Skip to content

Note · AI

What is the AI bubble? The debate behind the spending boom

The AI bubble is the worry that AI share prices and spending on chips and data centers have run far ahead of the profits AI will earn. Each side's case, dated.

By Graphite

An AI model, Claude Opus 5.5, wrote it for Graphite from the sources at its end.

· 7 min read

This note and the comments marked the Graphite mark are Graphite’s: an AI model, Claude Opus 5.5, wrote them on 5 November 2026, following Graphite’s own instructions, from the sources at the note’s end, and checked them against those sources. Check them as you would anyone’s.

This note and its comments are written by an AI model.

Contents · 7 sections

"The AI bubble" is the name for a worry, not an event: that what investors pay for artificial intelligence companies, and what those companies spend on chips and data centers, has run far ahead of the profits AI will earn. Since 2023 the question has split investors, central bankers and economists. One side sees the dot-com bubble again. The other sees a real technology built by companies that already make money.

In short

  • A bubble is a price that has run far ahead of the earnings that could justify it, and it is usually only certain in hindsight.
  • The case for a bubble rests on stretched valuations, a few companies carrying the market, and deals in which suppliers fund their own customers.
  • The case against rests on the profits the leading companies already earn and the cash they pay with, which most dot-com firms never had.

What is an economic bubble?

A bubble is a rise in the price of an asset, such as a share, a house or a coin, far beyond what its future earnings can reasonably support, kept going because buyers expect someone else to pay more later. When that expectation breaks, prices fall fast. Bubbles usually form around something real, which is why they are hard to call from inside: the internet did change the world, and most dot-com shares still collapsed. The dot-com bubble note walks through the usual stages, from a real breakthrough to euphoria to panic.

The simplest yardstick is the price-to-earnings ratio, or P/E: the share price divided by a year's profit per share. Suppose a company earns $10 a share and its shares cost $250. Its P/E is 25, so a buyer is paying 25 years of today's profit, an earnings yield of 1 in 25, or 4%. If profits double in five years to $20, the same $250 price is only 12.5 times earnings, and the buyer looks wise. If profits stay at $10, the buyer is earning 4% a year on an asset that can lose half its value in a bad year. Whether a price is a bubble depends on profits that haven't happened yet, which is why reasonable people disagree.

How big is the AI spending boom?

It is among the largest corporate building programs on record. Training and running large language models takes huge numbers of specialized chips, housed in data centers that need their own power supplies. In April 2026, Alphabet, Amazon, Microsoft and Meta together planned more than $700 billion of capital spending for the year, nearly double what they spent in 2025, according to Sherwood News tallying their own forecasts.

Share prices followed. Nvidia, which makes most of the chips used to train AI, became the first company valued at $4 trillion on 9 July 2025 and passed $5 trillion on 29 October 2025. In October 2025 the Bank of England's Financial Policy Committee noted that the five largest companies in the S&P 500 made up close to 30% of the index, a higher share than at any point in the past 50 years. When a few companies carry an index, a fall in them reaches almost everyone who owns an index fund.

Why do some say it is a bubble?

The first argument is price. The same Bank of England record called valuations "stretched" and said one long-run measure of US share prices, the cyclically adjusted price-to-earnings ratio, put the expected return from shares close to its lowest in 25 years, "comparable to the peak of the dot com bubble." It listed disappointing progress in AI's abilities or adoption, or more competition, as the risks that could bring a sharp fall.

The second is circular money. On 22 September 2025, Nvidia and OpenAI announced a letter of intent under which OpenAI would deploy at least 10 gigawatts of Nvidia systems and Nvidia would invest up to $100 billion in OpenAI, "progressively as each gigawatt is deployed." Critics see a supplier funding a customer to buy its own products, which can make demand look larger and more certain than it is. Deals like this also tie the fortunes of chipmakers, cloud providers and AI labs together, so trouble at one spreads to the others.

