AI bubble fears: why Wall Street is nervous about AI spending

Meta took in $64.6 billion in the three months to September and still spent more cash than its business brought in.

That result, published on October 21, 2026, was followed by a 7% fall in Meta’s shares the next day. It fed a worry that has run through markets all October: that the money going into artificial intelligence is running far ahead of the money AI earns. People call it the AI bubble question. In plain terms, investors are asking whether today’s spending on chips and data centers will produce enough profit to justify it, or whether some of it will have to be written off.

This is an analysis. It brings together reported figures, history and published research, and sets out who argues what, with attribution. It does not forecast where markets will go.

Key figure: Meta’s free cash flow, the cash left after paying for equipment and data centers, was negative $2.12 billion in the third quarter of 2026, according to its earnings release. Capital spending in the quarter was $34.91 billion, against $32.79 billion of cash from operations.

Where the AI bubble fears show up in the figures

The worry runs from the biggest tech groups to the start-ups that rent out AI computing power. The table brings the main October figures together.

Company or market	Figure	Reported by
Meta, Q3 2026	Revenue $64.6 billion; capital spending $34.91 billion; free cash flow negative $2.12 billion	Meta, October 21
Meta, full year 2026	Capital spending of $135 billion to $145 billion expected	Meta, October 21
OpenAI	Talks to raise up to $30 billion at a $1.25 trillion valuation before new money	CNBC, October 22
Anthropic	IPO valuation of $1.5 trillion to $2 trillion discussed, up from $965 billion in May	CNBC, Reuters, Anthropic
CoreWeave	About $50 billion in debt and lease obligations; some bonds yielding above 15%	Bisnow, SiliconANGLE
US IPO market	$148.3 billion raised by 118 IPOs in 2026, above 2021’s $142.4 billion	Renaissance Capital, as of October 25

Source: Meta earnings release; Renaissance Capital; reports as listed.

Meta’s quarter, and why a profitable company spooked investors

Meta’s revenue grew 26% from a year earlier, and it forecast $73 billion to $76 billion for the fourth quarter. The problem for investors was the bill. Meta narrowed its 2026 capital spending forecast to $135 billion to $145 billion and said that, beyond 2026, it expects “to continue investing aggressively” in its core business and new AI projects.

Shares closed down 7% at $648.85 on October 22, according to CNBC and other market reports. CNBC also reported that the cost of insuring Meta’s five-year debt against default, measured by credit default swaps, hit a record. Those swaps work like insurance on a company’s bonds, and their price rises when lenders grow more cautious. Meta held $82.45 billion in cash and marketable securities at the end of September. Against that cushion, the swap move reads less as a default warning than as a sign that lenders, like shareholders, want to know how long the spending will last.

The accounting choice that decides how expensive AI looks

A data center does not hit profits all at once. Companies spread the cost of servers over their expected useful life, and the length they choose has a large effect on reported earnings.

Meta’s annual report for 2025 shows how large. In January 2025 it extended the estimated useful life of most of its servers and network equipment to 5.5 years, from a range of four to five years. That single change cut 2025 depreciation expense by $2.92 billion and added $2.59 billion to net income, or $1.00 a share, according to the filing. Meta’s depreciation on servers and network assets still rose to $13.36 billion in 2025, from $7.32 billion in 2023.

This is where skeptics and defenders talk past each other. If AI chips stay useful for five or six years, today’s spending looks reasonable. If faster chips make them uneconomic in three, the true cost of the build-out is higher than current earnings show. Nobody yet has a full cycle of data on how long the newest AI hardware earns its keep, so the choice rests on management judgment, disclosed in a footnote that most readers never reach.

Valuations that keep climbing before the profits arrive

OpenAI, the maker of ChatGPT, is negotiating a round of up to $30 billion at a $1.25 trillion valuation before the new money, with Coatue and investors from the United Arab Emirates leading, CNBC reported on October 22. Reuters had reported days earlier that OpenAI was seeking more than $1.4 trillion. Its last round, in March, valued it at $852 billion. The terms are not final, and OpenAI declined to comment to TechCrunch.

Anthropic, which makes the Claude models, was valued at $965 billion in May. Reports now put the valuation it is seeking in a planned IPO at $1.5 trillion to $2 trillion. The Wall Street Journal reported that a draft filing showed a first-half 2026 net loss of $80.7 billion, of which about $72 billion was a non-cash accounting charge. Our Anthropic IPO explainer covers the filing process and what has been confirmed.

The neocloud squeeze, and Nvidia’s place in it

“Neoclouds” are companies that buy large numbers of Nvidia chips and rent the computing power to AI developers. Many have paid for that build-out with borrowed money.

The largest, CoreWeave, has close to $50 billion in debt and lease obligations, Bisnow and SiliconANGLE reported. Yields on some of its bonds passed 15% in mid-October. A bond’s yield rises when its price falls, so the jump meant investors were selling the debt and demanding a bigger return to hold it. CoreWeave shares fell more than 15% on October 14 alone, SiliconANGLE reported.

The stress reached new listings outside the US. Elice, a South Korean AI data center company, fell 32% on its first day of trading on the Kosdaq market on October 20, after raising 201.1 billion won ($151 million). Firmus, an Australian AI cloud company backed by Nvidia, shelved its planned listing, Yahoo Finance and Reuters reported.

