//
Podcast · 2026-09-14

The Real Eisman Playbook: The Big Short Partners Reunite: Rates, AI, Gold and Two Stock Picks | The Real Eisman Playbook Ep 75

Eisman, Daniel & Collins on fiscal strain, gold, OpenAI’s economics and live shorts.

Episode

The Real Eisman Playbook
ShowThe Real Eisman Playbook
EpisodeThe Big Short Partners Reunite: Rates, AI, Gold and Two Stock Picks | The Real Eisman Playbook Ep 75
GuestVincent Daniel, Porter Collins
HostSteve Eisman
Published2026-09-14T16:06:00Z
Duration52 min
Fidelity[partial]
StanceNEUTRAL
ListenEpisode link

Abstract

Steve Eisman reunites with former Big Short partners Vincent Daniel and Porter Collins, who now run a long/short book anchored in gold. The conversation covers the fiscal bind facing the Treasury — deficits near 6–7% of GDP, interest plus entitlements around a tenth of tax receipts — and why the partners hold their highest-conviction precious-metals position in years. On AI, they walk through OpenAI’s deteriorating economics versus Anthropic, hyperscaler revenue concentration, enterprise cost deflation, and why forced lab IPOs could flood equity supply. The back half explains why shorting is structurally harder now, then works through two live shorts and two idiosyncratic longs. [partial: YouTube auto-captions; ASR artifacts possible]

The Theses

8 claims
1. OpenAI’s economics are deteriorating while Anthropic accelerates away from it.
Evidence: OpenAI’s June-quarter revenue was $6.5B, up only 18% in three months against $12B of costs — revenue up $1B while costs rose $3B — while Anthropic posted $11.5B, up over 100% in three months with costs undisclosed.
2. Hyperscaler AI revenue is dangerously concentrated in two cash-burning customers.
Evidence: Wall Street reports put ~70% of hyperscaler AI revenue at Anthropic plus OpenAI — 25–35% of total cloud revenue — and Nvidia’s 10-Q Note 7 shows its top five direct customers at 70% of accounts receivable.
3. The AI buildout has crossed from internally funded to externally financed, and forced lab IPOs will flood equity supply.
Evidence: Hyperscalers moved from buybacks to debt and equity issuance — Google sold $85B of equity — while Collins argues new stock supply is what killed markets in 1929 and 2000 and that the labs must come public because they need the capital.
4. Enterprises are deflating AI unit economics by reserving frontier models for frontier queries.
Evidence: One public-company CIO (“Mr. X,” not for attribution) runs AI queries and tokens up ~30–40% with costs down ~60% by routing only the most important queries to Anthropic and OpenAI and the rest to Chinese open-weight models.
5. The Treasury is out of fiscal options and reduced to theatrical rate suppression; gold is the highest-conviction expression.
Evidence: Interest plus entitlements run ~10% of tax receipts against 6–7% deficits, and Collins calls Bessent’s “at least $4B” Treasury twist-buy program “theatricality and deception” — “they’re out of arrows” — backing his strongest precious-metals conviction in years.
6. Shorting is structurally harder because levered multi-strategy pods dominate the short book and cover fast.
Evidence: Pods run ~5:1 leverage with dollar-for-dollar, beta- and factor-neutral shorts, are happy with 3–5% spreads versus a fundamental short-seller’s ~150% target, and cover within weeks — squeezing fundamental shorts before the thesis pays.
7. Carvana’s reported profitability rests on an unexplained gain-on-sale premium paid by an undisclosed buyer.
Evidence: Average gain on sale ran 109–110 for two years while big buyers like Ally pay 102–104, implying a mystery buyer north of 110 — with 75–100% of pre-tax income from gain on sale — and Delaware Life’s related-party transactions jumped from ~3% to 30–40% per court filings.
8. FICO’s pricing power invited its own regulatory and customer backlash.
Evidence: The partners shorted FICO at ~$2,500 after Bill Pulte’s (“Bill Py” in captions) comments, arguing FICO raised prices 1,600% over five years — when inflation-plus-2% would have passed unnoticed — and charges 3–5x for mortgage credit pulls versus card and auto.

Key Math

  • ~10% of tax receipts — interest expense plus entitlements Porter Collins
  • ~$40T debt × prevailing rate “is a big number” and rising Vincent Daniel
  • 6–7% of GDP / ~$2T budget deficits — “the root cause of all of this” Vincent Daniel
  • Nvidia revenue up over 100%, but 70% of accounts receivable with top five direct customers (10-Q Note 7) Steve Eisman
  • ~70% of hyperscaler AI revenue from Anthropic + OpenAI, per Wall Street reports Porter Collins
  • ~25–35% of total cloud revenue — the Anthropic/OpenAI share Porter Collins
  • $11.5B — Anthropic June-quarter revenue, up over 100% in three months; costs undisclosed Steve Eisman
  • $6.5B — OpenAI revenue, up only 18% in three months; costs $12B Steve Eisman
  • OpenAI: revenue +$1B vs costs +$3B over three months Steve Eisman
  • GDP to grow 2%, at least half of it AI capex (Torsten Sløk) Steve Eisman
  • ~30–40% rise in AI queries/tokens, ~60% fall in AI costs — one CIO routing off-frontier queries to open-weight models guest
  • $85B equity issuance by Google Porter Collins
  • $25B capital levered 5:1 → $125B positions; annualized vol target 6–9%; longs/shorts dollar-matched Porter Collins
  • Pod shorts target 3–5% spreads vs the partners’ ~150% target Porter Collins
  • FICO shorted at ~$2,500 after Pulte comments; prices raised 1,600% over five years; mortgage pulls 3–5x card/auto pricing guest
  • 109–110 — Carvana’s average gain on sale over two years; Ally pays 102–104; mystery buyer north of 110; 75–100% of pre-tax income from gain on sale; stock ~30–40x earnings guest
  • ~3% → 30–40% — Delaware Life related-party transactions, per court filings guest
  • 4 FLNG boats on ~20-year toll contracts; capex $2.5–3.5B vs YPF’s $30–50B onshore build Porter Collins
  • Argentina rates down from 100%+ to ~30% Porter Collins
  • $90–100/lb production cost; $200/lb in California vs $600–700/lb in Germany; 50–65% EBITDA margins if export allowed; ~$500M–$1B market cap; federal rescheduling (Schedule 1 → 3) finalized last December — Glass House Vincent Daniel

