'Big Short' Investor Explains How the AI Bubble Will Burst

Finance-focused summary (AI bubble risk, concentration, and “picks & shovels” investing)

Macro / market setup

• Core thesis: The AI buildout chain is highly concentrated. If OpenAI and/or Anthropic stumble, the risk is not limited to tech-equity drawdowns—Eisman argues it could trigger a broad recession-like correction (framed as the U.S. going “into a recession overnight” if OpenAI fails).
• Key “tipping” idea: Concentration plus lack of durable pricing power (“no moats” at the LLM level) increases fragility.

---

Concentration risk across the AI value chain (key disclosures + linkages)

1) Nvidia (hardware)

• Eisman owns Nvidia (NVDA).
• Revenue growth: “over 100%” (approx. cited as ~110%) in a recent reporting period.
• Balance-sheet / customer concentration warning: A cited filing note indicates:
  • 70% of Nvidia’s accounts receivable (as of end of July) comes from five customers.
  • Framed as a “warning flag”, not “the end of the world.”

2) Hyperscalers (demand concentration)

• Hyperscalers named: Google, Meta, Amazon, Microsoft, plus Oracle.
• Claim: 70% of hyperscalers’ AI revenue is attributed to Anthropic + OpenAI.
• That AI revenue share is said to equal ~25–30% of total cloud revenue (excluding Oracle in the comparison).

3) Oracle backlog / RPO link to OpenAI

• Oracle uses RPO (backlog) as a key metric.
• RPO jump: from ~$150B to ~$400B in ~3 months (discussed around an August quarter earnings period).
• Stock reaction:
  • Oracle stock reportedly moved from ~230 to ~330 in 2 days after the RPO news.
  • Later, it fell back: around ~200, and “today” at ~150 (as described).
• Additional sell-side claim (paraphrased): ~50% of Oracle’s RPO is attributed to OpenAI.
• Eisman’s interpretation: Oracle’s future revenue is tied (via backlog) to an unprofitable LLM provider (“losing money like crazy”), implying high dependency on funding and growth.

---

Financial performance / unit economics (OpenAI vs Anthropic)

• Revenue (as cited via WSJ-reported assumptions):
  • OpenAI: ~$6.5B revenue in a quarter.
  • Anthropic: ~$11B+ revenue in the quarter.
• Loss / costs (as described):
  • OpenAI costs cited as roughly ~$12B (loss).
  • Eisman frames it as “upside down” economics: revenue rises, but costs rise faster.
• Growth rates cited:
  • OpenAI revenue growth: ~18% in 3 months
  • Anthropic revenue growth: >100% in 3 months
• Conclusion / uncertainty: Eisman views OpenAI as the weaker entity at that point in time, but emphasizes limited data (“not enough numbers”).

Bottom-line risk: The entire stack—hardware → hyperscaler → cloud → model pipeline—depends on LLM providers staying well-funded and scaling.

---

“Where does the money come from?” (financing circularity)

• Eisman rejects the idea that funding comes primarily from venture capital.
• Instead, he suggests much of the capital comes from Amazon, Google, Microsoft, Nvidia, and SoftBank.
• He describes a circular flow (commitments → capital → commitments), compared in spirit to structured-finance circularity (e.g., “Big Short”/CDO layering), but with the condition that LLMs keep growing rapidly.
• Failure mode: If OpenAI/Anthropic “messes up” or investors stop funding, the chain breaks.

---

Business risk: “token maxing” ends + no durable moat in LLMs

• Eisman argues there are no “moats” at the LLM model layer—providers rotate and developers can switch.
• Mechanism described:
  • Token maxing: early underpricing, then price increases once token costs threatened margins.
  • After price increases, customers become more cost-conscious, reducing usage budgets.
  • Examples:
    • One company reportedly burned through its entire AI budget in 3–4 months.
    • Another reportedly spent ~$500M before realizing it had spent that amount.
• Proposed “rotation” outcome:
  • Enterprises use top models only for “super important tasks,” and rely on cheaper openweight models for the rest.
• Speculation / caution: Eisman suggests the industry may be “manufacturing hysteria” to drive regulation that creates a legal barrier favoring a duopoly (potentially excluding certain Chinese models). He calls this his “conspiracy theory.”

---

Hyperscaler fundamentals risk: capex boom + cash flow loss

• Eisman argues hyperscalers used to be highly profitable and cash-generative, but:
  • AI data center capex is driving cash flow deterioration, sometimes negative cash flow.
• A key figure cited: hyperscalers projected to spend about $700B on AI capex “this year.”
• Example cited: an Oracle downgrade with S&P ~BBB- (“one level above junk”).
• He suggests investors are growing nervous across the hyperscaler complex.

---

Depreciation / chip obsolescence debate

• Michael Barr (referred to as “Michael Bur/Bur” in the transcript) raises a concern:
  • Data centers may take too long to ramp online.
  • Chips could become obsolete faster than standard depreciation schedules.
• Eisman’s rebuttal:
  • If AI demand succeeds, depreciation schedule changes (e.g., 3–4 years to 5–6 years) matter less than the macro/cycle reversal.
  • If AI demand fails, the correction is driven by demand collapse, not accounting depreciation.

