AI Bubble or Revolution?

What Happened

ChatGPT launched in November 2022 and reached 100 million users faster than any product in history. The generative AI moment triggered the largest technology investment cycle since the dot-com era. Microsoft, Google, Amazon, and Meta collectively committed to spending hundreds of billions on AI infrastructure — data centers, chips, power, and cooling.

Nvidia became the most valuable company in the world briefly in 2024 as demand for its H100 and Blackwell GPUs proved insatiable. The stock rose 10x from its 2022 lows. Every major enterprise announced AI initiatives. Every consulting firm published AI transformation reports. AI became the organizing framework for every technology investment decision.

The skeptical case: the capex cycle is running ahead of the revenue cycle by a widening margin. Sequoia estimated a $600B revenue gap to justify current infrastructure investment. The DeepSeek shock of January 2025 suggested frontier AI models could be built at a fraction of the assumed cost, potentially undermining the hardware demand thesis. The question: are we in 1995 (early days of a genuine revolution) or 1999 (peak of a speculative bubble)?

The Mechanism

The capex-revenue gap and the 1999 parallelThe dot-com bubble was driven by genuine technological transformation with investment running far ahead of revenue. The infrastructure built in 1999-2000 (fiber, routers, servers) powered the internet revolution of the 2000s — but investors who funded it mostly lost money. The AI investment cycle has the same structure: real technology, real investment, uncertain and delayed revenue. The telecom companies that funded fiber in 1999 were right about demand but wrong about timing. AI infrastructure investors face the same risk.

What the Consensus Believed

The prevailing view before the reckoning

AI represents the largest technology investment opportunity in history. The productivity gains are real and imminent. Companies that do not invest now will be left behind permanently. GPU demand will grow exponentially for a decade. The capex is justified by future revenue that will make current valuations look cheap.

What the Record Shows

Infrastructure investment is not the same as investment returns

The fiber built in 1999 was used. The companies that built it mostly went bankrupt. Amazon used the bandwidth to build AWS. Timing and competitive position matter as much as the technology.

The revenue cycle always lags the capex cycle

In every technology revolution the infrastructure is built before the applications that monetize it. The question is whether the lag is 2 years (reasonable) or 10 years (painful for investors).

Concentration risk is the hidden variable

Nvidia derives most of its data center revenue from 6 customers. If any one of them cuts guidance, the impact is catastrophic. Concentration is the AI bubble's most underappreciated risk.

DeepSeek changed the calculus

If frontier models can be built at a fraction of the assumed cost, the GPU demand thesis is undermined. The $350B capex cycle was predicated on continued scaling laws. DeepSeek suggested efficiency improvements may substitute for raw compute.

↑ Cognitive pattern: Exponential Extrapolation — Projecting current AI capability growth indefinitely

Key Voices

Bulls — Genuine Revolution

Jensen Huang

Nvidia

“We are at the iPhone moment of AI. The next wave will require 100x more compute than today. Demand is insatiable.”

January 2024 Pending

Sam Altman

OpenAI

“AGI is closer than most people think. The economic implications are transformative. This is the most important technology humanity has created.”

March 2024 Pending

Larry Fink