AI Bubble Risks Aren’t Equal: What Ordinary Investors Should Examine

In early July 2026, Economy Media asked how AI became more expensive than some of the workers it was supposed to replace. Later that month, Asian Dad Energy declared that “AI Is About to Crash”.

Two YouTube videos are not a financial indicator. But the arguments inside them raise questions investors should take seriously: How much debt is funding the AI buildout? What revenue would justify the spending? And what happens if the expected profits arrive later than the bills?

As a finance student interested in venture capital and AI companies, one calculation particularly caught my attention. Asian Dad Energy argued that the industry would need to profitably replace approximately 10 million American white-collar workers every year to support the investment behind it.

The broader concern is legitimate. But the calculation depends on assumptions that become less certain when examined individually. The strongest version of the AI-bubble argument is not that the entire boom is imaginary. It is that enormous amounts of money may be chasing the same optimistic expectations—and that some parts of the market are far less prepared for disappointment than others.

What the dot-com bubble actually teaches us#

The dot-com comparison is useful, but not because AI and the early internet are identical.

An NBER study of 356 internet-related initial public offerings found that only 36.8% remained publicly traded by March 2005. Another 38.2% had merged, while 25% fell into the study’s narrower delisted category. That does not mean nearly two-thirds went bankrupt, but it shows how few remained independent listed businesses.

The losses also reached eventual winners. In June 2001, the Federal Reserve reported that Nasdaq had fallen nearly 60% and internet stocks about 70% from their peaks. Amazon survived, but its 2001 annual filing reported a quarterly share-price high of $91.50 in early 2000 and a quarterly low of $5.97 in 2001.

The point is not that AI must replay 2000. It is that technological success and investor success were never the same thing. Investors could correctly predict the internet’s importance and still choose the wrong company or pay the wrong price.

Where did ten million workers come from?#

Asian Dad Energy’s calculation begins around 2:28. It starts with an estimated $2 trillion to $3 trillion of AI-related corporate debt. At interest rates of 3% to 4%, that would create between $60 billion and $120 billion in annual interest. The video rounds this to approximately $100 billion.

Assuming a hypothetical 10% profit margin, generating $100 billion in profit would require approximately $1 trillion in revenue. The creator argues that replacing white-collar labor is the only market large enough to generate revenue on that scale. At an assumed $100,000 per worker, $1 trillion corresponds to approximately 10 million workers.

The video does not specify whether that 10% margin is measured before or after interest. This shows how much revenue the scenario might require, rather than providing a complete account of the industry’s debt payments.

Because AI is not replacing anything close to 10 million white-collar workers annually, the creator concludes that the revenue and profit needed to support the claimed debt are not arriving.

Upon closer inspection, however, the hardest part of the calculation to prove is the starting debt figure that the creator claims. I could not find a reliable primary source confirming that the US AI industry already owes $2 trillion to $3 trillion.

Some estimates may be combining different kinds of money: spending that has already happened, projects announced for the future, money raised from investors, and money borrowed through loans or bonds. Those numbers may all describe the AI boom, but they are not all debt that needs to be repaid. “Investment” can also mean completed capital expenditure, announced infrastructure plans, equity financing, corporate borrowing, leases, or projected future spending. Those categories cannot be treated as one debt balance.

The calculation also assumes that replacing workers is the only or main way AI companies will make enough money. That is one possible source of revenue, but not the only one. AI could instead create revenue through cloud services, software subscriptions, advertising, new products, or productivity improvements inside companies that never eliminate the corresponding jobs.

The broader borrowing concern is still supported by evidence. The IMF reported in January 2026 that debt was becoming a more common way to fund technology and AI investment, increasing the possible damage if expected returns fail to appear.

The Bank of England found that Meta, Alphabet, Amazon, Microsoft, and Oracle represented 3% of outstanding US investment-grade debt at the end of 2025. Investment-grade debt refers to corporate borrowing considered relatively less likely to default.

By early May 2026, those companies accounted for more than 15% of year-to-date US investment-grade issuance—the new investment-grade bonds issued so far that year.

That does not cover every AI company or every type of loan. It shows that the five companies’ existing debt remained a relatively small part of that market, but their new borrowing was increasing quickly.

