The debate about an AI bubble often starts with the wrong question: Is AI overhyped?

AI can become one of the most important technologies of our time and still create an investment bubble. Railways, telecommunications and the internet transformed economies while destroying significant amounts of investor capital along the way.

The real question is:

„Will AI create enough economic value for customers, quickly enough, to justify the extraordinary amount of capital being invested in it?„

It starts with the customer

Companies will continue paying for AI if it saves or makes them more money than it costs.

But productivity is not the same as financial return.

If AI allows an employee to complete a task 30% faster, the company does not automatically save 30% of that employee’s salary. The saved time has to translate into lower costs, additional output, higher revenue or better margins.

Deloitte surveyed 1,854 executives and found that “85 per cent of organizations increased their investment in the past 12 months, and 91 per cent plan to increase it again this year”. Yet satisfactory ROI on a typical AI use case generally required two to four years (“typical payback period of seven to 12 months expected for technology investments”), and only 6% reported payback within one year. [1]

There is positive evidence too. McKinsey’s 2026 survey found that 44% of respondents were scaling AI across their enterprises, up from 38% a year earlier. [2]

AI adoption is real. The unanswered question is whether adoption will translate into sufficient cash returns.

Meanwhile, the investment bill is enormous

The International Energy Agency says capital expenditure by five major technology companies exceeded $400 billion in 2025 and was expected to increase another 75% in 2026 (“Capital expenditure of just five technology companies is now larger than global investment in oil and natural gas production”). [3]

Individual investment plans illustrate the scale:

• Microsoft expects roughly $175 billion of calendar-year 2026 capital expenditure after accounting for changes in lease classification. [4]
• Meta expects $130–145 billion. [5]
• Alphabet raised its 2026 capex forecast to approximately $195–205 billion. [6]
• Amazon spent $96.3 billion in cash capital expenditures during the first six months of 2026, compared with $55.6 billion a year earlier (“Amazon raised its 2026 cash capex outlook from about $200 billion to about $220 billion in July”). [7]

These numbers do not prove that AI is a bubble. But they show how much future economic value will have to be created to justify today’s spending.

AI does have real customers

The other side should not be ignored:

• Google Cloud revenue reached $24.8 billion in Q2 2026, up 82% year over year. [8]
• AWS segment revenue reached $42.2 billion, up 37% year over year. [9]
• OpenAI’s annualized recurring revenue was reportedly approaching $70 billion by September 2026. [10]

So the risk is not that nobody wants AI.

The risk is that AI revenue and profit may not grow fast enough to justify the capital being invested behind them.

How could the bubble burst?

A bubble rarely bursts simply because people suddenly decide that a technology is useless.

More often, expectations embedded in asset prices and investment plans become impossible to reconcile with realized cash flows.

The sequence could look like this:

• Phase 1: AI experimentation remains extremely high.

Companies deploy copilots, agents and generative AI systems throughout their organizations.

• Phase 2: CFO scrutiny increases.

Companies begin reviewing AI projects more critically. Renewals become dependent on demonstrated savings and revenue generation.

• Phase 3: weaker projects are cancelled.

Projects without measurable ROI are cancelled. Enterprise customers consolidate AI vendors and reduce overlapping subscriptions (that is good for customers but potentially painful for suppliers). 

• Phase 4: AI pricing comes under pressure.

Model competition, open-source models and cheaper inference reduce the price customers are willing to pay. 

• Phase 5: infrastructure utilization disappoints relative to assumptions.

Revenue still grows—but not fast enough to justify previously committed infrastructure.

• Phase 6: investors demand capital discipline.

Investors demand lower capex. Companies reduce data-center projects, GPU orders and speculative AI investments.

• Phase 7: credit tightens.

Financing becomes harder. Highly leveraged infrastructure operators and AI startups face refinancing difficulties. Startups struggle to raise money. Some data-center projects are postponed or cancelled. Eventually, major technology companies begin reducing AI investment plans.

• Phase 8: consolidation begins.

Strong companies buy technologies, talent, models and infrastructure from weaker companies at dramatically lower valuations.

That would constitute an AI bubble bursting even if worldwide AI usage continued increasing. The collection of individual problems could become an AI investment-cycle correction.

Shortly: the technology wins while the investment bubble bursts.

Who gets hurt?

Small AI companies would probably be the most vulnerable, particularly those with high inference costs, weak differentiation, negative cash flow and dependence on continuous financing.

