The artificial intelligence industry has reached a strange point.
The technology is improving quickly. More people are using AI at work, businesses are experimenting with automation, and new models appear almost every week. At the same time, the amount of money being invested has become difficult to comprehend.
AI companies are raising billions of dollars. Technology giants are building data centers that consume as much electricity as small cities. Private companies that are still losing money are receiving valuations that would have seemed absurd only a few years ago.
It is not surprising that people are asking whether AI has become a financial bubble.
A recent argument circulating online claims that the American AI industry has already accumulated trillions of dollars in debt and must replace millions of white-collar workers every year simply to cover its interest payments. (https://www.youtube.com/watch?v=xKuDvRRE7Bc)
The broader concern deserves to be taken seriously. The calculation behind it does not.
There may be an AI bubble, but it is important to understand where the real risks are.
The trillion-dollar number is being misused
One of the most dramatic claims is that between $3 trillion and $4 trillion has already been invested in the American AI industry, mostly using borrowed money.
There is little evidence to support that interpretation.
The Stanford AI Index estimated that private AI investment in the United States reached approximately $285.9 billion in 2025. That is an extraordinary amount of money, but it is nowhere near $3 trillion.
The much larger number appears to come from forecasts of global AI infrastructure spending over several years.
Morgan Stanley, for example, estimated that close to $3 trillion could be invested in AI infrastructure around the world by 2028. Most of that money had not yet been spent when the estimate was published.
A forecast of future investment is not the same as existing debt.
The estimate also includes more than the companies developing AI models. It covers data centers, computer chips, electricity infrastructure, networking equipment, construction and other parts of the AI supply chain.
It is therefore misleading to describe the entire amount as money borrowed by companies such as OpenAI and Anthropic.
AI does not need to replace 10 million workers
The argument usually continues with a simple calculation.
If the industry has around $2.5 trillion in debt and pays 4% interest, it would need approximately $100 billion a year just to cover the interest. At a profit margin of 10%, it would need $1 trillion in revenue.
From there comes the claim that AI must replace about 10 million white-collar employees every year.
The arithmetic works. The assumptions do not.
There is no reliable evidence that AI companies collectively have $2 trillion or $3 trillion in outstanding corporate debt.
The calculation also treats the AI industry as if it were one heavily indebted company with no other businesses or sources of income.
Much of the current infrastructure is being built by Microsoft, Amazon, Alphabet, Meta and Oracle. These companies already generate substantial revenue from cloud services, advertising, software, retail, databases and subscriptions.
Microsoft does not need every new data center to be paid for exclusively by employees losing their jobs to Copilot. Amazon does not need to classify every dollar generated by AWS as separate AI revenue before it can repay its debt.
More importantly, AI can create financial value without eliminating a job.
A company may use AI to answer customer questions more quickly, produce more software with the same team, detect fraud, improve advertising, translate content or reduce the time required to analyse documents.
The employee may remain in the company while producing more work.
It is therefore wrong to assume that the only possible return on AI investment comes from replacing salaries.
The real financial risk is still serious
The fact that the viral debt calculation is wrong does not mean the AI market is healthy.
The real numbers are already large enough to create concern.
OpenAI, Anthropic and other model developers are raising huge amounts of capital while continuing to spend heavily on training, research and computing infrastructure.
At the same time, Microsoft, Amazon, Alphabet and Meta are investing tens of billions of dollars every quarter in data centers and related equipment.
These investments are based on a major assumption: demand for AI services will continue growing quickly enough to justify the infrastructure being built today.
That may happen. It is not guaranteed.
Data centers are expensive to build and operate. They require chips, electricity, water, cooling systems and access to the power grid. The hardware inside them can become outdated surprisingly quickly.
If model efficiency improves faster than expected, companies may need less computing capacity than current forecasts suggest. If businesses decide that AI tools do not provide a sufficient return, some of the new infrastructure may remain underused.
This is where the strongest bubble argument begins.
The problem is not that the AI industry must replace a specific number of workers next year. The problem is that companies are spending enormous amounts today based on profits that may only appear several years from now.
Are AI companies losing money on every request?
Another common claim is that OpenAI and Anthropic lose money every time someone sends a prompt to one of their models.
That is possible for certain subscriptions or extremely heavy users, but it has not been demonstrated as a general rule.
There is an important difference between losing money on an individual request and losing money as a company.
An API request may be sold for more than the direct cost of running the model. The company can still report large losses after paying for research, model training, engineering salaries, sales, marketing and future infrastructure.
In other words, an AI service can have a positive gross margin while the company behind it remains deeply unprofitable.
The more useful question is not whether every request loses money. It is whether current revenue can eventually support the extraordinary cost of developing and operating the next generations of models.
For many frontier AI companies, the answer is not yet clear.
Chinese models are changing the economics
The rapid improvement of Chinese and open-weight models may be an even bigger problem for American AI companies than their operating losses.
A few years ago, it appeared possible that a small number of US companies would control the most capable models. They had access to advanced chips, large data centers and billions of dollars in funding.
That advantage still matters, but the gap has narrowed.
Chinese developers have released models that perform competitively in programming, reasoning and agent-based tasks, often at much lower prices.
Many of these models can also be downloaded and run on private infrastructure. They are often described as open-source, although open-weight is usually the more accurate term. The model weights may be available, while the training data and complete development process remain private.
Their growing popularity creates a serious pricing problem for companies such as OpenAI and Anthropic.
If customers can obtain similar results from a cheaper model, the expensive provider cannot simply increase its prices whenever it needs more revenue.
OpenRouter data has shown a sharp rise in the use of Chinese models among developers using its platform. This does not mean Chinese models already control most of the entire US market. OpenRouter represents only one part of the AI ecosystem.
