Boom, bubble, bust, boom. Why should AI be different? 🔗 **Source:** [Boom, bubble, bust, boom. Why should AI be different?](https://crazystupidtech.com/2025/11/21/boom-bubble-bust-boom-why-should-ai-be-different/) > [!note]+ Document Notes > tags: [AI, investment, leverage] > author: Fred Vogelstein > source: Crazy Stupid Tech > reference: [https://crazystupidtech.com/2025/11/21/boom-bubble-bust-boom-why-should-ai-be-different/] > date created: > > # 🏷️ Title: Boom, bubble, bust, boom. Why should AI be different? > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Include links cited in or relevant to the article. ## 📝 Summary - What happened: The authors argue today’s AI boom is a larger, faster, and riskier bubble than 1999’s internet era, driven by unprecedented spending, leverage, circular deals, and China competition. - Why it matters: A downturn could trigger massive write-downs across hyperscalers, chipmakers, LLM labs, and private credit markets. - Key stakeholders: Big Tech (Microsoft, Alphabet, Amazon, Meta, Oracle, NVIDIA), labs (OpenAI, Anthropic), infra (CoreWeave), investors, policymakers. - Risks and tradeoffs: Overbuild vs. missing the platform shift; vendor-financing contagion; geopolitical shocks; grid/power constraints. - Bottom line: AI is transformative long term, but the near-term mania is fragile and likely to correct before durable returns arrive. ## 📖 Main Content - Thesis: The AI cycle resembles 1999 but bigger and faster. Annual AI capex+VC is already ~$600B, with estimates up to $1.5T for 2025, versus ~“about $360 billion in 2025 dollars” for 1999’s internet bubble ([scale comparison](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0000G&endLexposition=0000M_00000-003S&emphasizeText=about+%24360+billion+in+2025+dollars)). - Market signs: Despite strong NVIDIA results, megacaps slid recently (MSFT -12%, AMZN -14%, META -22%, ORCL -24%, CoreWeave -47%), suggesting overspending fears ([recent declines](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=00008&endLexposition=0000E_00000-005g&emphasizeText=Investors+are+increasingly+worried+that+everyone+is+overspending+on+AI.)). - Valuation context: Index P/Es aren’t dot-com extreme, but the S&P near ~24 resembles pre-2000 levels ([valuation caution](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0000W&endLexposition=0000e_00000-0010&emphasizeText=the+last+time+the+P%2FE+ratio+of+that+index+was+as+high+as+it+is+now+%E2%80%93+about+24+%E2%80%93+was+right+before+the+internet+bubble+popped)). - Four vulnerabilities: “Too much spending. Too much leverage. Crazy deals. China. China. China.” ([the four risks](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0000q&endLexposition=00018_00000-004k&emphasizeText=Too+much+spending.%0AToo+much+leverage.%0ACrazy+deals.%0AChina.+China.+China.)) - Spending: Hyperscalers are each pouring $70–$100B/yr; Oracle ~$20B/yr, without proven unit economics. OpenAI projects half-trillion server rentals over 5 years, cashflow-positive only by 2029 ([OpenAI spend arc](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0001Q&endLexposition=0001U_00000-0049&emphasizeText=Eventually+we+need+to+get+to+hundreds+of+billions+of+a+year+in+revenue)). - Pricing pressure: “Training is expensive, serving is getting cheap, and pricing power is slipping” (Mary Meeker) ([pricing power slipping](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0001e&endLexposition=0001k_00006-003x&emphasizeText=Training+is+expensive%2C+serving+is+getting+cheap%2C+and+pricing+power+is+slipping.)). - Leverage/vendor financing: NVIDIA and OpenAI act like lenders/investors to secure demand/supply; Microsoft+NVIDIA up to $15B in Anthropic for $30B capacity; Oracle/CoreWeave heavy debt; private credit exposure rising ([circular deals snapshot](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0001m&endLexposition=0001q_00001_00000-003E&emphasizeText=have+begun+acting+like+banks+and+VC+investors)). - Crazy deals: Eye-popping first rounds/valuations (Thinking Machines $2B at $10B; SSI $3B in two years; Anthropic $13B at $170B; 498 AI unicorns) ([deal froth examples](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0002O&endLexposition=0002Q_00002-005J&emphasizeText=valued+the+company+at+%24170+billion)). - China: Tech race, power-grid advantage, open-source push, and narrative risk (e.g., DeepSeek), with Jensen Huang warning China is “nanoseconds behind” ([China race framing](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0002i&endLexposition=0002u_00000_00000-001Q&emphasizeText=China+is+going+to+win+the+AI+race)). - Timing: Mania can flip quickly; even bulls (e.g., Pichai) admit overshoot phases before durable impact ([overshoot admission](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0000k&endLexposition=0000o_00000-004Q&emphasizeText=there+are+moments+where+we+overshoot)). ### 🗒️ Notes - AI boom resembles 1999 but larger/faster; total spend already staggering; durability of demand unproven short term. - Market breadth fragile; S&P valuation reminiscent of 2000; hyperscaler capex at historic scale. - OpenAI: outsized burn, massive forward commitments, long runway to cashflow positivity; big bet on consumer+enterprise products and scientific discovery. - Vendor financing/circular deals raise contagion risk; private credit opacity adds fragility. - Deal heat at seed/Series A unprecedented; many unicorns birthed in 24 months. - China is both accelerant and tail risk: power, chips catching up, policy unpredictability, informational control concerns. ## 📚 References - JP Morgan CEO on private credit “cockroach” risk (FT coverage) - Gartner AI spend estimates (2025) - PitchBook: NVIDIA venture activity - CB Insights: 498 AI unicorns (July 2025) - Alphabet TPU strategy context - Pichai BBC interview on overshoot - Additional reading listed in-article (FT, Reuters, HBR, Brookings, Wired, Goldman Sachs, Platformer, Stratechery) ## 🗣️ Paraphrase - A: The AI surge is likely a bubble bigger than 1999’s, fueled by unprecedented capex, leverage, and geopolitics; transformative long term, messy near term. - B: Expect corrections: pricing pressure, vendor-financing contagion, and China shocks could force write-downs before sustainable revenues arrive. ## 📰 Newsletter Today’s read argues the AI boom echoes dot-com—only bigger. Key takeaways: hyperscaler capex is eye-watering, pricing power is slipping, vendor financing is everywhere, and China looms large. Builders: watch unit economics, avoid circular dependencies, and stress-test for power and supply shocks. For WordPress folks: explore AI responsibly—evaluate TCO, privacy, and latency; consider local inference pilots; and keep eyes on open models and TPU alternatives. ## ✅ Action Items #### DevRel - Publish costed AI reference architectures; bake in spot/fallback plans. - Create SDK examples with local+hybrid inference toggles; add eval suites. - Run office hours on AI TCO, latency budgets, and vendor lock-in. #### OSPO - Codify AI model licensing and data provenance checks. - Establish embargo/signing policies for AI supply chain artifacts. - Maintain SBOMs for AI components; set freshness SLOs. #### Community - Host governance clinics on AI usage; publish code of conduct addendum. - First-issue rotations on eval tooling; pairing for prompt safety. #### WordPress - Core AI: guardrails, eval datasets, opt-in pilots. - Performance: latency budgets for AI features; caching strategies. - Docs: examples-first AI patterns; privacy disclosures. - Hosting: cross-host checklists for GPU/CPU, quotas, and telemetry. #### The WPOCC - Microgrants for open eval tools; maintainer stipends for AI safety docs. - Publish transparent budgets and milestone reports. ### FAIR - Require provenance/signature checks for models/datasets. - Fund with anti-capture covenants and sunset clauses. - Ship reusable charters/MoUs for AI collaborations. ## 🧪 Hypotheses - H1: AI infra capex growth will outpace revenue growth through 2027. - H2: Vendor-financed AI customers show higher default/write-down risk by 2026. - H3: Model serving prices will decline >50% YoY for general-purpose LLMs through 2026. - H4: China-related shocks materially reprice US AI equities within 30 days of event. - H5: Vertically integrated stacks (chips+models+apps) achieve superior gross margins by 2028. ### 📊 Methodology - Compare hyperscaler capex vs. AI-attributed revenue in 10-Ks/8-Ks. - Track vendor-financing disclosures; correlate with distress events. - Collect public/spot prices for inference; regression on cost trends. - Event studies on China policy/tech milestones vs. AI equity baskets. - Margin analysis by stack integration proxies (TPUs, in-house silicon). ### ❓ Questions - What revenue mix (subscription, ads, usage) stabilizes AI unit economics? - How big is private credit’s hidden exposure to AI infra? - Which open-source models best compress serving costs without quality loss? ### 🔮 Future Research - Power grid constraints and AI deployment timelines. - Local/hybrid inference economics vs. pure cloud. - Evaluation benchmarks that predict willingness to pay. ### 💡 Implications - Practice: Stress-test AI features against price drops and supply shocks. - Policy: Increase transparency for vendor financing and