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

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

by Alex Kantrowitz

1h 1m•November 14, 2025
Listen to Episode

Overview of Big Technology Podcast (Friday edition)

This episode is an AI bubble special report hosted by Alex Kantrowitz with guest Gil Luria, Head of Technology Research at D.A. Davidson. They analyze whether current AI investment behavior is sustainable or represents a bubble — focusing on debt-financed data-center builds, accelerated depreciation of GPUs, losses at model companies (OpenAI, Anthropic), inference-pricing dynamics, and infrastructural constraints like power. Luria separates “healthy” hyperscalers from “unhealthy” debt-fueled players and explains the financial mechanics and systemic risks behind recent market turbulence.

Key takeaways

  • Both can be true: AI is transformational and in high demand, yet some market behavior around financing and buildouts is unhealthy and bubble-like.
  • Debt used to finance speculative data centers is a core worry — loans should be backed by predictable cash flows and durable collateral; many AI-related projects lack that.
  • Depreciation of GPUs/chips is being underestimated. If useful life is shorter (e.g., 3 years vs. 5–6 years), reported profits and valuations could be materially overstated.
  • A “prisoner’s dilemma” in inference pricing (race to the bottom / flat unlimited plans) could keep inference services unprofitable for a prolonged period.
  • Hyperscalers (Microsoft, Amazon, Google) can absorb risk and pause CapEx; smaller players (CoreWeave, some SPVs) and lenders are most exposed.
  • Power and other supply-chain bottlenecks are constraints but market incentives (storage, on-site generation, tradespeople) will drive solutions.

Topics discussed

  • What constitutes a “bubble” vs. legitimate growth in AI
  • Examples of healthy vs. unhealthy corporate behavior in AI buildouts
    • Healthy: Microsoft, Amazon, Google (cash-backed, diversified revenue)
    • Unhealthy: CoreWeave (heavy borrowing), Oracle (debt tied to speculative customer revenue), Meta (leveraging SPVs and large CapEx)
  • How debt should be used (Finance 101: debt for predictable cash flows; equity for speculative growth)
  • Specific lenders and structures: large banks (U.S. Bank, J.P. Morgan, Mitsubishi) lending to CoreWeave; Meta’s special purpose vehicle arrangements with firms like Blue Owl
  • Michael Burry’s critique: underestimating GPU depreciation is akin to accounting sleight-of-hand; estimated understatement of depreciation ~$176B (2026–2028)
  • Differences in chip usefulness vs. chip working — functioning hardware may not deliver equivalent revenue as newer generations dramatically outperform
  • Inference pricing dynamics and the risk of subsidizing “power users”
  • Power/grid constraints and mitigation strategies (behind-the-meter generation, storage, manual labor demand)

Notable data points & quotes

  • OpenAI allegedly committed revenue promises totaling roughly $1.4 trillion across multiple partners (e.g., $300B to Oracle, $200B to Microsoft; transcript cites aggregate commitments).
  • Gil Luria: “Both things are true” — AI is revolutionary, and simultaneously there is unhealthy financial behavior.
  • Michael Burry (tweet cited): “Massively ramping CapEx through purchase of NVIDIA chips and servers on a two- to three-year product cycle should not result in the extension of useful lives of compute equipment…By my estimates, they will understate depreciation by $176 billion from 2026 to 2028.”
  • Luria’s accounting point: “Does it work? vs. Can it generate revenue?” — old chips may work but not generate comparable revenue as newer chips.

Companies & examples (what they represent)

  • Microsoft: Large, diversified, cash-rich; can absorb CapEx and monetize AI via Azure + enterprise software; but materially exposed to OpenAI-related margins/metrics.
  • OpenAI: Rapid growth and large commitments; currently a negative gross-margin business for many queries; risks from over-committing to infrastructure and revenue claims.
  • CoreWeave: Cited as poster child for risky debt-funded expansion (borrowing to build data centers without proven long-term customers).
  • Oracle: Taking on debt tied to speculative revenue from OpenAI (example of lending-for-speculation risk).
  • Meta: Big spender and “maximalist” — using SPVs and signaling willingness to outspend rivals (investor concern).
  • Nvidia: Rapid chip-generation cadence (H100 → Blackwell → next gens) is central to depreciation debate.
  • Anthropic: Example of loss-making model companies supported by hyperscalers; Amazon recorded a big gain from its stake while Anthropic burns cash.

Main risks identified

  • Debt risk: Loans on speculative future AI revenue or on assets whose value could collapse quickly if chip economics shift.
  • Depreciation risk: Accounting life assumptions (5–6 years) may be unrealistic given rapid chip improvements; shortening life can materially reduce EBITDA and valuations.
  • Pricing dynamics (inference): Flat/unlimited plans to chase market share could depress prices and margins broadly.
  • Systemic contagion: Tens of billions in losses are manageable; hundreds of billions tied up in poorly collateralized loans could create contagion similar to 2008-style credit issues.
  • Power and infrastructure: Grid capacity and power provisioning are real constraints, though market responses (storage, behind-the-meter generation) will mitigate over time.

Actionable checklist / recommendations

For investors:

  • Quantify a company’s debt exposure tied to AI CapEx and special-purpose vehicles.
  • Examine depreciation policies and sensitivity: run scenarios with shorter useful lives (3 vs. 5–6 years).
  • Measure revenue dependency on a single model/customer (e.g., OpenAI) and stress-test those revenues.
  • Track lenders’ exposure (which banks have large loans to data-center builders).

For corporate decision-makers:

  • Prefer equity or internal cash funding for speculative AI infrastructure versus leverage tied to uncertain customers.
  • Use conservative depreciation assumptions aligned with observed chip performance cycles.
  • Focus on differentiated, monetizable products (e.g., Luria suggests OpenAI focus on ChatGPT/core model) instead of overextending into full-stack hardware ownership unless warranted.

For policy-makers / risk officers:

  • Monitor systemic credit exposure in AI/compute finance; require transparent reporting of SPVs and off-balance-sheet obligations.
  • Consider encouraging standards for asset life assumptions and disclosure around GPU depreciation.

Implications & conclusion

  • The AI opportunity is real, but the market currently mixes responsible, cash-backed expansion with speculative, debt-fueled growth. That mix creates meaningful downside risk for lenders and leveraged players.
  • Hyperscalers can likely weather a correction; smaller players and their lenders are most vulnerable. Widespread de-leveraging or repricing of depreciation would compress reported profits and could re-rate valuations.
  • Two main equilibrating forces: market discipline (lenders tighten standards) and consolidation (big players buy distressed assets). Either will reshape the space.
  • Short-term turbulence doesn’t negate long-term AI demand — but investors and decision-makers should distinguish healthy investment behavior from the kinds of financing and accounting practices that historically precipitate broader financial stress.