Wall Street has spent the past year pricing in an artificial-intelligence revolution. Now investors are starting to question whether the boom has gotten ahead of reality. Calling the AI boom a bubble may sound like one means artificial intelligence is fake or doomed to fail. But it actually means something far narrower. The price of AI-related assets may have climbed higher than the profits those assets can reasonably be expected to justify. On that definition, the evidence now points toward a bubble in AI investment.

The technology will still matter. Some people will get rich from AI. Consumers will get useful services, scientists will get better research tools, and companies will find more efficient ways to produce goods and services. But none of that responds directly to the valuation question. A great invention can still be a bad investment if buyers pay too much for a claim on its future profits.

A Revolution Can Still Be Overpriced

History is full of inventions that fit this pattern. Railroads transformed America, but many railroad investors lost money after the railroad bubble burst in the 1890s. The internet changed the nature of commerce, but most dot-com shares bought near the late-1990s peak were terrible investments. In other words, a technology can be useful, and even transformative, while the stocks and private-company valuations attached to it are too expensive.

That distinction is important because AI optimists often answer valuation concerns by pointing to what the technology can do. The demos are indeed impressive. But investor concern is separate. Will today’s owners of data centers, chips, and AI labs earn enough cash flow to justify today’s sky-high prices?

If the answer requires unusually fast adoption, unusually strong pricing power, or unusually cheap capital, the discussion is already about bubble risk. That is where the evidence points today.

The Costs Are Here. The Returns Are Not.

The evidence is hard to ignore. The Financial Times recently reported that Amazon, Alphabet, Microsoft, and Meta are on course to spend $725 billion on AI infrastructure. That spending splurge would push their combined free cash flow—the cash left over after expenses and investment—to a decade low. The report projected combined free cash flow of $4 billion in the third quarter of 2026. That’s compared with a post-pandemic quarterly average of $45 billion.

Put plainly, some of the world’s most cash-rich software companies are starting to look a lot like heavy industry. No longer are they just writing code. They are buying chips, land, power, and data centers at enormous scale.

Meta is a case in point. The Associated Press reported that Meta raised its 2026 capital-spending forecast to $125 to $145 billion, largely because AI chips and memory have become more expensive. Data center costs also added pressure. Meta’s stock fell after hours after the announcement even though revenue beat expectations.

Wall Street may eventually reward that spending. But the old software business model was far simpler. It involved writing code once, then selling each additional copy of software at little extra cost. AI operates very differently. It requires enormous upfront spending on physical infrastructure, with the hope of earning monopoly returns later. The payoff may come, but it’s a higher bar than software companies once had to clear.

Revenue Is Still the Hard Part

Sequoia Capital put the issue pointedly in 2024 when it wrote about AI’s $600B Question . VC investor David Cahn argued that the infrastructure buildout implies revenue expectations far above actual revenue. Since then, the spending side has grown faster than the returns.

OpenAI illustrates the strain pinching the industry. The Financial Times reported this spring that OpenAI secured up to $110 billion in funding at a $730 billion valuation. The same reporting tied that funding to huge compute commitments, including a $100 billion Amazon agreement and about $600 billion in compute commitments through 2030.

A valuation that large requires an extraordinary outcome. OpenAI would need to become one of the most profitable companies in history while also paying for some of the most expensive infrastructure in history. That could happen. But it is an enormous undertaking.

Anthropic offers the strongest reason for optimism. The Financial Times reported that the company expects a profitable quarter in 2026, with revenue above $10 billion in the second quarter. That is real evidence of demand. The same reporting suggested Anthropic’s valuation could reach $900 billion, while the company recently signed a $15 billion annual compute commitment to SpaceX. A profitable quarter is an important milestone, but it does not prove that today's prices for the whole AI sector make sense.

Tech companies have to show that AI pays for itself. A study discussed in Tom’s Hardware found that 95 percent of enterprise generative AI deployments had no measurable impact on profit and loss. A survey reported by TechRadar found that many senior leaders saw no clear productivity gain despite broad AI adoption. These findings may improve as AI tools get better and firms learn how to use them. For now, they weaken the case that current spending is firmly tied to fundamentals.

The Market Is Recycling Its Own Money

Another warning sign is circular financing. That means a supplier helps fund the customers or startups that buy from the supplier. MarketWatch reported that Nvidia put $18.6 billion into private, nonmarketable equity securities in only three months, much of it tied to AI startups and infrastructure firms. Its nonmarketable equity holdings rose to $42.3 billion from $3.2 billion a year earlier.

