Is AI a Bubble? The Money, the Debt and the Data

Is AI a bubble? It is the market question of 2026. Kunkafa answers market questions in one way: not with a yes or a no, but with how far a market could move, within what duration, in which direction, and how confident to be about each path.

So we did what Kunkafa does with prices. We stopped listening to the noise and read the numbers: company filings, quarterly results, and the record of spending, debt and revenue.

Watch: AI bubble fears

Play the film·4:29

1. The circle: who pays whom?

AI needs chips, cloud and power. Investors fund the buildout. In the end, customers have to bear the cost.

The twist: the same companies are often investor, supplier and customer at once.

Investments, contracts and warrants are different promises.

The circle even reaches profits. Microsoft’s net income for the year to June rose by $5bn from gains on its OpenAI investment. The year before, the same stake cost it $3.6bn.

NVIDIA, the main seller of AI chips, is now also a major investor in its buyers. Its latest quarterly report shows $47.9bn of stakes in private companies, up from $3.8bn a year earlier, plus $42.8bn in companies listed on stock exchanges. In August it also agreed to guarantee up to $105bn of data-centre leases for an OpenAI affiliate.

2. The backlog: promised, not yet earned

Bulls point to backlog: revenue customers have signed for but not yet paid. It is enormous.

Two things stand out. Much of it arrives years from now: only about 13% of Oracle’s is due in the next twelve months. And much of it comes from the circle above. Amazon’s own filing names the OpenAI and Anthropic expansions among its commitments.

A backlog is only as strong as the customer behind it. If a few AI labs are the customer, their fundraising is everyone’s revenue.

Steve Eisman, the investor who bet against the subprime mortgages behind the 2008 global financial crisis, and one of the heroes of The Big Short, puts a number on it. He estimates OpenAI and Anthropic account for roughly 70% of AI-related revenue at Microsoft, Amazon, Google and Oracle.

“The futures of these massive companies, in a sense, are a bet that OpenAI, Anthropic are going to succeed.”

3. The bill: bigger than the telecom boom

Here is what five companies spent on buildings and equipment in their latest quarter.

For the full year, Alphabet guides to $195–205bn, Microsoft to about $175bn and Meta to $130–145bn.

Scale it to the economy. Apollo’s chief economist puts hyperscaler spending at 0.3% of US GDP in 2019 and 1.4% in 2025, with forecasters expecting about 3% a year from 2027. The late-1990s telecom and fibre boom peaked at 1.2%, then collapsed in what became known as the dot-com bubble.

The people signing the cheques know the risk. Sam Altman has said that in bubbles “smart people get overexcited about a kernel of truth.” Mark Zuckerberg said he would rather Meta “misspend a couple hundred billion dollars” than be late. Both, as reported by Fortune, call the spending a bet worth making.

4. Who pays: cash, bonds, leases and promises

Operating cash helps pay for it. So do bond markets, landlords and new investors.

Not every bond is for AI, and an undrawn credit line is not a loan. The pressure shows lower down. CoreWeave, which rents out AI chips, more than doubled revenue to $2.6bn last quarter. It still lost $626m, and the interest on its loans alone was $640m.

5. Free cash, then and now

Big tech’s defence has always been that it pays from its own pocket, out of the spare cash it generates. The filings show that pocket getting thinner.

Cash from operations has roughly doubled at most of these companies. Spending has grown faster. Amazon and Oracle now spend more on buildings and equipment than their businesses bring in, and Alphabet and Meta keep less than they did in 2023. The cash is flowing to the seller: NVIDIA’s free cash flow is almost five times what it was.

*Alphabet excludes $80bn of SpaceX shares it holds as short-term investments. Several balances grew because of new bonds and share sales, not spare cash.

6. The demand is real

The bear case often skips this part. NVIDIA sold $96.2bn in one quarter, up 106%, and expects about $108bn next. AWS grew 37%. Microsoft’s Azure grew 41% over the year. Google Cloud added $11.1bn of quarterly revenue in a year.

The AI labs are growing too. Anthropic told investors its revenue was running at $65bn a year by July, about seven times a year earlier. That is a lot of customers paying real money.

That is the case the optimists make. “I don’t believe we’re in an AI bubble,” NVIDIA’s Jensen Huang told Bloomberg. Jeff Bezos calls it “a kind of industrial bubble”: money gets wasted, but what gets built stays useful.

Even the builders hedge. Google’s Sundar Pichai sees “elements of irrationality,” and told the BBC that if it bursts, “no company is going to be immune, including us.”

7. Apple: the company that chose not to build

One giant stayed out. Apple spent about $10bn on property and equipment in the last twelve months. Alphabet spent $132bn and Amazon $173bn.