The third is the gap between spending and payoff. In a survey of nearly 9,000 executives in seven countries, reported in an NBER working paper revised in October 2026, 78% of firms said they used AI, but executives estimated it had raised their labor productivity growth by only about 0.2% a year over the previous three years, and expected about 0.6% a year over the next three. Skeptics read that as a small payoff so far for spending on this scale.

Why do others say it isn't?

The strongest answer is profits. Asked about the 1990s at a press conference on 29 October 2025, Federal Reserve Chair Jerome Powell said "these companies—the companies that are so highly valued—actually have earnings," adding that in the dot-com period "these were ideas rather than companies." The biggest spenders pay for most of their building out of large, existing businesses in advertising, software and cloud computing, not out of money raised from share sales.

Prices are high, but not dot-com high on every measure. The Bank of England noted that the S&P 500's price relative to expected profits, about 25 times, "remained below the levels reached during the dot com bubble." And the demand is real enough that most firms in the NBER survey already use AI. Those who doubt the bubble add that even a fall in share prices would leave the chips, power lines and buildings in place, as the fiber laid in the 1990s carried the broadband that came after.

Dot-com and AI side by side

Dot-com era, 1995–2002 AI build-out, 2023–2026
What was bought Fiber-optic cable, network equipment, websites, advertising AI chips, data centers, power supplies
How it was paid for Share sales by young companies, and heavy borrowing by telecom firms Mostly cash from the largest companies' existing businesses, plus bonds, private loans and deals between suppliers and customers
Who earned money Equipment makers such as Cisco, briefly the world's most valuable company in March 2000; most start-ups never made a profit Chipmakers such as Nvidia and the cloud providers that rent out computing; whether selling AI services will cover the cost is still open
Valuations The Nasdaq Composite peaked at 5,048.62 on 10 March 2000 and fell 78% by October 2002 S&P 500 at about 25 times expected profits in October 2025, below dot-com levels, with its top five near 30% of the index

Still argued

Whether AI is a bubble is argued among central banks, investors and economists, and nobody can know in advance. The Bank of England in October 2025 warned that valuations looked stretched and that a correction could be sharp. Powell, the same month, said the leading companies' earnings made this different from the 1990s. Both readings fit the facts available in late 2026, because they differ about profits that haven't arrived. This note explains the debate; it isn't financial advice.

What would tell the two views apart?

The two sides make different predictions, so the evidence that settles it can be named in advance. These are the signs analysts on both sides watch:

  • Revenue from AI itself. Whether money paid for AI services, not for the chips and buildings, grows fast enough to cover what is being spent.
  • Productivity. Whether the gains executives expect show up in surveys like the NBER's and in official statistics.
  • Who pays, and how. Whether spending keeps coming from profits or shifts toward borrowing, which turns a fall in prices into losses for lenders too.
  • Where demand comes from. Whether chip orders come from end customers or from deals in which suppliers fund the buyer.
  • The price of computing. Falling prices to rent AI chips would point to more capacity than buyers need.

A bubble is a gap between price and future profit, and the gap only closes one of two ways: profits rise to meet the price, or the price falls to meet the profits. For the ideas behind the spending, start with what artificial intelligence is. Accounting matters here too: Enron showed how numbers can make a business look healthier than it is.

Check yourself

  • Explain, using a share price and its earnings, why a high P/E can be either wise or a bubble.
  • What is a "circular" deal, and why do critics think it makes demand look stronger than it is?
  • Name two differences between how the dot-com boom and the AI build-out were paid for.

Sources

Published by Graphite on , set the way Graphite publishes a note.

This note and the comments marked the Graphite mark are Graphite’s: an AI model, Claude Opus 5.5, wrote them on 5 November 2026, following Graphite’s own instructions, from the sources at the note’s end, and checked them against those sources. Check them as you would anyone’s.

About the author

Graphite

An AI model, Claude Opus 5.5, wrote it for Graphite from the sources at its end.

More in AISearch the MarketplaceWho writes here

Published with Graphite