Is AI Pricing a Bubble? Two AI Models Tested It

The short answer: both AI models reached the same verdict by different routes. AI share prices are not in a dot-com-style bubble: the leaders are profitable and the median AI infrastructure stock trades at about 22 times forward earnings. The bubble risk has moved into spending: record capital spending on data centers and chips, a growing share of it borrowed, and supplier profits that last only as long as that spending does.

Educational only, not financial advice. The figures on this page are as each AI model reported them, with the model's own sources listed at the end. We did not re-check them, and where the two models disagree we say so.

An article argued that "AI pricing is not a bubble." Its evidence: the median forward P/E of AI infrastructure stocks fell from 32 in April to 22, even as those companies kept delivering strong results, and Goldman Sachs reads this as pricing discipline rather than speculation.

Instead of taking the article's word for it, we treated the claim as a hypothesis and asked two AI models to test it against the bubble that defines all bubbles for traders: the dot-com (WWW) bubble of 1999 to 2000. Each model ran on its own paid professional account, with web research switched on, and received exactly the same prompt:

"I read in an article that 'AI pricing is not a bubble', explained by: valuation multiples for AI infrastructure stocks have fallen from a median forward P/E of 32 in April to 22 today, even as these companies continue to deliver strong results. Goldman Sachs argues this reflects strong pricing discipline rather than a speculative bubble. Now I want you to go out there and make deep research that confirms or does not confirm this statement. Do it by comparing it to the dot-com (WWW) bubble that hit traders. Make a table of your findings, call it AI vs WWW Bubble, and at the end summarize your verdict on the AI bubble. Test the statement as a hypothesis." (Lightly edited for spelling.)

• Model 1: GPT 5.6 SOL (OpenAI, ChatGPT). Its answer states its evidence runs to 7 October 2026.
• Model 2: Claude Sonnet 5.5 (Anthropic, Claude). Collected in the same week.
• Then a third opinion: Claude Opus 5.5, the model that built this page, read both answers and wrote the final verdict in section 05. It did no new research.

Headline

"The public-market AI boom in October 2026 does not resemble the final, extreme stage of the 1999 to 2000 dot-com bubble on valuation or corporate quality. But the AI capital-spending cycle increasingly does resemble the late-1990s infrastructure boom, and that is where the more serious risk now lies." GPT confirmed the 32x to 22x figure, but said its interpretation "needs considerably more nuance."

GPT's table: AI vs WWW Bubble

GPT marked each row with a color (green, red, orange, white). We wrote those out as "Against a bubble", "Bubble risk", "Mixed" and "Neutral" so the meaning does not depend on color. Wording shortened; numbers and sources as GPT gave them.

GPT's key argument: the earnings may be the bubble

GPT agreed that a P/E falling while earnings rise is a sign of discipline, the opposite of 1999. Its example: a stock at $320 with expected earnings of $10 trades at 32x; at $330 with expected earnings of $15 it trades at 22x. The price went up, yet the stock got cheaper.

But it warned that the earnings themselves could be inflated. AI companies buy chips, networking, memory, cooling, electrical equipment and data centers, and those purchases become other AI companies' earnings. That loop, heavy spending, then supplier profits, then lower P/Es, then confidence to spend more, can run for years. If spending slows, supplier earnings fall and "today's 22x wasn't really 22x." Its watch list: AI spending growth, AI revenue and return on investment, infrastructure use, debt-financed spending, supplier margins and earnings revisions.

GPT's conclusion

• AI technology: not a bubble. Adoption, revenue and demand are real.
• Large listed AI leaders: not comparable to 2000. Earnings and cash flows make the valuation picture different.
• AI infrastructure investment: increasingly bubble-like. Spending is enormous, past the 1990s share of the economy, more of it is borrowed, and supplier earnings depend on others continuing to spend.
• Some AI valuations: very aggressive. $27 trillion of added market value against a $9 trillion economic-value estimate is "a serious yellow or red flag."

GPT's verdict: "Not Dot-Com 2.0, but an AI capex and earnings bubble is now a credible risk." If the cycle breaks, it expects it to look like the telecom suppliers after 2000, not Pets.com.

Headline

Claude broke the statement into its four separate claims and tested each before comparing the two eras. Its headline: "AI stock prices are not in a dot-com-style bubble today. The AI investment cycle, however, carries real bubble risk that the 32x to 22x figure doesn't capture."

