Is AI a Bubble? The Debt, the Data Centres and the Warning Signs in 2026

For three years, artificial intelligence has powered the biggest investment boom in modern corporate history. Tech giants are pouring hundreds of billions of dollars into data centres, chipmakers are posting record profits, and AI companies are heading for some of the largest stock-market listings ever. Then, in the week to 16 October 2026, the mood cracked. The Nasdaq had its worst week since June, chip stocks suffered their longest losing streak in two decades, South Korea's Kospi had its worst week since the 2008 financial crisis, and the price of insuring one AI cloud company's debt against default hit a record.

So is AI a bubble? The honest answer is more interesting than a simple yes or no. This in-depth guide looks at the evidence on both sides: how much is being spent, how it is being financed, what the markets and regulators are signalling, whether the demand is real, and what to watch next. It draws on company results, central bank and IMF reports, and reporting by Reuters, Bloomberg, the Financial Times and others, all checked as of 20 October 2026.

Contents

1. The short answer
2. How much is being spent on AI
3. How the boom is being financed: the debt question
4. The weak links: neoclouds and the OpenAI web
5. What markets are signalling
6. The warnings from regulators and famous investors
7. The bull case: why this may not be 2000
8. Is the demand real?
9. Seven warning signs to watch
10. What it means for ordinary investors
11. Frequently asked questions

1. The short answer

There are strong signs of excess in some corners of the AI economy, particularly where borrowed money is funding data centres for customers who are themselves losing billions. But the core of the boom is being financed by some of the most profitable companies in history, demand for chips still exceeds supply, and AI revenues are growing at extraordinary speed. The most likely risk is not that AI turns out to be useless, but that too much is built too quickly with too much debt, and that returns arrive more slowly than investors have priced in. That is how many past investment manias ended: the technology survived, while many investors did not.

2. How much is being spent on AI

The scale of spending on AI infrastructure, chips, servers, data centres and power, has no modern precedent. Based on guidance given in their July 2026 earnings calls, the four biggest US cloud and platform companies plan to spend roughly $720–745 billion on capital expenditure in 2026, about 1.8 times what they spent in 2025.

*Microsoft's lower figure reflects an accounting change in how future data-centre leases are classified, not a spending cut; its finance chief said its 2026 investment expectations were unchanged. 2025 figures are approximate and the companies define capex differently.

Add Oracle, which guided for $90–95 billion of gross capex in its fiscal year to May 2027, and the top five spenders are on course for roughly $800 billion in a single year. Goldman Sachs estimates that Meta, Microsoft, Amazon and Alphabet will spend about $5.3 trillion on capex over 2025–2030. The Bank of England noted in July 2026 that consensus forecasts for hyperscaler capex in 2028 had risen above $1 trillion, from under $600 billion just seven months earlier.

Spending is outrunning cash flow

For years, Big Tech funded its investments comfortably from operating cash. That is changing:

• Amazon spent about $173 billion on capex in the twelve months to June 2026, against operating cash flow of $161.4 billion, leaving negative free cash flow of about $7.6 billion, compared with a positive $18.2 billion a year earlier.
• Oracle reported negative free cash flow of about $23.7 billion for its 2026 fiscal year.
• FactSet analysis in July 2026 expected free cash flow to be near zero or negative this year for the big spenders other than Alphabet and Microsoft, and found that new debt had risen from 9% of their capex in fiscal 2024 to about 32% in the year to June 2026.

The power problem

AI also needs enormous amounts of electricity. The International Energy Agency estimates that data centres used about 485 terawatt-hours of electricity in 2025, and that this could roughly double to about 950 TWh by 2030, around 3% of global demand. In PJM, the largest US grid region, capacity auction prices have hit their regulatory cap three times in a row, driven almost entirely by data-centre demand. Power is now a real constraint on how fast the boom can grow, and a cost that customers and households help to pay.

3. How the boom is being financed: the debt question

This is where many economists' concern is concentrated. Equity-funded booms can deflate with limited wider damage; debt-funded ones can spill into the financial system.

