Paul Kedrosky on the AI Bubble: Every Bubble Rolled Into One

The AI Bubble as a Meta-Bubble

• 💡 Paul Kedrosky describes the current AI boom as a meta-bubble, combining elements of every historical bubble: real estate, technology, exotic financing, and potential government backstops.
• 🚀 This unique confluence of factors makes the AI bubble unprecedented and suggests a difficult, non-gentle landing.

Data Centers: The Engine of AI Spending

• 📊 Data center construction has become a massive driver of US GDP growth, accounting for a significant portion of recent economic expansion.
• 💰 The immense capital expenditure (capex) required for AI build-out is a double-edged sword: it can create moats for successful companies but also poses a risk if revenues don't justify the investment.
• 🏠 Data centers are uniquely positioned at the intersection of industrial spending and real estate, making them an interesting, albeit complex, asset class.

Financing Structures and Risks

• 🏦 The AI build-out is increasingly reliant on private credit and complex financing vehicles like Special Purpose Vehicles (SPVs), which allow companies to raise capital without it appearing on their balance sheets.
• ⚠️ A significant risk lies in the temporal mismatch between long-term debt (e.g., 30-year loans) and the short lifespan of underlying assets like GPUs (potentially 18-24 months), creating constant refinancing risk.
• 📈 Blending different tenant types in data centers to improve yields can increase risk, leading to complex securitization structures like tranched securities (ABS).

Unit Economics and Future Demand

• 📉 Current AI models often exhibit negative unit economics, meaning they lose money on each transaction, with costs rising linearly with demand, unlike traditional software businesses.
• 🎯 Justifying the immense spend relies on faith-based argumentation about AGI or highly optimistic projections of market size (TAM), such as capturing a significant portion of the global labor market.
• 💡 While large language models (LLMs) are powerful, the future may lie in smaller, more efficient micro-models for mundane tasks, which are cheaper to train and operate but do not justify the current scale of investment.

The "Existential Competition" and Global Approaches

• 🌍 The AI race is framed as an existential competition, particularly between the US and China, leading to potentially unlimited spending and government backstops.
• 🇨🇳 China's approach, focusing on distillation and efficiency gains with smaller, open-source models, may offer a more sustainable path than the US's focus on massive, closed-source models.
• ⚡ A critical bottleneck is energy supply, with data centers struggling to secure sufficient power, leading some companies to consider behind-the-meter power generation, which introduces its own long-term risks.

The Nature of AI Investment

• 🤝 The current AI investment frenzy is characterized by compute hoarding, where companies lock up capacity to prevent competitors from accessing it, rather than purely for immediate operational needs.
• 🔮 Builders at the cutting edge often oscillate between framing AI as productivity-enhancing tools and the more speculative, faith-based pursuit of AGI or ASI.
• 📉 The massive private sector stimulus from AI capex, if not yielding economic returns, poses a significant negative wealth effect when unwound, impacting not just direct spending but also broader market valuations.