Claude's claim-by-claim test

Claude's table: AI vs WWW Bubble

"Supports the claim" means supports the statement that AI pricing is not a bubble. Claude's own score: 5 factors support it, 7 contradict it, 1 is unproven.

Claude's conclusion

The data rules out a dot-com-style bubble in listed AI stock prices: the leaders are profitable, trade at roughly half of 2000-era multiples, and there is no IPO mania. It does not show that AI as a whole is free of a bubble, because the risk has moved into spending, debt and private valuations:

• Spending is larger relative to GDP than in the dot-com era.
• AI revenue does not yet cover the cost of that spending.
• Borrowing and circular deals are rising.
• Private valuations are higher than Cisco's at its peak.

Claude's verdict: "The hypothesis is partly supported, and the reasoning behind it is flawed." The deciding question, in its view: does AI revenue reach the roughly $300 billion a year needed to break even before the debt-funded build-out slows?

Every factor from both tables, matched where the two models covered the same topic. A star (*) marks a topic that appears in only one model's table. GPT's 20 rows and Claude's 13 rows make 23 topics: 9 shared, 10 only in GPT's table and 4 only in Claude's.

* Covered by only one model's table. Of the 9 shared topics the models agree on 5, point the same way with different strength on 3, and disagree on 1. Where the other model raised the same idea outside its table, the cell says so.

• Where the numbers differ: the dot-com Shiller CAPE is 43.8 in GPT's answer and 44.19 in Claude's. GPT counts global AI investment (over $1 trillion in 2026), Claude counts hyperscaler spending ($800 billion), so both can be right.
• The biggest difference: GPT accepted the article's line that Goldman calls the de-rating "pricing discipline." Claude found no source where Goldman uses that framing, and read Goldman's note as investors doubting that AI profits will last.

I read both answers as a referee: no new research, only what the two models found and how they reasoned. Here is how I would grade the hypothesis.

What the two answers show when read together

• The one disagreement is the key to the whole question. GPT found the S&P 500 at a normal 19x forward earnings; Claude found the Shiller CAPE near its dot-com peak. Both are true. The forward P/E uses next year's expected profits, the CAPE uses ten years of inflation-adjusted profits. The gap between them exists because profits have jumped recently, and the jump is largely AI spending flowing through to suppliers. So the market looks cheap only if today's profit level is permanent. That is GPT's "earnings bubble" and Claude's "investors doubt the profits will last" in one number.
• Two methods, one answer, but not fully independent. GPT built a wide table; Claude took the claim apart first. They still landed in the same place. Both drew heavily on Goldman Sachs research, though, so their agreement is partly an echo of one source, not two separate confirmations.
• Claude was the stricter auditor. It was the only one to question whether the quote in the prompt was accurate, and it caught that a low P/E can be a warning. GPT went wider, adding the $27 trillion against $9 trillion gap and the Fed's stability warnings.
• The historical match is telecom, not Pets.com. Both models arrived at the same parallel: the internet was real, the fiber was overbuilt, and the suppliers' profits vanished when spending stopped. AI is following the build-out script far more than the stock-mania script.

My verdict

AI pricing is not a bubble today; AI spending might be. The article is right about prices and wrong about what its own number proves. Listed AI leaders are valued on real profits at roughly half or less of dot-com multiples, and the 32x to 22x de-rating shows investors are not paying any price for an AI label. But the profits that make 22x look reasonable are being paid for by the same AI build-out the bubble debate is about, and a growing share of that build-out is borrowed. If the spending slows before AI revenue catches up, today's reasonable P/E will turn out to have been high.

The signals both models point to, in one list:

• Hyperscaler spending growth: the first number to turn if the cycle turns.
• AI revenue against the roughly $300 billion a year that Claude says 2026 to 2027 spending needs to break even.
• The share of spending that is borrowed: about a third in 2026 on GPT's figures.
• Earnings estimates and margins at infrastructure suppliers, which fall before P/Es look expensive.
• The gap between the forward P/E and the Shiller CAPE: the wider it gets, the more the market is betting that today's profits are permanent.

For the physical side of the AI build-out, see AI Data Centers vs US Electricity, How AI Drives Metal Demand and How to Invest in AI.

Is AI pricing a bubble?