Some of the deals are vast. Meta sold a $30 billion bond in October 2025 and, the same month, financed its Hyperion data centre through a separate entity that issued $27.3 billion of bonds maturing in 2049, keeping the debt off Meta's own balance sheet. Alphabet raised $25 billion in August 2026 with an order book of about $115 billion. Oracle's total borrowings rose from about $93 billion to nearly $130 billion during its 2026 fiscal year.

Morgan Stanley has estimated that private credit, lending by funds rather than banks, could provide more than half of the roughly $1.5 trillion of data-centre financing needed through 2028. Data-centre asset-backed securities, bonds secured on the buildings and their leases, are also growing quickly.

Why the debt worries regulators

The Bank of England's July 2026 Financial Stability Report flagged several concerns: AI-related issuers accounted for 41% of non-refinancing US high-yield bond issuance in 2026; there is a mismatch between AI chips, which may become obsolete within a few years, and data-centre debt that runs for more than a decade; and "circular financing" links between AI companies and their suppliers. If AI revenues disappoint, lenders could be left holding long-dated debt backed by assets that have lost much of their value.

4. The weak links: neoclouds and the OpenAI web

CoreWeave: the canary in the coal mine

If the AI boom has a pressure point, markets think they have found it in the "neoclouds", specialist companies that buy Nvidia chips, often with borrowed money, and rent computing power to AI developers. The best known is CoreWeave.

A credit default swap price near 1,200 basis points implies, according to ICE Data Services, roughly a 60%+ chance of default within five years. CoreWeave's 2031 bonds fell to about 80 cents on the dollar. That is a stark signal from credit markets, even as Nvidia's chief executive Jensen Huang said on 15 October that CoreWeave's cost of capital was falling and Nvidia would not need to support it. On 19 October, CNBC summed up the mood: stock investors were ready to move on from the rout, but credit investors were not.

OpenAI: huge promises, huge losses

At the centre of the web of AI deals sits OpenAI. It has signed agreements reported to include about $300 billion of computing with Oracle over five years, $250 billion of Microsoft Azure services, $38 billion with Amazon Web Services, custom chips with Broadcom and AMD, and the Stargate data-centre project announced at $500 billion in January 2025. In February 2026 OpenAI said it was targeting about $600 billion of compute spending by 2030.

Its revenue is growing fast, but not as fast as some believed. The Financial Times reported in early October that OpenAI's annualised revenue was nearing $50 billion, lower than a widely circulated figure of about $70 billion that had counted some partner sales gross. Losses remain enormous: the FT reported that a company presentation projected cumulative negative free cash flow of about $278 billion over 2026–2030. OpenAI raised $122 billion at an $852 billion valuation in March 2026.

Circular financing

Critics worry about money going round in circles. Nvidia, for example, took a stake of about $30 billion in OpenAI's funding round, while OpenAI buys Nvidia-powered computing; AMD granted OpenAI warrants for its shares alongside a chip deal. According to reports on the IMF's October 2026 Global Financial Stability Report, which analysed 14 major AI deals, large tech firms had put about $300 billion of equity and $650 billion of debt into AI companies that, in turn, had committed around $1.4 trillion of future spending back to them. If one link weakens, the effects could ripple through suppliers, lenders and investors at once.

5. What markets are signalling

The October wobble

Markets bounced on 19 October, with the Nasdaq up 1.1%, a reminder that one bad week is not a crash. But the sell-off showed how sensitive AI stocks have become to any doubt about financing.

Valuations: two very different pictures

• Expensive on long-term measures: the cyclically adjusted price-to-earnings (CAPE) ratio for the S&P 500 was about 40.7 in September 2026, the second-highest reading in about 150 years and close to the dot-com peak of about 44. Its long-run average is around 17.
• Reasonable on forward earnings: FactSet put the S&P 500's forward 12-month P/E at 19.0 in mid-October, actually below its 5-year average of 19.8, because profits are growing fast (Q3 2026 earnings growth was tracking at about 29%). Nvidia traded at only around 13–14 times forecast earnings.
• Concentrated: the seven largest US tech companies make up roughly a third of the S&P 500's value, well above the roughly 26% peak share of the top stocks in 2000.