Quotes

I think I've had as much conviction in this precious metals trade as I've had in anything in a long, long time.

Porter Collins  ·  The Real Eisman Playbook  ·  ~00:02

Why it matters: The conviction statement behind the episode’s gold thesis — highest in years.

I make this joke that whoever buys open AI at out of bankruptcy, it's going to be a fantastic deal

Steve Eisman  ·  The Real Eisman Playbook  ·  ~00:21

Why it matters: How far the OpenAI narrative has turned: bankruptcy is now a punchline, not unthinkable.

if all of a sudden open AI failed and the whole thing reversed the economy would go into recession almost immediately.

Steve Eisman  ·  The Real Eisman Playbook  ·  ~00:24

Why it matters: The macro-stakes claim — half of GDP growth is AI capex — stated plainly.

That's that's in 2000 in 1929. That's what killed the market was new supply of stock.

Porter Collins  ·  The Real Eisman Playbook  ·  ~00:18

Why it matters: The supply-demand frame behind the forced-IPO warning on OpenAI and Anthropic.

proving that is going to be impossible. that has proven to be impossible for eight years.

Vincent Daniel  ·  The Real Eisman Playbook  ·  ~00:40

Why it matters: The honest tell on the Carvana short: the thesis is right but the payoff needs a catalyst.

Variant Perception

Priced in: Fiscal unease, gold’s long run, AI-capex concentration worries, and the difficulty of shorting crowded names are all consensus-adjacent; the episode confirms more than it discovers. Eisman himself pushes back on imminent fiscal doom, and Collins concedes the Carvana–Delaware Life link has been known to shorts for eight years.

What’s new: The concentration math is the variant datapoint — Nvidia’s 10-Q Note 7 puts 70% of accounts receivable with five direct customers; ~70% of hyperscaler AI revenue comes from Anthropic plus OpenAI, or 25–35% of total cloud revenue; and the quarterly split (Anthropic $11.5B, up 100%+; OpenAI $6.5B, up 18%, costs up $3B against revenue up $1B). The pod-leverage mechanics (5:1, 3–5% spread thresholds, weeks-long holding) quantify why squeezes keep happening. Carvana’s gain-on-sale arithmetic (109–110 average vs 102–104 from Ally) and the Delaware Life related-party jump (3% to 30–40%) are genuinely new to most readers, as is the Mr. X cost-deflation anecdote.

The bear case: OpenAI’s narrative could re-accelerate with a product or funding event, making the “in trouble” call look premature; gold is a crowded consensus long after a strong run; the Delaware Life/Carvana connection has been litigated for years without a payoff; FICO’s pricing may invite nothing more than noise; and Bessent’s twist program could pin long yields longer than skeptics expect.

Discount: The speakers are talking their book — long gold and idiosyncratic small caps, short FICO and Carvana (where they admit being squeezed). The Nvidia Note 7 point arrives via Ed Zitron (“Ed Zitron” in captions), an avowed AI bear, though Eisman says he verified it. The Mr. X anecdote is single-source, unaudited, and not for attribution. Auto-caption artifacts (names, percentages) mean any single figure should be checked against the primary source before use.

Why It Matters

This is three professional short-sellers arguing that the two risks Graham tracks — fiscal sustainability and the AI buildout’s funding — are converging: a Treasury reduced to signaling games on rates, and an AI economy whose revenue is concentrated in two cash-burning labs while enterprises quietly deflate the unit economics. The episode’s value is in the specifics — the Note 7 concentration, the lab quarterly splits, the gain-on-sale arithmetic — and in the rare mechanics lesson on why fundamental shorts keep failing against levered pod capital. It is a positioning read, not a call: bearish on the fiscal path and on OpenAI’s funding, bullish on gold, idiosyncratic elsewhere.

Positioning Read

Directional only
WEAKENED
AI capex supercycle
OpenAI is “in trouble” — $6.5B quarterly revenue up only 18% against $12B of costs, losing share to Anthropic — while the buildout’s demand story depends on two cash-burning labs; if funding cracks, Sløk’s arithmetic (half of 2% GDP growth is AI capex) makes the supercycle’s macro support fragile.
WEAKENED
Buildout financing is structurally sound
Hyperscalers have moved from buybacks to external financing (Google’s $85B equity sale), ~70% of their AI revenue sits with two unprofitable labs, and the labs must come public — adding equity supply exactly where the circularity critique bites.
Inference economics favor consumption pricing
One public-company CIO cut AI costs ~60% while usage rose ~30–40% by routing only frontier queries to frontier labs and the rest to Chinese open-weight models — a live datapoint on inference spend behaving like an optimizable consumption good, but it does not adjudicate the seat-vs-consumption question.