Not in the dot-com sense, on the evidence both AI models gathered in October 2026. Listed AI leaders are profitable and the median AI infrastructure stock trades at about 22 times forward earnings, far below dot-com peaks of 50 to over 100 times. Both models found real bubble risk elsewhere: in record capital spending, rising debt, private-company valuations, and earnings that depend on that spending continuing.

What does the fall from 32 to 22 times earnings show?

Both models confirmed the figure: Goldman Sachs found the median AI infrastructure stock fell from 32 times forward earnings in April to 22 times. Prices rose less than expected earnings. That rules out a price-only bubble, but a falling P/E can also mean investors doubt today's profits will last, so on its own it does not prove there is no bubble.

How is the AI boom different from the dot-com bubble?

The leaders make large profits, valuations are roughly half or less of 2000 levels, there is no IPO mania, and investors now separate companies that earn from AI from those that only spend on it. The similarity is the infrastructure build-out: spending on data centers, chips and power is now as large relative to the economy as the 1990s telecom boom, which left most of its fiber unused at first.

Where is the AI bubble risk now?

In spending and the earnings it creates, not in share prices. Hyperscalers plan around 800 billion to over 1 trillion dollars of AI spending, a growing share is borrowed, and AI suppliers' profits depend on that spending continuing. Claude estimated about 300 billion dollars a year of AI revenue is needed to break even on 2026 to 2027 spending, and current revenue is still below that.

Did GPT and Claude agree on the AI bubble?

On the verdict, yes. GPT 5.6 SOL called it not dot-com 2.0 but a credible AI capex and earnings bubble risk. Claude Sonnet 5.5 called the hypothesis partly supported with flawed reasoning, and placed the risk in spending, debt and private valuations. They disagreed on broad market valuation, and Claude found no source where Goldman Sachs called the de-rating pricing discipline.

Why did the two AI models report different numbers?

They measured different things. GPT used the S&P 500 forward P/E of about 19 times, near its 10-year average, while Claude used the Shiller CAPE of about 42, close to the dot-com peak. GPT counted global AI investment of over 1 trillion dollars in 2026, while Claude counted hyperscaler spending of 800 billion. Both can be right at once.

The sources each model cited, as it gave them, collected in early October 2026. We did not re-check each figure against its source.

Cited by GPT 5.6 SOL

• Goldman Sachs research: Are US stock market valuations outpacing fundamentals?, Can AI investment drive S&P 500 earnings even higher?, Global investment is forecast to exceed $1 trillion in 2026, Why AI companies may invest more than $500 billion in 2026, and How AI debt is reshaping the credit market.
• The 32x to 22x figure: citybiz, Goldman Sachs: AI infrastructure valuations fall as spending climbs toward $1.1 trillion.
• Federal Reserve: Financial Stability Report, May 2026 (overview) and near-term risks, Governor Jefferson's speech of 21 November 2025, Monetary Policy Report, February 2001, a speech of 20 January 2000, and FEDS working paper 2007-15.
• Dot-com valuations and IPOs: Nasdaq, Is AI another bubble for the Nasdaq-100? and NBER Reporter, Fall 2002.
• News: Bloomberg, 25 September 2026: Goldman sees hyperscaler AI capex rising 50% to $1.2 trillion, Reuters, 5 October 2026: Goldman sees US data center growth intact, and two Financial Times articles (one, two) on AI spending and earnings, and on cooling IPO demand.

Cited by Claude Sonnet 5.5

• Goldman's guide to the AI valuation reset (Seeking Alpha, via TradingView)
• Goldman Sachs hikes S&P 500 target (MarketWatch, via Yahoo Finance)
• Goldman Sachs says AI not yet in a bubble, but private valuations raise red flags (Wealth Professional)
• Fear of an AI bubble: this time is different, Goldman says (Benzinga, October 2025, a year older than the rest)
• AI stocks and the dot-com bubble compared (The Next Web)
• Is this 1999? What the AI bubble question gets wrong (ETF Trends)
• Dot-com bubble vs AI bubble (Lambda Finance)
• AI bubble vs dot-com bubble (Complete Trader's Edge)
• How to spot an AI bubble (Capital Group)

Images from Pixabay (free to use).

Partner platforms (sponsored). We may earn a commission if you sign up. Not a recommendation or financial advice.

This guide is free to use, supported by affiliate partnerships. Some broker and tool links are sponsored, and we may earn a commission if you sign up, at no extra cost to you. This never affects what we cover or how we explain it.