Inside The AI Bubble: Debt, Depreciation, and Losses — With Gil Luria

Episode Summary

Executive Summary: The episode is a deep dive into whether AI is in a bubble, with Gil Luria arguing both that AI demand is real and revolutionary, and that parts of the financing behavior are unhealthy. The central risks he identifies are excessive debt, aggressive depreciation assumptions, speculative revenue promises from OpenAI, and the possibility of systemic spillover if hundreds of billions are tied to assets that lose value quickly.

Main Topics: AI demand is real, but bubble behavior exists (Priority: 5/5): Luria argues the market can simultaneously have genuine AI adoption and unhealthy financial speculation. He sees Microsoft, Amazon, and Google as mostly prudent, while warning that other actors are stretching the rules. Debt financing for data centers is the main risk (Priority: 5/5): He says debt is appropriate for predictable cash flows and durable assets, but dangerous when used to fund speculative AI infrastructure backed by uncertain revenue from startups like OpenAI. Oracle, CoreWeave, and leverage on OpenAI (Priority: 5/5): Oracle and CoreWeave are presented as examples of risky behavior, especially when borrowing against promised future revenue that may not materialize, creating exposure to a fragile customer base. Depreciation and AI chip obsolescence (Priority: 5/5): Michael Burry’s critique is largely endorsed: if GPUs become obsolete in 2–3 years rather than 5–6, companies may be overstating profits by under-depreciating equipment, which would hit valuations. OpenAI’s losses and overcommitment (Priority: 4/5): OpenAI is praised for creating ChatGPT but criticized for making huge commitments it may not be able to honor. Luria says it should focus on ChatGPT and frontier models instead of building its own full-stack infrastructure. Game theory, market share, and persistent losses (Priority: 4/5): The large tech firms are acting as if AI is winner-take-all, which drives prolonged spending, flat-rate pricing pressure, and a race to the bottom in inference economics. Power and infrastructure bottlenecks (Priority: 3/5): Electricity and grid capacity may slow deployment, but Luria says the market will work through bottlenecks via storage, behind-the-meter generation, and specialized labor.

Key Arguments: AI is both a transformative technology and a source of bubble-like behavior; the two can coexist. Debt is appropriate for mortgages and predictable assets, but speculative AI infrastructure should generally be funded with equity or cash flow, not leverage. OpenAI’s promised revenue streams are too uncertain to justify hundreds of billions in debt across partners and suppliers. If AI chip useful life is closer to 3 years, then depreciation is understated and reported profits are overstated. Big tech firms with huge cash flows can absorb AI losses and stop capex quickly; smaller, leveraged firms cannot. The biggest danger is not isolated losses, but a collapse in asset values across a large stack of debt-financed AI investments. Markets may eventually correct by forcing irresponsible lenders to tighten standards, leaving better-capitalized firms to buy distressed assets cheaply. AI pricing may stay under pressure because firms are prioritizing market share over margins, especially in inference. Power constraints are real, but economic incentives will likely drive workarounds and new supply. OpenAI has been highly successful at product creation, but its current ambitions may exceed what it can sustainably finance.

Data Points: ChatGPT weekly active users: 800 million - Used to illustrate OpenAI’s consumer traction and rapid adoption OpenAI forecasted revenue commitments: $1.4 trillion total commitments - Luria cites OpenAI’s promised obligations to Microsoft, Amazon, CoreWeave, and others OpenAI annual revenue (current year estimate): About $15 billion - Compared against its enormous commitments and losses OpenAI annual losses (current year estimate): More than $20 billion - Used to argue the company cannot support its commitments through current cash flow Lisa Su AI compute market projection: $1 trillion - Mentioned as a bullish forecast supporting the demand-side argument CoreWeave loan rate example: 9% - Luria describes lenders being attracted by high interest rates on speculative lending GPU rental price example: $4 per hour - Illustrates current economics of renting AI compute capacity GPU rental decline scenario: $0.40 per hour in three years - Example used to show how chip revenue can collapse as newer generations arrive Burry depreciation estimate: $176 billion understatement from 2026 to 2028 - Michael Burry’s estimate for under-depreciation across hyperscalers Burry overstatement example: Oracle 26.9%, Meta 20.8% - Burry’s claim that profits could be overstated if asset lives are too long OpenAI commitments to specific companies: Microsoft $200B, Amazon $38B, CoreWeave $25B - Examples of the scale of promised future spending OpenAI customer concentration/overcommitment: $300B to Oracle (mentioned as a promise) - Used to highlight speculative revenue assumptions AI revenue growth contribution: 65% - WSJ cited figure for OpenAI losses equating to 65% of the rise in underlying earnings of Microsoft, NVIDIA, Alphabet, Amazon, and Meta together Meta ad revenue growth: 25% - Used to show Meta is highly profitable even while increasing AI spending AI adoption pricing example: $20/month to potentially $200/month or $10,000–$20,000/year - Illustrates how AI products could become more profitable as users realize their value

Pivotal Quotes: "Both things are true." — Gil Luria: His core framing: AI is genuinely transformative, while some financing behavior is still unhealthy "If you're borrowing money to make a speculative investment based on a speculative customer, that's bad behavior." — Gil Luria: Explains why debt-financed AI buildouts tied to OpenAI commitments worry him "The fact that the chip works after three years doesn't mean it's going to generate the same revenue." — Gil Luria: Responding to counterarguments about GPU warranties and hardware lifespan

Implications: The AI boom may continue, but the market is likely to punish leveraged or overcommitted players first. Expect tighter underwriting, more scrutiny of depreciation, and a split between cash-rich hyperscalers and fragile AI infrastructure bets.