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

Paul Kedrosky discusses why the current AI moment represents a historically unique convergence of bubble-creating forces. Unlike previous bubbles focused on single factors—railroads, electrification, or real estate—the AI bubble simultaneously exhibits loose credit conditions, a genuinely transformative technology story, massive real estate allocation (data centers), and supportive policy environments. This multi-factor intersection makes the bubble's scale and consequences difficult for humans to comprehend, particularly given our cognitive limitations with exponential functions.

Kedrosky emphasizes that AI adoption began with an atypical early user base: software engineers. Software development represents an unusual domain because it features tight grammar (code either works or fails), rapid feedback loops (gradient descent), and expansive rather than compressive applications. Most white-collar work, by contrast, is compressive—taking large amounts of information and distilling it into actionable insights. Because AI was optimized for software engineers' needs, the projected data center capacity and GPU requirements now appear massively misaligned with actual token usage, which Kedrosky estimates runs 35-40% utilization rates despite narratives of scarcity.

The financing structure has fundamentally shifted. Through mid-2026, more than 50% of hyperscaler data center financing comes from external sources (asset-backed securities, private credit, sovereigns, SPVs) rather than internal cash flows. This represents a critical inflection point because companies must now service debt obligations while facing unprecedented deflation in token pricing—falling 70-80% annually. To maintain flat unit economics, frontier AI companies must grow revenues 400-500% year-over-year just to stand still, yet Wall Street expects 50-100% growth. This creates a structural Wile E. Coyote scenario where executives must maintain unsustainable growth rates.

Kedrosky discusses how this financing shift is distorting sovereign debt markets. Data center credit has become more attractive to lenders than sovereign debt because hyperscalers have implicit credit backing and 12-year renewable leases, creating secure cash flows. This reverses historical patterns where sovereign fundraising crowds out private financing. The resulting capital flows are now so large they're pushing longer-term rates higher and creating bond market friction.

On technological convergence, Kedrosky argues that large language model performance has essentially plateaued. Year-over-year composite benchmark improvements have fallen from 10-12% to 1-2%, with variance across frontier models collapsing—Claude, Qwen, and DeepSeek produce nearly indistinguishable results behind harnesses (wrapper applications that mask underlying model stagnation). This convergence, combined with the homogeneity of training data (median data source: 37-year-old Reddit males), means the industry cannot justify multi-billion dollar training runs on performance grounds alone.

Regarding semiconductor supply, Kedrosky anticipates a boom-bust cycle typical of capital-intensive industries. Unprecedented cash flows into Taiwanese and Chinese chip manufacturers suggest a tsunami of GPU supply arriving in early 2028. Once capacity is locked in, semiconductor prices historically collapse as manufacturers must move inventory to cover fixed costs, following historical patterns from previous semiconductor cycles.

On public offerings, upcoming IPOs from SpaceX and Anthropic alone would create $4-5.5 trillion in new issuance, more than all post-World War II IPOs combined (inflation-adjusted). This forces long-only managers to sell liquid, high-performing stocks to fund new allocations, creating selling pressure on the exact names most supporting recent market gains. This anticipated reallocation likely contributed to market friction, including pressure on situational awareness (an AI-focused portfolio fund).

Kedrosky concludes that at high valuations, failure becomes overdetermined—many low-probability failure modes combine to create high overall failure probability. With 20+ ways the AI complex could unwind, each individually unlikely, the statistical probability of implosion within the relevant period exceeds 60%. The system exhibits Nike-like characteristics: high PE multiples mean almost any business outcome could trigger decline.

On public sentiment, Kedrosky notes Americans uniquely oppose AI compared to other developed nations, attributing this partly to healthcare being tied to employment (job loss means healthcare loss and personal bankruptcy risk). Developing nations view AI positively as a pathway to economic participation, while developed nations fear homogenization and lost opportunity.