Nvidia sits on both sides of many of these deals. In September 2025, it agreed to buy up to $6.3 billion of CoreWeave’s unsold computing capacity. In January 2026, it invested a further $2 billion in CoreWeave at $87.20 a share. Critics call this “circular financing”: the chipmaker puts money into customers who then spend it on its chips. CNBC described the October slide in neocloud stocks as the first big test of Nvidia’s role as a financier of the AI build-out.

What the 1840s railways and the 1990s fiber boom suggest

Investors reaching for comparisons usually land on two earlier building sprees.

The first is Britain’s Railway Mania. Andrew Odlyzko, a University of Minnesota mathematician who studies the history of technology manias, has called it by many measures the greatest technology mania in history. His study of the period argues that reliable figures existed at the time to show there would not be enough rail traffic to earn the profits investors expected, and that critics said so loudly. Their warnings failed, he found, partly because the most prominent critics were themselves wrong about where the danger lay. For today’s skeptics, the awkward point is that being early and partly right did not stop the mania, which ran from 1844 until its collapse began in late 1845.

The second is the fiber-optic build-out of the late 1990s, when telecom companies laid cable on the belief that internet traffic was doubling every three or four months. Odlyzko and his colleague Kerry Coffman checked the data and found traffic was roughly doubling once a year: fast, but nowhere near enough to fill what was being built. Global Crossing filed for bankruptcy in January 2002 and WorldCom in July 2002, then the largest bankruptcy in US history.

The parallel is closer for some companies than for others. The biggest spenders, such as Meta, pay for data centers mainly with cash from businesses that already exist: Meta’s operations generated $32.79 billion in the third quarter alone. That is a different position from a debt-funded carrier selling a single product. The neoclouds look more like the 1990s carriers: young companies borrowing heavily to build capacity on the strength of contracts with a small number of customers. That is why the first visible strain in this cycle has appeared in CoreWeave’s bonds and in small data center IPOs, and only lately in a giant like Meta, whose quarterly cash flow just turned negative.

The case that the spending will pay

Supporters point to revenue that is growing fast as well.

Meta’s chief executive, Mark Zuckerberg, said in the earnings release: “Our investments are paying off, the core business is strong, and we’re seeing velocity across the board.” Meta’s finance team wrote that its AI investments are already delivering substantial extra advertising revenue.

Anthropic told investors its third-quarter revenue rose to $16.8 billion from about $11.5 billion and that it was profitable on an adjusted basis, the Financial Times and Reuters reported. That is an increase of about $5 billion in a single quarter.

On the debt side, Goldman Sachs upgraded CoreWeave to buy after the sell-off, arguing that because many of its deals use asset-backed financing, its cost of borrowing should not rise significantly, CNBC reported. CoreWeave’s chief executive, Michael Intrator, told staff in a memo that the company’s order backlog more than covers its debt, Bloomberg reported. And demand for big listings has held up: SpaceX raised $75 billion in June in the largest IPO on record, and US IPOs have raised more money in 2026 than in any full year since at least 2016, according to Renaissance Capital.

The case that it won’t, at least not soon

Skeptics focus on the gap between spending and the profits needed to justify it, and some of the sharpest arguments are not new.

In June 2024, David Cahn, a partner at the venture firm Sequoia Capital, published an essay called “AI’s $600B Question.” He worked from Nvidia’s chip sales to estimate the total cost of the data centers they go into, then asked how much revenue AI products would need to earn to pay for them. His answer was about $600 billion a year, far more than the industry was then making.

The economist Daron Acemoglu of the Massachusetts Institute of Technology, a Nobel laureate, reached a cautious view from a different direction. In a 2024 paper, “The Simple Macroeconomics of AI,” he estimated that AI would raise total factor productivity, the economy’s efficiency in turning inputs into output, by no more than 0.66% over 10 years. That is a total for the decade, not a yearly figure.

The research firm Gartner said on October 14, 2026, that more than $3.5 trillion of AI infrastructure spending could be exposed to write-downs by the end of 2030 if spending continues at its current pace. It predicted that more than 80% of neoclouds will stop operating independently by 2030, through takeover, partnership or failure.

New Constructs, an independent research firm, titled an October 6 report “Anthropic Is The Most Ridiculous IPO of 2026.” Its lowest estimate valued Anthropic at about $144 billion, and it called the offering “an unprecedented test of investor gullibility.” Its middle case, about $1.1 trillion, was still well below the reported target.

The dates that will test both sides

• 27 to 28 Oct 2026: The US Federal Reserve holds its policy meeting. Interest rates affect the cost of borrowing, including the debt behind many data center projects; see how interest rate decisions work.
• 28 Oct 2026: Microsoft reports quarterly results after the US market close, a week after Meta.
• 30 Oct 2026: End of the week in which OpenAI told investors it aimed to sign its new funding round, according to CNBC’s October 22 report.
• Week of 9 Nov 2026: Anthropic’s IPO roadshow could begin, Bloomberg and Reuters reported, if a public filing comes in time.

One number is worth tracking through all of it: each big spender’s capital spending set against its cash from operations. Meta’s was $34.91 billion against $32.79 billion last quarter. While cash from operations stays above spending, the build-out pays for itself. When it does not, the money has to come from cash reserves or from lenders, which is how this cycle’s strains surfaced first at debt-funded companies like CoreWeave. For how these moves feed into the share prices most savers hold, see our stock market index explainer.

This article is for information only and is not financial advice.