The AI bubble in late 2026: how big, and what to watch

Everyone agrees AI is real. Nobody agrees on the price tag attached to it.

That is the actual debate right now. The technology works, revenue is showing up, and yet warnings about an AI bubble keep getting louder, not quieter, the more money pours in.

What the money is buying

Start with what is actually being built. Amazon, Microsoft, Alphabet, and Meta are guiding toward roughly $725 billion in combined capital expenditure for 2026, up about 77% from around $410 billion in 2025.

That number is already climbing further out. The Bank of England says estimates for AI hyperscaler capital spending in 2028 have jumped from under $600 billion at the end of 2025 to more than $1 trillion. That is not a forecast drifting slightly higher. It is a projection that has nearly doubled in less than a year.

And the revenue backing it up is real. Nvidia reported $96.2 billion in quarterly revenue for the quarter ended July 26, up 106% year over year, while its data centre business alone generated $89 billion, up 117%. Whatever else is true about this market, money is actively moving through the AI supply chain in enormous amounts. The company also guided to another year of steep growth.

The dotcom comparison

History offers a blunt warning here. The internet was genuinely revolutionary, and the dotcom bubble was also genuinely real. The fact that the internet eventually reshaped the global economy did nothing to stop investors from paying absurd prices for companies with no real path to the growth those prices assumed.

AI could follow the same pattern. The technology can keep transforming industries while parts of the market attached to it become dangerously optimistic. The European Central Bank raised almost exactly this question in August, comparing the current AI rally with the late 1990s technology boom and asking whether today's valuations reflect rational expectations about AI's economic potential or another speculative cycle.

Concentration widens the hit

Concentration makes this riskier than usual. The Bank of England says AI related companies now account for around half of the S&P 500, compared with roughly a quarter in 2022. A correction concentrated in a small group of stocks would therefore not stay contained to investors who deliberately bought into AI. It could move through portfolios that simply have exposure to the broader US market.

From cash to debt

For a while, the biggest tech companies funded their AI buildouts largely from their own cash. That picture is changing. The Bank of England says AI companies are increasingly turning to public debt, private credit, leveraged finance, and structured financing as their infrastructure needs expand. The OECD estimates that the share of private credit financing going toward AI investment rose from 9% in 2024 to 34% in 2025, while more than half of the external financing needed for global data centre investment between 2026 and 2028 could come from debt.

That does not guarantee a crash. It does mean the consequences of disappointment are getting bigger. An investment boom funded by cash reserves can absorb a slow quarter. One increasingly dependent on borrowed money has far less room to absorb a prolonged shortfall in returns.

When safe assets start paying

The AI trade has thrived on investors being willing to pay today for profits expected years from now. That gets harder when safer assets start paying well too. The US 10 year Treasury yield is approaching 5%, while the S&P 500's forward price to earnings ratio has fallen from 22.2 earlier in the year to 19.7, still above its long term average. Higher yields can put additional pressure on valuations, particularly for growth companies whose investment case depends heavily on future cash flows.

This does not make AI companies unattractive on its own. It just raises the bar. If earnings keep beating expectations, high valuations can survive. If growth slows while yields stay elevated, investors may stop being willing to pay today's prices for tomorrow's story. Cheaper competition can push in the same direction, as we wrote when Chinese labs began giving models away.

The circular part

One of the stranger features of this market is how tightly connected everything has become. Hyperscalers spend on chips and data centres. Chipmakers supply the infrastructure. Startups rent the compute. Investors fund the startups. Some of those same technology companies invest directly in the startups buying their own compute.

The Bank of England describes some of these arrangements as self reinforcing capital loops and warns they could amplify a shock if expectations suddenly shift. None of this means the revenue is fake. It means a company's growth may depend more heavily on another AI company's continued financing than the headline number suggests, and that is worth knowing before treating every AI revenue line the same way.

“AI-related companies' use of credit markets has accelerated rapidly, including in public markets, private credit, leveraged and structured finance, and is set to increase further as financing needs continue to expand. This pace of investment is unprecedented historically.”

— Bank of England, Financial Stability Report, July 2026

Quote source: Bank of England, Financial Stability Report, July 2026

The numbers that settle it

The next chapter of this market probably will not be settled by another flashy model demo. It will come down to a handful of unglamorous numbers:

• revenue against spending, since rising capital expenditure keeps demanding proof that it is producing comparable returns
• free cash flow, which the Bank of England says is declining among major AI hyperscalers
• data centre utilisation, since building capacity is not the same as using it profitably
• margins, because AI is creating enormous demand while competition and infrastructure costs continue to pressure returns

The winners will not necessarily be the companies with the most users. They will be the ones that can turn those users into durable profits.

A slowdown is enough

None of this needs a full collapse to hurt investors. A slower period could be enough. If hyperscalers pull back on capital spending, startups struggle to raise their next round, or corporate customers discover that some AI projects deliver less value than promised, the market could reprice expectations without AI itself going anywhere. That is closer to what happened after the dotcom crash than most people remember. The technology won. Not every company that rode the hype did.

Nobody can tell you exactly when or if this breaks. The honest answer is that both outcomes are still on the table. AI could keep outrunning expectations and make today's spending look reasonable in hindsight, or it could simply meet expectations and leave today's valuations struggling anyway.

Either way, the question worth tracking is not whether AI matters. Almost everyone already agrees it does. It is whether the value it creates arrives fast enough to justify what has already been priced in.

This article is for informational purposes only and does not constitute investment advice.