Highly leveraged data-center developers would also be exposed if utilization fell below expectations.

Semiconductor companies could suffer if customers slowed purchases to absorb existing capacity.

Large technology companies are different. Microsoft, Alphabet, Amazon and Meta have highly profitable businesses outside AI. They could absorb a period of disappointing AI returns more easily than startups dependent on their next financing round.

An AI crash could make Big Tech even stronger. The largest companies could acquire distressed startups, infrastructure, models and talent at lower prices.

Think about railways

Some time ago, I have heard the railway analogy which may be particularly useful here.

Railways were revolutionary. But during Britain’s Railway Mania of the 1840s, enormous amounts of speculative capital flowed into railway projects. Companies failed, investors suffered losses and the sector eventually consolidated. [11]

The railway tracks, however, remained.

AI infrastructure could follow a similar pattern:

• Imagine a data center built for $5 billion whose owner eventually fails and sells it for $2 billion.
• The infrastructure still works. But its economics for the new owner are completely different.
• That creates an interesting possibility:

An AI bubble bursting could ultimately make AI cheaper and accelerate its adoption.

Investors lose money. Infrastructure changes hands. The cost of providing AI falls. The technology survives the financial excess surrounding it.

What comes after the bubble?

AI itself probably does not disappear: the models, the data centers, the chips will remain. The companies will continue using AI.

What disappears are some valuations, weak companies, excessive leverage and uneconomic projects.

This is what could be called AI investment-cycle correction.

Investment may then move from building AI toward making money with AI.

That could favor robotics, industrial automation, healthcare, engineering, cybersecurity, logistics, biotechnology, specialized enterprise AI and smaller domain-specific models—areas where ROI can be demonstrated more clearly.

The red flags to monitor

In my opinion, several indicators are worth following.

1. Customer ROI and payback periods. Deloitte currently finds that satisfactory AI payback typically takes two to four years, while only 6% of surveyed organizations reported payback within one year. [1] Are organizations shortening the current two-to-four-year AI payback period? If not, increasing expenditure becomes progressively harder to justify.

2. Enterprise scaling. McKinsey’s current 38% to 44% enterprise-wide scaling figure provides a useful baseline. The percentage needs to rise materially. [2]

3. AI revenue versus AI capex. Revenue should progressively catch up with infrastructure expenditure.

4. Free cash flow. Falling free cash flow (example: of Amazon’s trailing-12-month free cash flow moved from +$18.2 billion to -$7.6 billion, largely because of higher purchases of property and equipment, primarily reflecting AI investment [12]) despite rapidly increasing AI-related revenue is a warning that capital intensity may be overwhelming monetization.

5. Debt issuance. The approximately $220 billion of 2026 debt issuance reported for Alphabet, Amazon and Microsoft deserves continued attention. [13]

6. Data-center cancellations and postponements. Announced projects matter less than projects that actually receive financing, power connections and customers.

7. GPU utilization. Falling utilization would be one of the clearest indications of overbuilding.

8. GPU depreciation assumptions. If technological progress shortens useful economic lives faster than expected, infrastructure economics will deteriorate dramatically.

9. AI pricing. Rapidly falling prices are good for customers and adoption but can put pressure on suppliers.

10. Vendor financing and circular financing. When suppliers increasingly finance customers so customers can buy suppliers’ products, reported demand deserves greater scrutiny.

11. Startup down-rounds and failures. A sustained increase would indicate that private capital is repricing AI risk.

12. Capex guidance from hyperscalers. The first synchronized reductions from several major technology companies would be an important turning point.

The counterargument: why the bubble may not burst

Any scientific assessment must test the opposite hypothesis – let’s try to think about it. 

There is substantial evidence supporting the growth case: AI revenue is already growing rapidly: Google Cloud revenue increased 82% year over year to $24.8 billion in Q2 2026. [8] AWS revenue increased 37% to $42.2 billion.[9] OpenAI’s annualized recurring revenue reportedly approached $70 billion by September 2026. [10]

These numbers are inconsistent with the hypothesis that nobody wants AI.

So, let’s speculate about a different scenario:

• Customer ROI improves.
• Agentic AI automates complete processes rather than merely assisting employees.
• Inference becomes cheaper.
• AI usage expands dramatically.
• New applications emerge.
• Revenue catches up with infrastructure expenditure.

