The “AI bubble” debate is the running argument over whether the money pouring into AI has outrun the value it produces. The bearish case holds that the pattern rhymes with past technology bubbles: enormous capital spending, circular deals between a few firms, soaring valuations, and a revenue base far too small to justify the outlay. The bullish case holds that this is a justified buildout of durable infrastructure - like railroads or electricity - where heavy upfront spending precedes, rather than follows, the returns.
The bears point to several documented facts. Sequoia’s David Cahn estimated in June 2024 that AI products would need around 600 billion dollars in annual revenue to justify the infrastructure being built, far above what the industry earned. Deals among the central players became deeply interlinked - NVIDIA saying it intends to invest up to 100 billion dollars in OpenAI, a major buyer of NVIDIA chips, is the emblematic case - raising the worry that demand was partly circular. Surveys showed enterprise AI adoption still in the low tens of percent. And on January 27, 2025 a single cheap model erased nearly 600 billion dollars of NVIDIA’s value in a day, showing how fragile the consensus was.
The bulls counter with their own facts. Per-token costs are collapsing (LLMflation), which by the Jevons-paradox argument should expand total demand rather than shrink it. Leading labs are posting fast revenue growth - Anthropic went from about 1 to over 5 billion dollars of run-rate revenue in eight months in 2025. And data centers and power, unlike dot-com-era fiber, are general-purpose assets that retain value even if specific bets fail. The honest answer is that the same observations are visible to both sides; what differs is the forecast.
Why a business reader should care: “bubble or not” is not an abstract question. It bears on whether to invest in AI-exposed equities, how aggressively to build AI into a business, and how much to trust headline deal sizes. The most useful posture is to watch the gap between real revenue and real spending close or widen over time, rather than to bet the whole question on a single number or a single bad day.