The case for an AI bubble: the strongest bear arguments

What is the strongest bear argument for an AI bubble?

The first bear argument is arithmetic. The EQOS Asia AI study counts about $200B a year invested in AI research and deployment (EQOS Asia AI study, April 2026). Spending at that scale needs matching revenue from buyers who keep paying. The bear observes that measurable results lag the spending, and reads the lag as a gap that investors currently fill with expectations instead of earnings.

A fair bear states the test: revenue per customer must rise until it carries the build-out. If AI revenue grows faster than spending, the argument weakens each quarter. If spending grows faster, the argument strengthens. This makes the bear case falsifiable, which separates it from a mood. You can track both lines on a single sheet and write the date beside each reading.

Why do failed projects matter to the case for an AI bubble?

The second argument concerns deployment. MIT 2025 studied generative-AI projects in companies and found 95% failing to reach production, a figure the study cites. The study ranks the causes: insufficient or biased data at 42%, no alignment with strategy at 31%, team resistance at 24%, unproven ROI at 22%, and ethics or bias issues at 18%. Shares overlap because one project can have several causes. The causes are structural, so teams fix them slowly.

For the bear, this means revenue growth depends on fixes that money alone cannot buy quickly. Each failed project is a customer who delays the next purchase. For the bull, the same table shows a learning curve. The disagreement therefore sits in the speed of learning, which you can observe by counting how many pilots reach production each year in your own organisation.

How does fast price growth support the case for an AI bubble?

The third argument concerns price. The study records valuations up ×4 over 18 months, and a rise that fast needs earnings to follow within a similar span. A bear compares the speed of the price with the speed of the cash flow. When the price multiplies by 4 and cash flow moves less, the difference is expectation, and expectation can change quickly.

Credit adds a second lens. A price rise funded by customers is a boom. A price rise funded by new investors is a bubble. The bear asks who pays for each new dollar of spending, and public filings and funding announcements answer that question. Single deals belong to the dated filings themselves, so the method here is to read who pays and to record the date of each reading.

What do AI hallucinations add to the bear case for a bubble?

The fourth argument concerns trust. The study reports that 47% of enterprise users took at least one major decision based on hallucinated content, and that language models produce false answers between 0.7% and 30% of the time depending on model and domain (EQOS Asia AI study, April 2026). A bear expects buyers who meet this problem to slow their spending until controls exist.

Controls cost time and money. The study measures the weakest dimension of the enterprise AI market, communication, at 0.305, and the explainability of deployed models sits at 0.25 out of 1. When nobody can explain why a model recommended an action, managers hesitate to scale it. Slower scaling means slower revenue, and the bear links that delay back to the first argument.

What does a fragile score say about the AI bubble?

The fifth argument rests on structure. The study scores the enterprise AI market at φ(S) = 0.337, in the fragile zone where the system works and a shock can tip everything. The score aggregates 9 dimensions through a geometric product, so the weakest dimension weighs as much as the strongest. A bear reads a fragile score as a market that tolerates little bad news.

Communication sits only 0.005 above the critical threshold of 0.3 (EQOS Asia AI study, April 2026). A discovered bias, a public hallucination, or a data leak could push it below and pull trust down with it. The study’s predicted failure rate of about 94% sits close to the 95% that MIT observed. The score frames the risk and leaves the timing open.

What observations would weaken the case for an AI bubble?

A bear case earns trust when it names its own disproof. Four observations would weaken it: revenue growing faster than spending, more pilots reaching production, valuations tracking earnings, and the φ(S) score rising toward the 0.5 viability threshold. The study lists 4 corrections that lift the score to 0.401 and lower the predicted failure rate from 94% to 82%. Each correction is an action a company controls.

Each observation is public or computable, so you can check the case every quarter. Mark a pillar green when its disproof appears and red when it stays absent. Three or more green pillars mean the bear case is losing ground. The same logic applies in reverse to the bull side, which has its own page in this series. You can recompute the score on the Λ simulator.

How can you use the bear case for an AI bubble in a decision?

Use the bear case as a stress test, not as a forecast. Ask what your plans look like if each pillar holds. Which AI purchases still pay back if revenue lags? Which contracts allow a pause? Which suppliers depend on one customer? These questions turn an argument into a list of decisions that you can stage and review, and they cost little to ask.

EQOS Asia applies the same scoring to a single decision. If your decision involves AI spending in Singapore, ASEAN, or Hong Kong, a 30-minute pre-diagnostic returns a score and a ranked set of options, and the method page shows how the score is built. This page is general information, not investment advice.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile