“Is AI a bubble?” Bulls and bears often hold the same numbers and argue their way to opposite conclusions.
Nobody can reliably predict whether or when a bubble will burst. So this piece doesn’t give an answer. It gives you a ruler: four metrics to read together, plus the main arguments from each side, so you can check for yourself. To see who is actually making money, read it with the AI Stocks Money-Flow Map.
Core-Data Snapshot
| Metric | Data | Nature |
|---|---|---|
| Big-four cloud capex, 2026 (Amazon, Alphabet, Microsoft, Meta) | About $720B–$745B combined | Official guidance (July 2026) |
| NVIDIA CFO’s estimate of top-five cloud capex | Nearly $800B in 2026, about $1.3T in 2027 | Official earnings call (August 2026) |
| NVIDIA data-center quarterly revenue | About $89B, up about 117% year over year (as of July 2026) | Official filings |
| Oracle remaining performance obligations (RPO) | About $664B (end of August 2026) | Official filings |
| Sequoia’s “$600 billion question” (2024) | Annual revenue needed, as estimated then: about $600B | Venture view |
| Bain estimate | About $2T in annual revenue needed by 2030 | Consulting report |
| OpenAI annualized revenue | About $50B (end of September 2026, company’s own method) | According to the FT |
| Anthropic annualized revenue | Over $47B (May 2026) | Official announcement |
Annualized revenue (ARR) is not the same as full-year revenue. The two companies count revenue sold through partners differently, so the numbers can’t be compared directly.
Why Some People Cry Bubble
The bears’ core worry fits in one line: spending too fast, earning it back too slowly.
The model companies are still burning cash. According to the WSJ, OpenAI’s operating loss in Q2 2026 was about $12.3 billion (including stock-based compensation), more than its revenue for the same quarter. CoreWeave, a neocloud that rents out GPUs, also posted a net loss of about $630 million that quarter.
The giants are also raising outside money. According to Axios, citing S&P, by early June 2026 the five largest cloud companies had raised about $255 billion that year through stock and bond sales. Meta’s free cash flow in Q2 was only about $780 million. In September 2026 the head of the International Monetary Fund (IMF) also warned that if profits disappoint, leverage and circular financing could turn that disappointment into a broad correction.
Why Some People Say It’s Not a Bubble
The bulls’ core rebuttal also fits in one line: demand is real, and the people paying can afford it.
NVIDIA’s data-center revenue is still growing fast. Oracle’s order book (revenue already signed but not yet recognized) keeps getting thicker, which bulls read as real demand already placed. The model companies have good news too: according to Bloomberg, Anthropic’s adjusted operating profit turned positive in Q2 2026 (excluding stock-based compensation).
The big spenders this round are Alphabet, Amazon, Microsoft, and Meta, giants whose core businesses make money, not startups living on borrowed money. So bulls compare it to electrification or the railways: scary upfront costs, but the foundation lasts a long time.
A Ruler: Four Metrics to Read Together
Watch these four together; even one worsening can carry risk.
- Capex vs. payback. Look at capex as a share of revenue. In February 2026 Morgan Stanley estimated this ratio (including finance leases) for several cloud giants at about 38%, 44%, and 45% for 2026 through 2028. If it runs far ahead of revenue for several years in a row, be alert.
- Revenue quality. Private model companies’ annualized revenue is not audited, and the methods differ. In September 2026 the market circulated a figure of nearly $70 billion for OpenAI’s annualized revenue. In October, according to the FT, OpenAI’s own method put it at only about $50 billion. The gap comes down to whether revenue sold by cloud partners counts in full.
- The degree of circular financing. When chipmakers and clouds invest in model companies and also collect money from them, demand can be pushed up by insiders’ own money. The full money web is in the AI Circular Financing Map.
- Chip-depreciation life. Are book profits padded by depreciation assumptions that run too long?
Why Depreciation Life Is So Pivotal
Cloud companies spread the cost of the GPUs they buy over several years (depreciation), mostly 5 to 6 years. But AI chips turn over fast, and some argue their economic life is only 2 to 3 years. If the assumed life is too long, the yearly cost comes out too low and book profits get padded.
Examples: in 2025 Amazon cut some servers’ life from 6 years to 5, which reduced net income by about $1 billion. Meta went the other way, extending most servers to about 5.5 years and booking about $2.9 billion less depreciation that year.
Can History Offer a Lesson?
The most common analogy is the dot-com bubble of the late 1990s: telecom companies borrowed heavily to lay fiber, demand didn’t keep up, and many went under. The difference is that the spenders back then were mostly indebted startups, and fiber lasts for decades while AI chips turn over much faster. The 19th-century railway mania is another reminder: investors may lose money, but the rails stay. Which script AI follows depends on whether what is built today can be used long enough and fully enough.
Penchan’s summary
Having a ruler is more useful than listening to bulls and bears shout: capex vs. payback, revenue quality, circular financing, and depreciation life. All four are worth watching over time.
This piece lays out a framework and both sides’ arguments. It makes no call on any stock or the broader market and is not investment advice. The judgment is yours.