Paul Kedrosky: AI is the First Bubble With Every Ingredient at Once | #648

Episode Summary

Executive Summary: The conversation argues that AI is driving a bubble-like capital cycle similar to past manias, but with a unique mix of loose financing, policy, technology hype, and real-estate-style infrastructure buildout. Paul Kodrowski says early AI adoption in coding misled markets into overbuilding data centers and power systems, while token prices are deflating rapidly, models are converging, and financing is shifting from cash flow to debt and exotic structures.

Main Topics: AI as a historically unusual bubble: Kodrowski compares the current AI boom to past U.S. bubbles and says it uniquely combines loose credit, a powerful technology narrative, real estate/data-center buildout, and policy support. Exponential growth, deflation, and misunderstanding scale: He emphasizes that humans struggle with exponentials and that AI is simultaneously scaling fast in usage and deflating fast in unit economics, making the situation hard to interpret. Data-center financing and the shift from cash flow to external capital: A central claim is that AI infrastructure spending has moved beyond internally funded capex and is now increasingly financed through ABS, private credit, sovereigns, and SPVs, creating bubble dynamics. Tokens, unit economics, and deflationary pressure: Kodrowski argues AI tokens are a hyper-deflationary commodity, with prices falling so fast that frontier companies must grow at extraordinary rates just to maintain revenue. Model convergence and the end of model differentiation: He says model performance gains have largely flatlined, frontier models are converging, and harnesses make products look better without meaningfully changing the underlying models. Data centers, chips, and oversupply risk: The discussion covers hoarding, over-ordering, low GPU utilization, and the coming wave of chip and data-center oversupply that could eventually collapse prices. Public backlash and agency concerns: Kodrowski links opposition to data centers and AI to people’s sense of lost control, especially in the U.S., where job displacement is tied to healthcare loss and financial insecurity.

Key Arguments: The current AI cycle resembles past bubbles because it combines multiple ingredients at once: cheap capital, a compelling technology story, policy implications, and heavy infrastructure investment. Early AI market projections were distorted by coding use cases, which are unusually favorable to LLMs and are not representative of broader white-collar adoption. AI tokens are deflating at a historically unprecedented rate, so companies need explosive unit growth just to offset price declines. Financing has shifted from internal cash flows to external funding, which signals that the cycle has become financialized and more bubble-like. Model performance improvements have slowed sharply and frontier models are converging, reducing the case for massive new training runs. The data-center buildout is creating distortions in capital markets, including competition with sovereign borrowing and pressure on longer-term rates. The eventual correction may come from multiple overdetermined failure modes, including financing stress, oversupply, regulatory reaction, or simply valuation compression. Public resistance to AI and data centers is partly rational because people associate them with lost agency, job risk, and, in the U.S., healthcare insecurity.

Data Points: AI token price decline: 70% to 80% per year - Used to describe tokens as a hyper-deflationary commodity and the basis for extreme growth requirements. Minimum growth to stand still: 400% year-over-year - If token prices fall 80% annually, frontier companies need roughly 4x unit growth just to maintain revenue. GPU utilization in a VPS warehouse: 35% to 40% - Presented as evidence that large amounts of AI hardware capacity may be underused despite scarcity narratives. Investment-grade market share from data centers: 15% to 18% - Kodrowski says data center-related issuance now makes up a major share of the IG market and exceeds financial services if treated as a sector. OpenAI recent quarterly growth: 18% quarter-over-quarter - Cited as being viewed as disappointing, illustrating how high expectations have become. Year-over-year model improvement: 1% to 2% recently, down from 10% to 12% - He argues frontier model performance gains have flattened substantially over the last six months. AI share of GDP growth: More than 50% - He says AI/data centers accounted for over half of U.S. GDP growth in the relevant period. New issuance from a narrow subset of companies: $4 trillion to $5.5 trillion - Used to illustrate the scale of forthcoming IPO and financing pressure from major AI companies. IPO scale comparison: Potentially larger than all post-World War II IPOs combined - He says a few expected AI IPOs could exceed the total historical issuance of the modern era. Completion of financing transition: First half of 2026 - He says by mid-2026 more than half of data center financing came from external sources rather than internal cash flows.

Pivotal Quotes: "It’s going to take a lot of AI to solve the problems created by AI in climate alone." — Paul Kodrowski: He is explaining that AI’s own power and emissions footprint is creating downstream problems even as people claim AI will solve climate issues. "This moment is the first one that sits at the intersection of all of the forces that created the largest bubbles in U.S. history." — Paul Kodrowski: He frames the current AI/data-center boom as unusually bubble-prone because multiple historical bubble ingredients are present at once. "Tokens are the first hyper-deflationary commodity in the history of modern economies." — Paul Kodrowski: He uses this to argue that AI economics are structurally deflationary and unusually dangerous for frontier companies.

Implications: Listeners should expect falling AI unit economics, more financing stress, and a likely re-rating of AI companies and utilities. The biggest risks are oversupply, leverage, and overhyped adoption assumptions; the main winners may be consumers and AI users, not frontier investors.