So I agree with the warning behind the calculation: if companies borrow heavily because they expect enormous AI profits, but those profits do not arrive, some may struggle to make their payments.

What the evidence does not show is that the entire industry already shares one $2 trillion to $3 trillion debt problem—or that replacing 10 million workers every year is the only way to pay for the boom.

AI is not one balance sheet#

The phrase “AI bubble” makes the market sound more uniform than it is. Large technology companies, data-center developers, model providers, application companies, and businesses adopting AI do not share the same economics.

Big technology companies: real demand, extraordinary spending#

Alphabet spent $91.4 billion on long-term assets in 2025. By July 2026, it expected that spending to reach $195 billion to $205 billion, even as second-quarter Google Cloud revenue rose 82% to $24.8 billion.

Microsoft similarly reported $41 billion in quarterly capital spending, with roughly two-thirds going toward shorter-lived assets such as CPUs and GPUs. Azure and other cloud-services revenue grew 43%, while demand continued to exceed capacity.

These are profitable companies responding to real cloud demand, some of it AI-related. Their central investor risk is whether this extraordinary spending will eventually earn enough—and whether their stock prices already assume that it will.

Infrastructure: power and financing#

Economy Media says around 4:52 that nearly half of US data-center projects planned for 2026 were expected to be delayed or canceled.

I could not verify the underlying project list or research method through an accessible primary source, so I would not treat that as an established cancellation rate.

Separate research from the International Energy Agency supports the broader concern. It estimates that approximately 20% of planned data-center projects globally could face delays unless grid problems are addressed. It also notes that new transmission lines can take four to eight years to complete in advanced economies.

A delay can increase costs and postpone customer revenue without disproving the demand for computing. This makes electricity access, construction timing, borrowing costs, and customer contracts especially important for infrastructure projects.

American models and applications: value can move#

OpenRouter’s analysis found that Chinese-authored models surpassed US-authored models’ share of token volume on OpenRouter in early June 2026.

This measures traffic routed through one platform, not all AI use by American companies. Still, competition from capable alternatives could weaken model providers’ ability to maintain high prices while making AI cheaper for application companies and other businesses.

Enterprise adoption shows the same unevenness. A Sinch-commissioned survey found that 74% of surveyed enterprises had rolled back or shut down at least one deployed AI customer-communications agent after a governance failure. Yet 98% said they were increasing investment in AI communications during 2026.

Failures and continued investment can coexist. The same development can destroy value in one part of AI and create it somewhere else.

What ordinary investors can examine#

Most individual investors cannot audit a private model provider’s computing contracts or predict when market sentiment will change. They can still ask better questions:

• Which part of the AI market am I actually investing in?
• Is evidence about one company or platform being generalized to the entire market?
• Is demand recurring use, a pilot, a backlog, or management guidance?
• What remains after computing, support, and other operating costs?
• How much must the company keep spending before revenue becomes cash it can retain?
• Is growth funded by the company’s own cash, new investors, or borrowing?
• What future growth does the current stock price already assume?
• What evidence would make me revise my view?

Index investors should add one more: how much common AI risk is already inside a supposedly broad portfolio?

Investor.gov explains that diversification can reduce the effect of one investment performing badly, but it cannot guarantee protection when the broader market falls.

That is not an argument for abandoning index funds or trying to time a correction. It is an argument for understanding what diversification does—and does not—provide.

My provisional conclusion#

Some of these crash-content creators may be right about the broad problem: The AI buildout is becoming more expensive and more dependent on future profits arriving at the right time. Borrowing is increasing, data centers face real power constraints, and companies are spending extraordinary amounts of money.

But those pressures will not affect every part of AI equally. A large profitable company can survive the investment cycle and still disappoint shareholders. A heavily indebted infrastructure project can fail even while cloud demand grows. A model provider can lose pricing power while application companies benefit from cheaper inputs.

That is why the dot-com comparison remains useful. The internet transformed the economy, but value was distributed unevenly. Correctly identifying the technological shift was only the beginning of the investment problem.

The most credible AI-bubble thesis is not that every part of the market will collapse. It is that too much money may be chasing the same optimistic expectations—and that some parts of the market are far less prepared for disappointment than others.

AI may be a technological boom. Financially, it is not one bubble.