Instead of building its own frontier model, Apple rents one. In January it signed a multi-year deal to base its models on Google’s Gemini, reportedly for about $1bn a year. Apple called Google’s technology “the most capable foundation for Apple Foundation Models.”

The result shows in the cash. Apple produced the most free cash flow of the group, $136.7bn, and spent about $82bn buying back its own shares over the same twelve months.

It is a bet both ways. If AI models become cheap and plentiful, Apple rents the best one for a fraction of the build cost. If a few labs pull far ahead, Apple depends on a rival for the brain of its products.

8. When intelligence gets cheap

The deepest bubble question is not about spending. It is about price. What if the thing being built becomes a commodity?

The first warning came in January 2025. China’s DeepSeek released a cheap, open model that matched top rivals on many tests. NVIDIA fell nearly 17% in a day, a record loss of close to $600bn in market value.

The gap has kept closing. Epoch AI found in May that the best open-weight models trail the closed frontier by about four months. Its earlier work found that the price of reaching a given level of performance falls between 9 and 900 times a year, depending on the task.

“The Chinese open-end models, open-weight models are much cheaper. And if they start really taking a lot of market share … you could have a big price war.”

The other side: cheaper answers can mean far more questions. Falling prices have not stopped the frontier labs’ revenue from multiplying, and businesses still pay a premium for the best model. But a lead that lasts four months is a hard thing to charge for over a ten-year lease.

9. Three gaps that could still bite

• Revenue. Bain estimated AI needs about $2 trillion a year in revenue by 2030 to pay for its compute, and could fall about $800bn short.
• Power. The IEA expects data centres to use about 950 TWh of electricity in 2030, double 2025. Grids, turbines and transformers are already bottlenecks.
• Accounting. Big buyers spread the cost of AI chips over five or six years. Michael Burry, the other famous name from The Big Short, says the chips are replaced every two to three years. If he is right, costs are understated today and profits flattered.

“Understating depreciation by extending useful life of assets artificially boosts earnings.”

Burry estimates the gap at about $176bn of missing costs from 2026 to 2028, with Oracle’s profits overstated by about 27% and Meta’s by about 21% by 2028. It is his estimate, and hard to prove: accounting rules give companies leeway. The defence is that older chips keep earning, moving from training to lighter work.

Bubble fears and boom facts, side by side

What could go wrong, and what could go right

If growth disappoints, the strain could travel through the same links. Orders could slow, funding could tighten, projects could stop. Lower earnings expectations could pull down share prices, and indices heavy in technology with them.

Who would catch the fall? Not Washington, said David Sacks, then the White House adviser on AI: “There will be no federal bailout for AI.” If one lab fails, he argued, others take its place. Weeks later he warned that a reversal in AI investment “would risk recession.” NPR reported both posts.

JPMorgan’s Jamie Dimon put the history plainly: “AI in total will pay off. Just like cars in total paid off.” But most people involved “didn’t do well.”

If paying customers grow into the capacity, the same links become a supply chain built just in time. A useful technology can still be an expensive investment. An expensive investment can still pay off. That is why we look both ways.

Where the experts stand

The experts disagree, and each has a stake in the answer. Kunkafa weighs seventy views of the same market before settling on one, then counts how it turned out.

What Kunkafa’s record shows on the Nasdaq 100

No one knows for sure whether this is a bubble, and Kunkafa does not claim to. Kunkafa puts a confidence on possible moves in both directions, for every duration, and counts how each one turned out.

From 11 September to 2 October, 24,875 Nasdaq 100 forecasts at Kunkafa’s confidence of 95% and above settled: 9,628 on rises and 15,247 on falls, with durations of up to 12 days. Each is measured by how far price travelled toward its level within its own duration.

100% means price reached the level. It is a distance, not an investment return. The forecasts that fell short are in the same chart. Every band, by direction and duration, is in the full Nasdaq 100 report, and how to read forecast results explains each band.

Six questions to ask about the AI bubble

• Who is the customer? Is demand coming from outside the circle of investors and suppliers?
• Is it cash, a contract or a warrant? Each is a different promise with a different risk.
• When is the backlog due? A contract paid in 2030 is not revenue in 2026.
• What is a four-month lead worth? Open models catch up fast and cost less.
• How long does the hardware last? The answer moves profits by billions.
• What does the other direction look like? A fall is only half the picture. Both outcomes stay possible.

Bubbles are argued in headlines. Moves can be measured. Create a Kunkafa account and read the Nasdaq 100, both directions, beside its record.

Figures from company filings and releases, Apollo, the IEA, Bain and CNBC, checked through 10 October 2026.