E155.似乎没什么人再提「AI 泡沫论」了

Summary

This 面基 episode argues that the “AI bubble” debate has cooled because generative AI can now be tracked through a business flywheel: CAPEX improves models, better models create more token demand, token demand converts into paid usage, and ARR or deferred revenue validates the loop. The discussion contrasts OpenAI’s consumer-route possibilities with Anthropic’s enterprise and coding focus, especially around Model Context Protocol, AI Skills, Claude Code, and agent products. It then extends the investment frame from software into Human Resource Deflation Compute Infrastructure Inflation, CAPEX OPEX Substitution, energy constraints, Holo Assets, Nvidia, and Hang Seng Tech Index repricing.

Key Claims

• The episode treats tokens as a better leading indicator for generative AI than MAU, DAU, or time spent, because token growth can reflect model usefulness, usage depth, and production adoption.
• Jevons Paradox In AI means lower token cost can increase total token demand as users, applications, and agents call models more often.
• The best answer to “AI bubble” anxiety is not rhetorical optimism but improving AI Investment Metrics such as CAPEX, contract liabilities, deferred revenue, AI-native revenue, and ARR.
• The guest frames Anthropic as more enterprise- and coding-oriented than OpenAI, while OpenAI is treated as more consumer-entry oriented through ChatGPT, possible traffic monetization, take rate, or advertising.
• Coding is treated as the first widely successful agent domain because code is standardized, toolable, and testable; the episode expects this pattern to generalize into finance, legal, medical, and other vertical workflows.
• Model Context Protocol is described as a unifying connector layer for databases, GitHub, Slack, ERP, and other external systems; together with AI Skills and Claude agent products, it becomes part of Anthropic’s ecosystem strategy.
• Language User Interface weakens traditional software’s GUI stickiness because users can ask an agent to submit a travel request, reimburse an expense, or operate a workflow without touching the original UI.
• B2B software know-how may be partly externalized into skills, reducing the moat of workflow-heavy SaaS products even when the products themselves do not disappear.
• Human Resource Deflation Compute Infrastructure Inflation summarizes the episode’s labor-market and investment thesis: AI can reduce demand for some white-collar labor while raising demand for compute, storage, power, networking, and data centers.
• CAPEX OPEX Substitution turns current labor or operating savings into AI infrastructure investment that may lower future OPEX and improve model performance.
• Energy, chips, data centers, cooling, and power grids are treated as hard constraints on AI output, making AI Compute Continuity, Data Center Physical Resilience, and Holo Assets part of the AI investment map.
• The discussion of Hang Seng Tech Index and Chinese internet assets argues that low valuation is not enough if the assets lack hard-infrastructure scarcity, Holo attributes, or new separately priced AI-native businesses.
• The social extension is ambivalent: AI may release human time and create conditions for original innovation, but wealth may concentrate unless institutions, redistribution, or UBI-like mechanisms adapt.

Key Quotes

“MCP 是 AI 世界里的 USB Type-C” — the episode’s metaphor for unified agent connectivity.

“人力资源行业的通缩正在转化为算力基础设施的通胀” — the investment thesis linking labor savings to compute demand.

“AI 的尽头是电力” — the physical constraint behind token production.

“Language 成为新的 UI” — the episode’s shorthand for natural language as the user interface.

Connections

• 面基 — show context for the episode.
• AI Investment Metrics — tokens, CAPEX, contract liabilities, deferred revenue, AI-native revenue, and ARR as the episode’s tracking framework.
• Jevons Paradox In AI and AI Inference Cost Structure — why cheaper tokens can still expand total compute demand.
• Anthropic, OpenAI, Google, ChatGPT, and Claude Code — frontier company and product comparison set.
• Model Context Protocol, AI Skills, Agentic Workflow, and Agentic Software — agent ecosystem and software-interface layer.
• Headless Software, Agent-Facing Interfaces, and Language User Interface — software-entry and GUI-displacement implications.
• Human Resource Deflation Compute Infrastructure Inflation and CAPEX OPEX Substitution — labor-to-infrastructure investment thesis.
• MaaS Infrastructure, AI Compute Continuity, and Data Center Physical Resilience — infrastructure layer behind token supply.
• Holo Assets, Nvidia, and Hang Seng Tech Index — hard-asset and public-market interpretation.
• AI Equity Valuation Risk, Investment Risk Management, and Value Investing — adjacent valuation and risk-control frames.

Contradictions

• No direct contradiction found. The source complements earlier AI Equity Valuation Risk pages by arguing that some AI fundamentals are becoming more measurable, while still leaving valuation, data-source quality, and entry-price risk unresolved.
• Data caution: the episode cites ARR, user-count, token-growth, ETF-return, and energy-use figures that should be independently verified before being used as current market facts.