---

Power / electricity bottleneck (real assets beneficiary call)

• Macro constraint: power availability.
• Eisman cites Elon Musk’s framing: China chip problem vs U.S. power problem.
• Eisman argues power is physical infrastructure and may be a bigger bottleneck than chips.
• Equity example: GE Vernova (GEV)
  • GE’s restructuring into:
    • GE Healthcare (symbol referenced as “GE”)
    • GE Vernova (GEV) = energy grid/electrical + gas turbines
    • GE = aerospace
• Key claims/numbers (as cited/described):
  • US electricity growth: ~3–4% per year
  • Prior analyst view: Vernova “negative value,” including a cited “-$3B” sum-of-the-parts view
  • Backlog said to be stretching past 2030
  • Orders/upturn reportedly up about ~85% in the last quarter (per transcript)
• Eisman emphasizes turbine orders and gas-turbine-driven generation/electrical equipment as core.
• Mentions other generation approaches: nuclear (general) and Bloom Energy (named).
• Elon anecdote: for Colossus (Memphis), Elon used 35 portable gas turbines due to slow grid readiness.

---

Investing guidance: “picks & shovels” vs high-risk AI layer

Eisman’s explicit positioning (as stated)

• High-risk: “I would not invest in an LLM” (no moat).
• Caution: be wary of hyperscalers “at this point” due to capex-driven cash flow loss.
• Potential beneficiaries (“picks & shovels”):
  • Nvidia (NVDA)
  • Micron
  • GE Vernova (GEV)
  • Arista Networks
  • Cisco
  • Eaton (electrification)

Software risk example (agentic AI)

• Agentic AI could reduce the value of traditional intermediaries.
  • Example: if an AI agent books flights/hotels directly via travel sites, travel platforms may lose traffic.
• He flags the payment world as potentially “dicey,” while enterprise-embedded software may be more resilient.

---

Rates / sovereign risk / performance sensitivities

• US debt numbers cited:
  • ~$40T debt
  • Average interest rate: 1.77% (2020) → 3.45% (today)
  • Interest expense: $523B/yr → $1.22T
• Risk framing:
  • Eisman doubts a “debt spiral,” citing the reserve currency nature of the U.S. and the role of Treasuries in global finance (repos).
  • Still prefers smaller deficits.
• “Showstopper” discussion:
  • Focus on 10-year yield (not fed funds).
  • He suggests ~5% on the 10-year is a rough threshold (“Rubicon”).
  • A figure mentioned: 10-year ~4.98%, described as “over 5% yesterday,” and markets rallied as it fell.
• Political/manipulation angle:
  • Criticism of a Treasuries buyback plan (referenced via Scott Bessant in the transcript).
  • Characterized as an attempt to bring rates down but “failed miserably,” with $6B small versus a $40T deficit.

---

Precious metals / Bitcoin view (risk framing)

• Precious metals:
  • Not opposed, but he has never owned it; it doesn’t generate income (“sits there”).
• Bitcoin:
  • Calls it worse than gold-like non-yield assets because it trades inversely to the stated thesis:
    • When inflation worries rise and rates rise, Bitcoin doesn’t behave like a hedge.
  • Notes contradictory day-to-day correlations with Nasdaq and rates.

---

Framework / methodology mentioned (portfolio/company analysis)

• Eisman emphasizes a story-driven but rational approach:
  • “Story” matters, but he dislikes overpaying.
  • He states he is not a pure value investor.
• Red flags / filters:
  • Management selling stock
  • Preference against cyclical companies (predicting the economy), with exceptions
• Moat preference (not absolute):
  • He likes moats (example: Moody’s, Visa) but says it’s not a complete requirement.

---

Key explicit recommendations / cautions

• Do not invest in LLMs (no moats; high risk from competition and switching).
• Caution hyperscalers due to capex-driven cash flow deterioration and dependence on LLM providers.
• Prefer “picks & shovels” tied to durable infrastructure demand (chips, networking, power/electrification).
• Watch 10-year yield for a potential ~5% “Rubicon” threshold.
• Watch for customer/backlog concentration (e.g., Nvidia receivables; Oracle RPO exposure to OpenAI).

---

Disclosures / disclaimers

• No explicit “not financial advice” language appears in the provided subtitles/text.

---

Presenters / sources mentioned

• Steve Eisman (guest; “Big Short” investor; “Real Eisman Playbook” host/creator)
• Michael Lewis (The Big Short author; mentioned as the book writer)
• Elon Musk (referenced via interviews/quotes)
• The Wall Street Journal (reported revenue figures; mentioned as a source)
• CNBC (mentioned generally)
• Interviewer/host: referred to as “host of the Real Eisman playbook podcast” / “Mr. Anders” (full name not clearly provided)
• Michael Barr (referred to as “Michael Bur/Bur” in transcript)
• Scott Bessant (mentioned regarding a Treasuries buyback plan)