The gap between these measures is the heart of the debate. If forecast AI profits arrive, today's prices may look sensible. If they do not, the forward P/E will rise sharply as earnings estimates fall.

Rates are no longer helping

Bubbles tend to inflate when money is cheap. In 2026 it is not. The US Federal Reserve raised rates in September for the first time since 2023, and the 10-year Treasury yield hit about 5.35% in early October, its highest since 2002 (see our report on Treasury yields). Higher yields raise the cost of the debt funding AI and reduce what investors will pay for distant future earnings. With a forward earnings yield of about 5.3% and a 10-year yield above 5%, the extra reward for owning shares rather than bonds is unusually thin.

6. The warnings from regulators and famous investors

• IMF (13 October 2026): its Global Financial Stability Report warned of "a correction in AI-related asset prices", amplified by "rising leverage and concentrated exposures". The IMF's World Economic Outlook estimated that an AI correction scenario could cut global GDP by 1.1% and US output by 1.8% by 2028. We explored what this means for portfolios in our report on the IMF's warning.
• Bank of England: its Financial Policy Committee said in July 2026 that valuations had "become more stretched", driven by a "narrow set of AI-related companies", and that AI companies' use of credit markets "has accelerated rapidly".
• Bank for International Settlements (June 2026): warned that disappointment in returns "could trigger a sudden pullback in financing" and a "hangover" once AI spending peaks.
• Jeremy Grantham (GMO, January 2026): argued the US market met his firm's definition of a bubble and compared AI to the railways: a transformative technology that still produced a crash.
• Michael Burry: the investor famous for betting against US housing before 2008 has disclosed bets against several AI-linked stocks.
• Sam Altman (August 2025): even OpenAI's chief executive said investors as a whole were "overexcited about AI" and compared the moment with the dot-com boom.
• Barclays (16 October 2026): strategist Emmanuel Cau said "the best of the AI thematic trade may be behind us", although he saw the sell-off as mainly about positioning rather than a collapse in fundamentals.

7. The bull case: why this may not be 2000

Comparisons with the dot-com era are everywhere. Between March 2000 and October 2002 the Nasdaq fell about 78% and did not regain its peak until 2015. Telecom companies had borrowed hundreds of billions of dollars to lay fibre-optic networks that sat largely unused, and many went bankrupt. (Read the full story in our dot-com bubble explainer.)

But there are important differences:

Bulls such as Wedbush's Dan Ives describe the spending as an arms race in which the big players cannot afford to cut back, and expect a phase in which AI investments are increasingly turned into revenue. JPMorgan argued in July 2026 that investors should buy dips in AI chip stocks because demand is strong and little new chip capacity arrives before 2028.

8. Is the demand real?

The supply chain says yes

• Nvidia reported revenue of $96.2 billion for the quarter to July 2026, up 106%, with data-centre revenue of $89 billion, and guided to about $108 billion for the next quarter. Jensen Huang said AI "has reached its inflection point".
• TSMC posted a record profit in the third quarter of 2026 and said AI demand still exceeds its supply capability (see our TSMC report).
• ASML raised its 2026 outlook and said some customers are already discussing equipment orders for 2029 and 2030.
• High-bandwidth memory used in AI chips has been reported sold out through 2026.

Cloud and AI revenues are surging

Anthropic's results, and investors' reaction to slower quarter-on-quarter growth ahead of its planned listing, are covered in our Anthropic IPO report.

But business adoption is slower

• The US Census Bureau found that only about 20% of US firms used AI in their business as of May 2026, rising to 37% among firms with 250+ employees.
• McKinsey's 2025 survey found 88% of organisations use AI somewhere, but nearly two-thirds had not scaled it across the business, and only 39% reported any impact on operating profit, mostly small.
• An MIT study published in 2025 estimated that about 95% of corporate generative AI pilots delivered no measurable profit-and-loss return.