The Financial Times described a broader Nvidia investment spree of roughly $90 billion across more than 145 companies in 16 months. In other words, Nvidia has moved from selling shovels in a gold rush to helping finance the miners, some of whom then turn around and buy its shovels.

The International Monetary Fund has warned that circular AI financing can inflate revenues and valuations by tying buyers, suppliers, and investors together in an artificial manner. These structures are not proof of any wrongdoing. They do, however, make underlying fundamental demand harder to see. A sale financed by the seller is still a sale, but it is weaker evidence of demand than spending funded by a customer's own operating cash flow.

The temptation to rely on vendor-financing is easy to understand. The Associated Press reported Nvidia quarterly revenue of $81.62 billion and noted that its market value had climbed from about $365 billion at the end of 2022 to $5.4 trillion. Nvidia may remain an excellent company. The bubble question depends on whether the rest of the AI ecosystem can earn enough profits to make all those chip purchases justified.

Luck Will Choose Many Winners

Economists often say the social return to innovation is larger than the private return captured by those who fund or create it. In plain English, the benefits of invention spill far beyond the inventors. One firm may publish the research. Another may build the tools. A third may find the profitable business model. Competition pushes down prices for consumers, while workers get bid away from the leading firms. For all these reasons, the original investors may help change the world but still earn poor returns.

That is one reason basic research often requires public support. The people who ultimately benefit may have had far less to lose than the earliest supporters, which makes investors less likely to invest in the first place.

AI may follow this pattern. Model improvements may commodify quickly. Open source models may put pressure on prices. Chips and servers depreciate. Data centers built for one generation of models may need expensive upgrades for the next. The eventual winner may emerge only after today’s investors have absorbed the costs of developing the technology.

Someone will become a billionaire. Later, that success will probably be told as a story of vision and genius, because markets like clean narratives. But the outcome may depend as much on luck and timing as foresight. The winning founder may arrive after compute costs have fallen, after users have learned what they actually want, or after earlier firms have failed and revealed the business model the works.

Taxpayers Helped Pave the Road

The public sector is already part of the AI boom. NSF says it has funded AI research since the early 1960s, helping build the technical foundations for today’s boom. More recently, NSF’s National AI Research Institutes have supported university-based AI research around the country since 2020. The CHIPS and Science Act put tens of billions of dollars behind semiconductors and research capacity. The official White House AI portal emphasizes there will be further federal support for AI innovation.

That public role can make sense. Basic research produces knowledge that no single company can fully capture to profit from. The payoff may not appear for many years, but it eventually emerges in forms that pay for the initial investments many times over.

The awkward part is the distribution of outcomes. Taxpayers help fund the knowledge base at the riskiest stage. Yet they do not receive common stock in the private company that eventually wins the competition. They may receive broad social benefits, but they do not get the venture-capital return. If an AI fortune grows out of a publicly supported knowledge stack, much of the financial upside will be privatized.

That does not mean government should stop funding AI research. It means taxpayers should have a way to share in the upside when publicly supported research helps create private fortunes.

Both sides of the AI debate have a point. Skeptics are right to question the fundamentals. Booms often pass through busts before reaching financial sustainability. But optimists are also right that AI has real promise. It could transform the economy even if today’s valuations and investment plans prove too aggressive.

The clearest sign of fragility is the gap between spending and returns. Capital expenditures are surging while free cash flow is weakening. Private valuations are outrunning durable profits, and many companies are still struggling to show clear returns on their AI investments. Financing loops are becoming more common. These are classic warning signs.

AI does not have to fail as a technology for the investment outlook to be grim. The most dangerous manias often form around technologies that really do work. Investors see the future coming, then rush too far ahead of it—paying too much, too early, for the wrong opportunity.

By the valuation test that matters, AI investment looks overheated. The technology may succeed. Someone will become very rich. The economy will benefit. But today’s prices assume that the companies spending the most will capture most of the eventual gains.

History gives reasons to doubt that. The people who lay the foundation often do not get to own the penthouse. Taxpayers helped finance the science. Early investors are now laying the brick. But the fortune may go to someone who arrives later, after costs fall and the business model becomes clear. When that happens, commentators will call it genius. Instead, much of it will be luck, timing, and the simple fact that markets often reward the last mile more generously than the first.