Paul Kedrosky: Nvidia Is Now the AI Bubble's Single Point of Failure

The current surge in AI-related capital expenditure represents a systemic risk reminiscent of historical infrastructure bubbles, fueled by the dangerous intersection of loose credit, government policy, and speculative real estate. Hyperscalers and their financial partners are increasingly relying on off-balance-sheet financing and special purpose vehicles to fund data center expansion, effectively transforming prime corporate credits into subprime risks. Meanwhile, large language models are rapidly commoditizing, with performance variance collapsing and token prices facing hyper-deflationary pressure. This creates a mathematical trap where companies must achieve exponential growth just to maintain current margins. Furthermore, the industry’s reliance on software engineers as a primary market segment provides a distorted view of future demand, as most enterprise applications are compressive rather than expansive. Consequently, the massive buildout of AI infrastructure faces overdetermined failure as market realities clash with speculative investment models.