In other words, the money flowing into AI infrastructure is real, and the revenue of AI model makers and cloud providers is rising fast. The open question is whether end customers will earn enough from AI to justify the trillions being committed. That gap is where a bubble, if there is one, lives.

9. Seven warning signs to watch

1. Credit spreads: CoreWeave's and Oracle's credit default swap prices are real-time stress gauges.
2. Capex guidance: any cut in 2027 spending plans by the big four would change the story. Meta reports on 21 October, with Microsoft, Alphabet and Amazon expected in the final week of October.
3. Nvidia's results on 17 November, especially guidance and data-centre growth.
4. AI IPOs: how the market receives Anthropic's planned listing, reportedly targeted for November, and whether other AI-linked IPOs are delayed.
5. Interest rates: the Fed meets on 27–28 October; higher yields make debt-funded data centres more expensive.
6. Revenue versus run-rate claims: watch for further downward revisions like OpenAI's.
7. Power and permitting: delays in grid connections could strand expensive projects.

10. What it means for ordinary investors

• Check your concentration. If you own a US index fund, a third of it may already be in a handful of AI-linked giants.
• Distinguish between profitable leaders and leveraged followers. The risk is highest in heavily indebted, loss-making companies dependent on a few customers.
• Do not try to time the top. Bubbles can inflate for years; Grantham notes that the classic signs of a final top had not yet appeared in early 2026.
• Diversify across sectors, countries and asset classes, including bonds, with 10-year US Treasuries now yielding around 5%.
• Think long-term. Even if prices fall, the technology may still transform the economy, as the internet did after 2000.

This article is general information, not investment advice. Consider speaking to a qualified financial adviser before making decisions.

11. Frequently asked questions

Is AI a bubble in 2026?

Parts of it show bubble-like features, especially debt-funded data centres and very high expectations for loss-making AI companies. But the biggest spenders are highly profitable, and demand for chips still exceeds supply, so it is not a simple repeat of 2000.

How much is being spent on AI?

The four largest US tech companies guided for roughly $720–745 billion of capital spending in 2026, and forecasts for 2028 exceed $1 trillion.

Why is debt such a concern?

A growing share of AI investment is funded by bonds and private credit. If returns disappoint, losses could spread to lenders and the wider financial system, as happened with telecom debt after 2000.

What could pop an AI bubble?

Possible triggers include disappointing AI revenues, a big customer cutting spending, a credit event at a leveraged AI company, rising interest rates or power shortages delaying projects.

Would an AI crash be as bad as the dot-com crash?

Nobody can know. The leading companies today are far more profitable than most dot-com firms, which may cushion a fall, but valuations on some long-term measures are close to 2000 levels.

Sources

• Goldman Sachs – How AI debt is reshaping credit markets (Aug 2026)
• Bank of England – Financial Stability Report, July 2026
• IMF – Global Financial Stability Report, October 2026
• Reuters via Yahoo Finance – IMF flags risks to global economy from AI market correction (13 Oct 2026)
• Amazon – Second quarter 2026 results
• CFO Dive – Microsoft holds the line on AI spending plans (30 Jul 2026)
• CNBC – Stock investors are ready to move on from the AI rout. Credit investors are not (19 Oct 2026)
• AP via ABC News – How major US stock indexes fared, 16 Oct 2026
• Benzinga – Chip stocks endure longest losing streak since 2006 (16 Oct 2026)
• FactSet – S&P 500 earnings season update (16 Oct 2026)
• NVIDIA – Financial results for second quarter fiscal 2027
• US Census Bureau – AI use by businesses (May 2026)
• GMO – Valuing AI: extreme bubble, new golden era, or both? (Jan 2026)
• IEA / Utility Dive – PJM capacity auction and data-centre demand