# 【旧世代读书】AI泡沫现在到哪一步了？以史为鉴的《泡沫逃生》读书与感想

- Channel: [旧世代的Lunamos](https://www.youtube.com/channel/UCxIFQMoQu8M-i3PogVz_ynw)
- Published: 2026-11-13
- Source video: [Watch on YouTube](https://youtube.com/watch?v=-Jx-rSSfT70)
- Read on ReadTube: https://www.read.tube/p/videos/-Jx-rSSfT70
- Language: English

## Summary

**AI Bubble Lessons From Two Centuries**

The video argues that AI can genuinely transform society while still producing severe losses for early investors, as shown by railways, electricity, oil, automobiles, telecommunications, computing, and the internet. The greatest risks are excessive infrastructure spending, leverage, uncertain future demand, falling prices, and choosing the wrong technical or commercial winner. AI may avoid a single catastrophic bubble because demand is still evolving and semiconductor capacity remains constrained, but weaker infrastructure providers and overvalued companies could still face consolidation and painful write-downs. The speaker’s broader advice is to study historical patterns, avoid treating technological importance as proof of investment value, and remain cautious rather than going all in.

The video uses *Engines That Move Markets* to examine whether the current AI boom resembles earlier technology bubbles. Its central warning is that a technology can transform society while most early investors still lose money, because technological success, corporate profitability, and investment returns follow different timelines.

- **Real technological progress does not make every related asset a good investment.** British railways became essential infrastructure, but early success on profitable routes encouraged excessive construction, weak projects, leveraged share purchases, and unreliable dividend payments. When interest rates rose and liquidity tightened, many companies failed even though railways continued to transform the economy. The same pattern appeared in oil, electricity, automobiles, and telecommunications: the industry survived, but many early shareholders did not.

- **Leverage and capital-intensive expansion are the most immediate risks for AI infrastructure.** The video compares AI data centers and cloud-computing providers with railway companies, independent telephone networks, and early electricity firms. Companies may borrow against long-term customer contracts before that revenue is secure, while GPU depreciation, power costs, leases, and debt payments remain fixed. If demand slows or customers renegotiate, providers such as CoreWeave or Nebius could be forced to cut prices, restructure debt, or sell assets cheaply, causing losses for existing shareholders while stronger buyers acquire the infrastructure at a discount.

- **Choosing the winning technology or company is exceptionally difficult.** The history of electric lighting shows investors repeatedly backing the wrong route: gas lighting, arc lamps, Edison’s direct-current system, and eventually Westinghouse’s alternating-current system each had periods of apparent dominance. Even technically successful founders could lose control of their companies because construction costs and debt overwhelmed cash flow. Similar uncertainty exists in AI, where NVIDIA faces alternatives such as ASICs and Cerebras, while large companies with distribution advantages could eventually use their platforms to overtake specialized AI firms, much as Microsoft defeated Lotus.

- **Competition can destroy profits even when demand is genuine.** Oil producers expanded output so aggressively that prices collapsed, while early television and automobile manufacturers entered destructive price wars and generated excess capacity. By contrast, radio found an unexpected mass-market model when broadcasting turned a lack of privacy into an advantage, and personal computing created durable software economics around Microsoft and Intel. The lesson for AI is that today’s visible applications may not be the ultimate commercial breakthrough: a new “broadcast moment” or other unforeseen use could accelerate demand, while cheaper models, local hardware, or open systems could compress prices.

- **The speaker sees AI as unusually powerful but not necessarily headed for one catastrophic collapse.** He argues that AI’s capabilities are already producing meaningful scientific and engineering results, and that demand may keep expanding as users expect increasingly capable systems. Unlike some earlier connection-based infrastructure revolutions, AI is presented as a node-centered expansion of machine intelligence, with semiconductor and memory supply still constrained. His base case is therefore uneven profit redistribution, consolidation, and repeated soft landings rather than a total industry failure—but he stresses that this is a historical framework, not a direct investment recommendation.