It does show that developers are willing to switch models when the difference in price or performance becomes attractive.
That makes it harder for any single provider to build a lasting monopoly.
Local AI is improving, but it is not free
The same competitive pressure is coming from models that can run on local computers and private servers.
Local models offer clear advantages. Data can remain inside the company, the system can work without an internet connection, and users are not dependent on the pricing or policies of an external provider.
For common tasks, a smaller local model may be good enough.
That could reduce the number of people willing to pay for premium AI subscriptions.
Still, local AI is not automatically cheaper in every situation. A company must purchase hardware, pay for electricity, maintain the system, install updates and monitor security.
Smaller models may also require more corrections or perform less reliably on complex work.
The most likely outcome is a mixed market.
Routine tasks may be handled by local or inexpensive models. Difficult problems may be sent to more capable cloud systems. Traditional software will continue to handle predictable processes, while people will remain responsible for decisions where errors are expensive.
This would create a large AI market, but not necessarily one with extremely high margins for the companies that build the underlying models.
Productivity has not matched the hype
The most important question is whether AI is creating enough productivity to justify the money being spent.
The evidence so far is mixed.
A study of more than 5,000 customer-support employees found that access to a generative AI assistant increased productivity by about 14% on average.
The largest benefits were seen among less experienced workers. AI helped them access knowledge and techniques that more experienced colleagues had learned over time.
That is a meaningful result.
Other research has been less encouraging.
In a study conducted by METR, experienced open-source developers completed certain tasks more slowly when using AI tools. The developers believed the tools were helping them, but the measured results showed that they took longer.
The difference may come from the type of work being performed.
AI is often useful when a task involves drafting, summarising or applying familiar patterns. It can become less useful when an expert is working inside a complicated system and must carefully inspect suggestions that are almost correct.
The time saved in writing code can easily be lost while finding a subtle mistake.
This helps explain why AI demonstrations can appear spectacular while real business deployments move more slowly.
A demonstration usually shows the best result under controlled conditions. A production system must deal with incomplete information, unusual customers, privacy rules, security risks and legal responsibility.
A tool that is correct 95% of the time may be impressive in a video and unacceptable in a bank.
Why companies are withdrawing AI customer-service systems
A 2026 survey from Sinch found that 74% of the organisations questioned had withdrawn or shut down at least one AI customer-communications agent after it was launched.
That number sounds disastrous, but it needs context.
It does not mean that 74% of all AI agents failed or that most companies abandoned AI. It means that many organisations discovered problems in at least one deployment and decided to stop, redesign or replace it.
The finding is still important.
Customer service was supposed to be one of the easiest areas to automate. Conversations are often repetitive, companies already possess large databases of answers, and the cost of human support is significant.
Yet real customers do not always ask predictable questions. They may provide incomplete information, misunderstand policies or face situations that do not appear in the training examples.
An AI agent can also make promises that the company cannot honour, disclose private information or repeatedly misunderstand an already frustrated customer.
These failures do not prove that customer service will never be automated. They show that reliable automation requires more testing, supervision and human involvement than many businesses expected.
Government support raises uncomfortable questions
AI companies have also asked governments to support investment in energy, semiconductor production and data-center infrastructure.
This has been described by some critics as a request for a bailout.
That description is too simple.
Governments regularly use tax incentives, loan guarantees, procurement contracts and infrastructure programs to support industries they consider strategically important.
The semiconductor industry, renewable energy, aviation and transport have all benefited from similar policies.
The more relevant question is who carries the risk.
If private investors receive the profits when demand is strong, should taxpayers absorb part of the loss if the infrastructure becomes uneconomic?
AI may be strategically important enough to justify public investment. That does not mean every private project should be protected from failure.
Requests for government support are not proof that the AI industry is about to collapse. They do suggest that the planned expansion may be too large to finance comfortably using private capital alone.
The dot-com comparison makes sense
The most useful comparison is probably the internet boom of the 1990s.
Investors were correct that the internet would transform business and society. They were often wrong about how quickly profits would appear, which companies would survive and how much those companies were worth.
The result was a financial bubble built around a genuinely revolutionary technology.
AI could follow a similar path.
The technology may improve productivity, create new products and become part of everyday work. At the same time, many AI companies may be dramatically overvalued. Some providers may disappear. Model prices may fall. Data centers may be sold below construction cost. Investors may discover that the largest profits are captured by companies that were not considered the obvious winners during the boom.
A collapse in valuations would not prove that AI had failed.
It would show that investors paid too much for expected profits that had not yet arrived.
So, is AI a bubble?
Parts of the market almost certainly display bubble-like behaviour.
Valuations are extremely high. Infrastructure spending is growing faster than proven demand. Companies are committing capital based on optimistic forecasts about adoption, productivity and future pricing.
Competition is also becoming more intense. Chinese models, open-weight alternatives and local systems are making AI cheaper, which is good for users but potentially bad for the profit margins of model developers.
At the same time, the argument that the industry has already borrowed several trillion dollars and must replace 10 million workers every year is not supported by the available evidence.
The stronger case is less dramatic.
AI companies and their partners are spending extraordinary amounts of money today. To justify those investments, they will need much greater revenue, much wider adoption and much clearer productivity gains than we can currently observe.
They may achieve that.
They may also discover that AI changes the world while producing disappointing returns for many of the investors who financed it.
Those two outcomes are not contradictory.
The internet transformed almost every industry. That did not prevent investors from losing billions during the dot-com crash.
AI does not need to be a fraud for an AI bubble to burst. It only needs to be priced as if nothing can go wrong.
This article is for informational purposes only and does not constitute investment advice.