situation is very different from previous bubbles, and that the bear case doesn't take into account what is really being built.

Ben Schindler of Goldman Sachs says that the S&P 500's long-term performance is mostly based on earnings, not valuations. He also says that the index's gains in 2025 were mostly due to real earnings growth. Larry Fink, the CEO of BlackRock, said that the spending was not too much but necessary for the company's strategy: "Investing in AI isn't just buying GPUs and chips; it's also buying HVAC systems, IT systems, power grids, and power supplies." He was sure that the biggest hyperscalers would come out on top. In December 2025, JPMorgan used a five-factor diagnostic framework to look at the AI rally and found that the sector has real structural utility instead of just speculation. They found that capital inflows are directly linked to measurable business growth and revenue generation.

Analysts from BlackRock and Goldman Sachs say that the median cash flow and capital reserves of the top 500 US firms are about three times higher than they were during previous bubbles. They also say that the current market leaders have stable earnings streams and strong margins, which is different from the revenue-negative valuations of the late 1990s. In 2001, the tech leaders of the dot-com era saw their annual net income drop by 65%. Now, everyone agrees that the hyperscalers will make 17% more money over the next year, which is a very different starting point.

Citigroup analysts say they can see signs of a "bubble," but they think there is still enough liquidity to keep it going. They say that it might be a bubble, but it hasn't burst yet. The most important question is what happens when momentum changes. Citi says that a "burst" could cause a lot of trouble, a recession, and aggressive interest rate cuts by central banks. At that point, investors would probably move into defensive sectors like healthcare and consumer staples, as well as sectors that are sensitive to interest rates, like utilities and real estate.

The Geopolitical Aspect

The AI bubble isn't just a financial thing. It's at the crossroads of technology, national security, and global competition, and these factors add levels of risk that stock prices alone can't show.

China's technology ecosystem gets a lot of different kinds of funding, like state investment, subsidies, and private capital. Innovation happens in universities, small and medium-sized businesses, and hedge funds. Chinese companies don't promise revolutionary breakthroughs; instead, they focus on small, practical improvements and open-source models that are easier to adopt and less expensive. If an AI bubble pops in the West, this strategy—less leveraged and less reliant on speculative private valuations—could help China make a lot of progress.

The Bank of England has warned that the risks of a global market correction are rising because leading AI companies may be overvalued. They also said that investors are not being properly warned about the risks if AI does not meet market expectations. The IMF has compared the current levels of valuation to the excitement over internet companies that led to the dot-com crash in 2000. S&P Global Market Intelligence says that the bursting of an AI bubble could cost the US tech sector 2.5 million jobs. This shows how deeply the AI story has become part of the real economy, not just the stock market.

The Bubble Model with Two Layers

The most intellectually accurate characterization of the present circumstances is that there exist two distinct layers, each in a different stage of development.

The first layer, which is stock price valuations, has already changed in a big way. P/E ratios have dropped from their highs. The DeepSeek shock was a violent reminder that the assumptions behind those valuations were always more fragile than they seemed. In this way, the valuation bubble has at least lost a lot of its air.

The second layer, which includes expectations for earnings and commitments to capital expenditures, is almost completely unchanged. In fact, it has continued to rise even as valuations fell. If the market for AI were to keep growing at the same rate, the industry would quickly build too much capacity, the debt would become worthless, and banks would lose a lot of money. This is similar to what happened with fiber optics in the late 1990s.

When price bubbles pop, they cause sharp but short-lived corrections. When expectation bubbles burst, the effects are deeper and last longer than when they burst. These bubbles are based on wrong assumptions about things like corporate strategies, capital structures, government policies, and employment levels. The market may have already taken the warning about the valuation into account. A much less comfortable question is whether it has started to factor in the risk of earnings expectations.

As OpenAI CEO Sam Altman said, "Are we in a phase where investors are too excited about AI?" I think so. Is AI the most important thing to happen in a long time? I also think yes. It's hard to read this moment because both parts of that sentence can be true at the same time.

The honest answer to the question "Has the AI bubble burst?" is the easy part may already have happened. The hard part may still be ahead.