Why Paul Kedrosky Says AI Is Like Every Bubble All Rolled Into One

Episode Summary

Executive Summary: The episode argues that the AI boom is increasingly a capital-intensive credit and real-estate story, not just a tech story. Guest Paul Kedrosky says data centers combine multiple ingredients of past bubbles—real estate, loose credit, technology hype, and implied government backstops—while GPU depreciation, power constraints, and SPV financing create serious rollover and refinancing risk. He also warns that current unit economics and demand forecasts may be overstated, especially if cheaper, more efficient models reduce compute needs.

Main Topics: AI capex as a macro force (Priority: 5/5): The hosts open by framing AI data-center spending as a major driver of GDP growth and a potential macro risk if returns fail to materialize. Data centers as a bubble-like asset class (Priority: 5/5): Kedrosky argues data centers uniquely combine speculative real estate, technology hype, loose credit, and quasi-government backstops, making the current cycle bubble-prone. SPVs, private credit, and balance-sheet engineering (Priority: 5/5): Discussion of how hyperscalers and AI labs use SPVs and private credit to fund projects off-balance-sheet, shifting AI from an equity story to a credit story. Depreciation, GPU lifespan, and refinancing risk (Priority: 5/5): The episode explains that GPU-heavy workloads can wear out hardware far faster than traditional cloud storage, creating a mismatch between short-lived collateral and long-term debt. Unit economics and demand uncertainty (Priority: 4/5): Kedrosky argues many AI businesses have negative unit economics and that revenue projections rely on fragile assumptions about consumer and enterprise adoption. Compute hoarding and vertical integration (Priority: 4/5): Meta, CoreWeave, Anthropic, and others are portrayed as racing to lock up scarce compute capacity, not just to use it immediately but to prevent rivals from getting it. US vs China AI strategy (Priority: 4/5): The conversation contrasts the U.S. model of expensive frontier training with China’s more efficient distillation/open-source approach, which may imply less compute demand than forecast.

Key Arguments: AI data-center spending has become a major and perhaps unintended source of U.S. economic growth, functioning like a private-sector stimulus program. This cycle is unusually risky because it combines several bubble ingredients at once: technology hype, speculative real estate, loose credit, and backstop expectations. SPVs are being used to keep debt off the parent company’s balance sheet, but the legal recourse and future ownership structures can become very messy if projects underperform. GPU collateral can be financed by long-duration debt even though the useful economic life of those GPUs may be much shorter, especially under heavy training workloads. The mismatch between the lifespan of AI hardware and the tenor of the financing creates significant refinancing and stranded-asset risk. Current AI unit economics are weak because costs rise roughly with usage, unlike classic software businesses with strong operating leverage. Many revenue projections assume unrealistically broad adoption, such as consumers or workers paying at enormous scale, or AI capturing a large share of global labor value. Compute is being treated like a scarce hoardable commodity, leading hyperscalers to lock up capacity through deals like CoreWeave even when they may not need all of it immediately. Cheaper, more efficient model-training approaches—especially distillation—could reduce the need for massive compute buildouts and undermine current capex assumptions. The likely future of AI may be more mundane and internal: small models handling tasks like onboarding, matching records, and workflow automation rather than transformative AGI. The AI boom’s potential failure would have broad spillovers because equity, private credit, REITs, and retirement/institutional money are increasingly exposed. An existential national-security framing around AI may encourage unlimited spending and government support, increasing bubble risk.

Data Points: Anthropic U.S. data-center commitment: $50 billion - Mentioned as a recent headline illustrating the scale of AI infrastructure spending. Hyperscaler free cash flow share going to data centers: around 50% - Kedrosky says roughly half of hyperscaler free cash flow is being diverted to data-center capex. Private credit industry size: $1.7 trillion - Used to show how large and systemically relevant private credit has become. First-half GDP growth contribution from data centers: on the order of 50% - Kedrosky says data centers made up roughly half of U.S. GDP growth in the first quarter and similarly in the second quarter. GPU useful lifespan under heavy training use: about 18 months to 2 years - Estimated lifespan when chips are run flat out for model training. Traditional cloud storage hardware lifespan: 6 to 8 years - Longer lifespan for storage-oriented data-center assets like S3-type use cases. Data-center loan tenor vs collateral life: 30-year loans vs ~2-year depreciation - Highlights the temporal mismatch between financing and hardware economics. Commercial real estate cap rates for large hyperscaler data centers: 4.8% to 5.3% - Kedrosky says these yields are unattractive relative to risk. AI revenue projection cited: $70 billion in revenue by 2028 - Referenced in relation to Anthropic forecasts. Anthropic API revenue concentration: 35% from current revenues tied to APIs; 35% of that from software developers - Illustrates customer concentration risk, especially in tools like Copilot and Cursor. Potential consumer subscription requirement: $50 per iPhone user - A back-of-the-envelope example for how large consumer monetization would need to be. Global labor TAM example: $5 billion - Used sarcastically in the discussion of top-down TAM-based justifications. Potential capex needed to justify returns: $3 to $4 trillion - Kedrosky says many people are projecting AI infrastructure needs around this range. Natural gas turbine delivery lead time: to 2030 - Used to illustrate supply constraints in power infrastructure for data centers.

Pivotal Quotes: "This is the first bubble that has all of that." — Paul Kedrosky: He is explaining that AI/data-center financing combines real estate, tech hype, loose credit, and a government-backstop mentality. "It's not really our debt, it's in an SPV, I don't have to roll it back onto my own balance sheet" — Paul Kedrosky: He is describing how sponsors structure data-center financing to keep liabilities off balance sheet. "Compute as a hoardable commodity" — Paul Kedrosky: He uses this phrase to explain why hyperscalers strike deals for capacity they may not immediately need.

Implications: Listeners should expect more scrutiny of AI capex, financing structures, and power constraints. If demand or pricing disappoints, the fallout could affect tech equities, private credit, REITs, and the broader economy.