Odd Lots Why Paul Kedrosky Says AI Is Like Every Bubble All Rolled Into One
573 snips
Nov 14, 2025 
In this discussion, Paul Kedrosky, a venture investor and tech cycle analyst, dives into the current AI landscape, likening it to a 'meta bubble' that combines elements from past bubbles like real estate and tech. He argues that the significant spending on AI, including massive data center investments, is unsustainable. Kedrosky explains intricate financial mechanisms like SPVs and highlights risks associated with GPU assets and securitization. With contrasting insights on US and China’s strategies, he prompts a reevaluation of AI's future profitability and potential macroeconomic impacts.
AI Snips
Chapters
Books
Transcript
Episode notes
Securitization Hides Underlying Risk
- Securitization creates a flywheel where buyers ignore what happens inside data centers and focus only on yield.
- That disconnect can amplify mispriced risk, as buyers don't need to understand AI demand to buy tranches.
Poor Unit Economics For LLMs
- Current-generation LLMs have roughly linear cost scaling with usage, producing poor unit economics.
- That contrasts with typical software where fixed costs spread across users, making AI revenue models fragile.
Justifying CapEx Requires Heroic Assumptions
- You can justify huge AI CapEx only with optimistic TAM or consumer subscription assumptions.
- Kedrosky cautions those top-down models rely on unrealistic shares of global labor or massive subscriber fees.