Under that scenario, what currently appears to be overinvestment could eventually resemble the early construction of the internet or electricity grid.

Conclusion

So let me repeat the question from the very first paragraph of this article: „Will the economic value created for AI customers grow quickly enough to pay for the infrastructure being built for them?”

If yes, today’s extraordinary investments may eventually look visionary.

If not, valuations and investment will have to adjust.

And if that happens, AI itself will probably survive.

The weaker balance sheets may not.

Sources

[1] Deloitte. AI ROI: The paradox of rising investment and elusive returns, 2025. Survey of 1,854 executives covering AI investment, ROI and payback periods.

[https://www.deloitte.com/dk/en/issues/generative-ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html]

AWS Executive Insights. TechTarget. Evaluating the ROI of Major Tech Investments, October 2024. 

[https://d1.awsstatic.com/onedam/marketing-channels/website/aws/en_US/whitepapers/approved/executive-insights/esg-research-insights-paper-aws-tech-roi-report-pdf.pdf]

[2] McKinsey & Company. The State of AI: Global Survey 2026. Enterprise AI adoption and scaling.

[https://www.mckinsey.com/~/media/mckinsey/businessfunctions/quantumblack/ourinsights/thestateofai/the-state-of-ai-in-2026-on-the-road-to-roi-aug-2026.pdf?shouldIndex=false]

[3] International Energy Agency (IEA). Key Questions on Energy and AI, 2026. Data-center and major technology-company investment and energy requirements.

[https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf]

[4] Microsoft. FY2026 Q4 Earnings Conference Call and investor materials. Capital expenditure and AI infrastructure investment.

[https://es.tradingview.com/news/stocktwits:8b6860adc094b:0-msft-stock-soars-after-results-cfo-dismisses-ai-overcapacity-narrative-says-can-rein-in-spending-if-demand-changes/]

[5] Meta Platforms. Q2 2026 Results. 2026 capital-expenditure outlook.

[https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx]

[6] Reuters. July 22, 2026. Reporting on Alphabet’s Q2 2026 results and revised capital-expenditure guidance.

[https://www.reuters.com/business/google-quarterly-cloud-revenue-growth-beats-expectations-2026-07-22/]

[7] Amazon. Q2 2026 Form 10-Q. Cash capital expenditure and infrastructure investment.

[https://www.moomoo.com/news/notice/307810751/amazon-10-q-q2-2026-earnings-report?level=1&data_ticket=1791278846587719]

Yahoo. Gian Estrada. Amazon Raised Its 2026 Capex Guide to $220 Billion.

[https://finance.yahoo.com/markets/stocks/articles/amazon-raised-2026-capex-guide-114017457.html]

[8] Alphabet. Q2 2026 Results and SEC filings. Google Cloud revenue and year-over-year growth.

[https://s206.q4cdn.com/479360582/files/doc_financials/2026/q2/2026q2-alphabet-earnings-release.pdf]

[9] Amazon. Q2 2026 Results. AWS revenue and year-over-year growth.

[https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/]

[10] Reuters. September 29, 2026. Reporting on OpenAI’s annualized recurring revenue.

[https://www.reuters.com/technology/openais-annual-recurring-revenue-nears-70-billion-axios-reports-2026-09-29/]

[11] Journal of Applied Corporate Finance. Research on Britain’s Railway Mania, railway investment, financial losses and subsequent industry consolidation.

[https://onlinelibrary.wiley.com/doi/full/10.1111/jacf.12673]

[https://www.cambridge.org/core/journals/financial-history-review/article/abs/deriving-the-railway-mania/2EA377886FE48B4BF2DD620D4F845991]

[12] Amazon. Q2 2026 Results and cash-flow disclosures. Trailing-12-month free cash flow and increased purchases of property and equipment.

[https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/]

[13] Reuters. September 10, 2026. AI debt splurge is warping credit spreads.

[https://www.reuters.com/commentary/reuters-open-interest/ai-debt-splurge-is-warping-credit-spreads-marty-fridson-2026-09-10/]

Keywords

AI Bubble, AI Bubble Burst, Artificial Intelligence, AI Investment, AI ROI, AI CapEx, AI Infrastructure, Generative AI, Enterprise AI, Data Centers, AI Economics, AI Productivity, AI Monetization, GPU Economics, Technology Bubble, Investment Risk, AI Consolidation, Railway Mania, Future of AI.