Why Paul Kedrosky Says AI Is Like Every Bubble All Rolled Into One
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this episode of Odd Lots, hosts Joe Weisenthal and Tracy Alloway explore the massive AI infrastructure buildout with Paul Kedrosky, a partner at SK Ventures and longtime tech observer. The conversation delves into why Kedrosky believes we're experiencing the ultimate "meta bubble" that combines all the elements of previous financial bubbles. (06:40) The discussion covers how data center spending has become a significant driver of US GDP growth, representing about 50% of first-quarter growth. (07:05) The episode examines the complex financing structures emerging around AI infrastructure, including SPVs (Special Purpose Vehicles) and the role of private credit in funding these massive capital expenditures.
- Main Theme: The AI infrastructure boom represents an unprecedented combination of real estate speculation, technology investment, loose credit, and potential government backstops, creating what Kedrosky calls a "meta bubble" that incorporates elements from every major historical bubble.
Speakers
Joe Weisenthal
Co-host of Bloomberg's Odd Lots podcast and Bloomberg Opinion columnist. He has been covering markets, economics, and finance for over a decade, providing insights into complex financial phenomena and market dynamics.
Tracy Alloway
Co-host of Bloomberg's Odd Lots podcast and Bloomberg Opinion columnist. She specializes in markets, credit, and financial innovation, with extensive experience covering global financial markets and economic trends.
Paul Kedrosky
A partner at SK Ventures and fellow at the MIT Institute for the Digital Economy. He is a longtime internet blogger, writer, and investor who has been analyzing technology markets and financial trends for decades. Kedrosky has become particularly focused on the data center buildout and AI infrastructure investment in recent years.
Key Takeaways
The AI Bubble Combines All Historical Bubble Elements
Paul Kedrosky argues that the current AI boom represents a "meta bubble" that uniquely combines every major ingredient from historical financial bubbles. (09:00) This includes real estate speculation (data centers as commercial real estate), technology hype, loose credit conditions, and the potential for government backstops. Unlike previous bubbles that typically featured one or two of these elements, the AI buildout incorporates all of them simultaneously, making it potentially more dangerous and harder to manage when it eventually unwinds.
Massive Temporal Mismatch Creates Refinancing Risk
There's an unprecedented mismatch between the 30-year loan terms used to finance data centers and the 18-24 month useful lifespan of GPUs that generate the income to service those loans. (29:30) This creates constant refinancing risk, as operators must continuously replace their income-producing assets while maintaining long-term debt obligations. The situation becomes more precarious when considering that GPU depreciation schedules were extended just as the technology shifted toward more intensive, shorter-lived applications.
Unit Economics Remain Fundamentally Broken
Current AI models suffer from negative unit economics, meaning costs rise linearly with usage rather than being spread across more users like traditional software businesses. (25:51) Various models to justify the spending require unrealistic assumptions, such as every iPhone user paying $50 annually or capturing 10% of the global labor market. These mathematical justifications rely on either impossible consumer adoption rates or fragile customer concentration among a few large API users.
Compute Hoarding Drives Irrational Market Behavior
Established hyperscalers like Meta are purchasing compute capacity from neo-clouds not because they need the services, but to prevent competitors from accessing that capacity. (37:00) This hoarding behavior treats compute as a commodity to be cornered rather than a service to be used efficiently. The strategy reflects the scarcity mindset driving the industry, where companies prioritize denying resources to competitors over optimizing their own operations.
Energy Constraints Create Stranded Asset Risk
The power grid cannot support the massive electricity demands of AI data centers, forcing companies to build their own natural gas plants with 25-30 year lifespans. (30:57) This creates a dangerous mismatch between the long-lived energy infrastructure and the uncertain future of AI demand. Companies risk creating stranded assets—expensive power plants that outlive the data centers they were built to serve, leaving operators with massive sunk costs and no revenue source.
Statistics & Facts
- Data center spending represented approximately 50% of US GDP growth in the first quarter, with even larger impacts when including externalities and secondary effects. (07:05) This statistic demonstrates how AI infrastructure has become a major driver of economic growth, functioning as an unintentional private sector stimulus program.
- Around 50% of hyperscaler companies' free cash flow was being directed toward data center spending, reaching approximately $500 billion in total investment. (12:04) This represents the point at which public companies felt they needed alternative financing structures to continue expanding without further diluting earnings per share.
- The private credit industry has grown to approximately $1.7 trillion, which is larger than many components of the traditional lending market combined. (10:47) This growth has enabled much of the AI infrastructure financing through non-traditional channels, essentially replacing what was once called "shadow banking."