private credit. - Security: Track supply-chain and model provenance rigorously. - Education: Teach TCO modeling for AI workloads. - Funding: Tie grants to open evals and reproducible results. ### 🧠 Learning Styles Assessment - Reading level: Flesch–Kincaid Grade ≈ 11–12. #### 👀 Visual learners - Present: Risk lists, examples. - Improve: Capex vs. revenue charts; leverage maps. #### 🔊 Aural learners - Present: Quotes. - Improve: 3-minute audio summary. #### 🗣️ Verbal learners - Present: Definitions, comparisons. - Improve: Glossary for finance terms. #### 🧩 Kinesthetic learners - Present: N/A. - Improve: TCO calculator exercises. #### 🧠 Logical learners - Present: Causal chains. - Improve: Decision trees for build/buy. #### 🧑‍🤝‍🧑 Social learners - Present: Discussion prompts. - Improve: Cohort study groups. #### 🧘‍♀️ Solitary learners - Present: Reflective takeaways. - Improve: Self-check quizzes. #### 🌿 Naturalistic learners - Present: Grid/power context. - Improve: Lifecycle footprint analysis. ### 📊 Accessibility Assessment - Screen reader semantics: Likely adequate headings. - Improve: Alt text for diagrams; language attributes. - Keyboard navigation: Ensure visible focus/skip links. - Contrast/readability: Maintain ≥4.5:1; verify line-height. - Text resizing: Support 200% zoom/reflow. - Media/captions: Provide captions/transcripts; avoid autoplay. - Forms: Labels, error hints; generous hit targets. - Standards: Map to WCAG 2.2 AA. ### 🌍 DEIB Assessment - Cultural breadth: Add global case studies beyond US/China. - Inclusive language: Minimize jargon; add plain-language summaries. - Neurodiversity: Strengthen consistent structure and icon+text labels. - Representation: Balance quotes/sources; inclusive imagery. - Psychological safety: Clear community guidelines and escalation paths. ### ✅ Logical Validity - Strong reasoning: Historical analogy (1999 vs. 2025) supported by spend/valuation data; causal chain from leverage to contagion risk. - Fallacies: Potential availability bias via high-profile deals; mitigate with broader datasets. - Tone of voice: Sober, cautionary, occasionally wry. ### 🛡️ DARVO Analysis - Deny: No denial of risks; acknowledges overshoot. - Attack: None. - Reverse victim/offender: None; focuses on systemic incentives and behaviors. ## 🎯 Conclusion - Bottom line: AI’s long-term impact is profound, but today’s funding/spending/leverage mix is brittle and likely to correct. - Next steps: Build with unit-economics discipline, reduce circular dependencies, diversify compute, and scenario-plan for China/power shocks. ## 🧱 Refactor Appendix (append after 🎯 Conclusion) ### 📚 Literature Split Pad - Purpose: Stage Konik CEE units that become per-source Literature notes. - Split level: H3. - Include verbatim quotes under 25 words and a [view highlight] link. ### [H1] AI bubble exceeds dot-com scale > [!quote] [Boom, bubble, bust, boom. Why should AI be different?](https://crazystupidtech.com/2025/11/21/boom-bubble-bust-boom-why-should-ai-be-different/) by [[Fred Vogelstein]] via [[Crazy Stupid Tech]]. Published 2025-11-21. Accessed . > “Latest estimates… might top $1.5 trillion.” Two lines: context and use. Cite data if relevant. [view highlight](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0000G&endLexposition=0000M_00000-003S) ### [H2] Four systemic AI risks concentrate fragility > [!quote] [Boom, bubble, bust, boom. Why should AI be different?](https://crazystupidtech.com/2025/11/21/boom-bubble-bust-boom-why-should-ai-be-different/) > “Too much spending… Too much leverage… Crazy deals… China.” Two lines: context and use. [view highlight](https://read.readwise.io/read/01kav06hxhqvv8sn0hvadpkrrh?startLexposition=0000q&endLexposition=00018_00000-004k) --- ### 🔖 Atomic Split Pad (Zettel Candidates) #### [Z1] Vendor financing amplifies AI contagion risk - Applies to hyperscaler–LLM circular contracts. - Pointer: NVIDIA/Microsoft–Anthropic arrangement. - Links: [[Vendor-Finance]] [[AI-Systemic-Risk]] #### [Z2] Serving costs fall faster than demand hardens - Applies to general-purpose LLMs 2025–2026. - Pointer: Meeker quote. - Links: [[AI-Unit-Economics]] [[LLM-Pricing]] --- ### ✏️ Fleeting Queue (Extract by Selection) - [F1] Chart AI capex vs. AI revenues for top 5 hyperscalers — context/https://crazystupidtech.com/2025/11/21/boom-bubble-bust-boom-why-should-ai-be-different/ #fleeting #inbox - [F2] Map vendor-financing exposures across NVIDIA/OpenAI/Oracle/CoreWeave — context/{ARTICLE_URL} #fleeting #inbox - [F3] Build inference price tracker for GPT-class APIs — context/{ARTICLE_URL} #fleeting #inbox --- <!-- deep-read:sections end -->