Is the AI Bubble About to Be Tested?

Finance-focused summary of the subtitles (AI IPO “bubble” risk)

Key market/timing context

• NASDAQ: trading at a record high during the period discussed.
• US growth: a purchasing managers survey indicated US business output growing fastest in 5 years.
• IPOs getting pulled anyway:
  • Jay Ritter (University of Florida IPO initiative) says it’s especially surprising given NASDAQ’s strength.
  • Anthropic filing: expected late August but had not appeared at the time of narration.
  • OpenAI: listing pushed to next year.
  • WholeTech (nuclear power): postponed IPO, citing market conditions.
  • SB Energy (data center developer, SoftBank): filed but not started marketing.
• Since Labor Day:
  • 3 IPOs occurred.
  • 5 of the 10 largest listings were trading below offer price.

Interest rates / macro risk driving valuation pressure

• 10-year Treasury yield: 5.23% (highest since 2004).
• Fed: raised rates earlier in the month.
• Rising rates create a “double whammy”: 1) Discount rates lower the present value of future profits. 2) Financing to build data centers becomes more expensive.
• Historical parallel noted: last time the Fed raised rates into a similar boom was 1998.

Company valuation discussion: Anthropic ~$2T vs “reasonable pricing”

• Anthropic was expected to go public soon and (reportedly) seek about a $2 trillion valuation.
• The narrator argues this relies on optimistic assumptions and “new math,” implying the valuation is more a bet on growth than near-term cash flows.
• Implied/mentioned sizing:
  • Anthropic revenue: “around $65B/year” (August running rate).
  • At $2T, that implies about ~31x revenue.

Valuation mechanics explained (framework)

Bankers commonly combine:

• Discounted Cash Flow (DCF):
  • Forecast future cash flows
  • Discount them back to today using an interest rate
• Comparable multiples:
  • Compare similar public companies’ sales/profits/metrics
  • Apply relevant industry multiples
• Then price via a road show:
  • Solicit bids from large investors
  • Set price based on order book demand

“McNeely test” / price vs fundamental economics

• An analogy attributed to Scott McNeely (Sun Microsystems co-founder) from the dot-com era: paying 10x revenue was “insane” because the assumptions for payback ignored:
  • costs, taxes, R&D, and the time value of money.
• The critique emphasizes time value of money:
  • At dot-com peak, the 10-year rate ~6.5% (mentioned).
  • Under those conditions, even “costless” revenue would take ~17 years to pay back 10x revenue—not 10.
• Applied to Anthropic:
  • Even with absurdly generous assumptions (e.g., no costs/taxes/staff/electricity; perpetual payout), $2T still doesn’t “add up” without extreme growth.
  • Conclusion: almost all of $2T is effectively a bet on growth, and higher interest rates make that growth thesis harder to justify.

“TAM inflation” as the argument for higher valuations

• TAM (Total Addressable Market) is introduced as an alternate justification:
  • TAM = the total revenue pool if a company captures the entire industry.
• Examples cited of TAM expansion:
  • May: SpaceX reportedly told investors the enterprise apps (including AI products) market is ~$22.7T.
  • Wall Street Journal: Anthropics filing may cite $30T market.
  • Morgan Stanley estimate: generative AI could be ~$60T, framed as about half of everything produced on the planet in a year.
• Counterpoint:
  • Valuations assume companies capture large TAM shares.
  • History suggests this often doesn’t happen (e.g., Uber TAM ~ $122.3T vs revenue < $60B, plus a large “remaining market” concept).

Recursive self-improvement narrative challenged

• Mentioned research: DeepMind AlphaEvolve found improvements to matrix multiplication.
• Critique:
  • Gains relied on benchmarks with known answers and were “gamed” in part.
  • Production-scale improvement was described as about half a point (within noise).
• Takeaway: lab “results” may not translate cleanly into real-world economic breakthroughs.

---

Data center infrastructure bubble: SB Energy / SoftBank / financing risk

SB Energy valuation and leverage/rate shock

• SB Energy (SoftBank data center developer) seeks an IPO valuation of about $50B.
• Major caution: it had not switched on a single data center.
• The narration cites:
  • Valued at about 400x annual EBITDA (per Lex).
  • Needs $174B to build promised projects (per prospectus).
  • New debt price yield: 9.75%.
  • Described as the largest junk bond offering on record.
  • Lenders restrict dividends until 2029.
• Result: IPO was pushed back pending better investor sentiment.

Circular funding chain (counterparty/guarantee entanglement)

SoftBank borrows at junk-ish rates to fund OpenAI-related needs:

• SoftBank marketing >$11B of new bonds at about ~9.75% (FT).
• OpenAI cash burn: expected to burn about ~$280B by end of 2030 (FT).
• Lease model:
  • OpenAI leases SB Energy’s Ohio campus for 20 years, supporting SB Energy revenue.
• Warrants/guarantees:
  • OpenAI holds warrants paying if SB Energy hits an $80B valuation.
• Nvidia:
  • Bought $1.5B of SB Energy shares in August at a 10% discount to IPO price; planned to buy another $1.5B at IPO.
  • Agreed to guarantee up to $105B for the Ohio campus.
  • Nvidia won’t record a liability until leases begin in 2028.
  • Campus operates exclusively on Nvidia hardware for 20 years.
• Critique: the sector is portrayed as mutually supporting valuations through guarantees, rent arrangements, and private marks—creating a “can’t bet against it” dynamic.

---

Nvidia as “shovel seller”: why it’s treated as cheap despite profits

Reported performance and valuation

• Nvidia is described as the “company making all the money from AI.”
• Sales:
  • About $27B four years earlier → estimated ~$410B this year (Bloomberg).
• Net income:
  • Expected to almost double (Bloomberg).
• Valuation:
  • Trading at <17x expected next-year profits.
  • Yet framed as near its cheapest in more than a decade.
  • Bloomberg claim: “cheapest” relative to expected profits over a decade.

4 explanations offered for Nvidia’s discount (investment logic)

1) Cyclical at peak - Gross margin ~75% last quarter; expected below 72% by year-end. - Memory chip cost pressure; Micron up >180% this year. - Customers (e.g., Meta, Alphabet) building chips in-house. 2) Spending dependency - Nvidia’s multiple assumes AI spend continues. - If labs slow due to regulation or sentiment, Nvidia suffers. 3) Uncertainty premium - Early-stage uncertainty can sometimes increase valuations, but it can reverse once outcomes become clearer. 4) Different pricing mechanisms / constraints - Nvidia is repriced continuously by public market liquidity (including short sellers). - Anthropic priced privately with fewer downside arbitrage opportunities; “pessimists can’t bet against it,” allowing optimists to set price.

---

Price war / cost curve compression (model commoditization)

Cheaper models arriving around IPO windows

• Newer releases:
  • Anthropic and OpenAI released cheaper models ~40% and ~50% cheaper than prior ones (same day).
• Epoch AI study:
  • Cost to achieve given performance fell ~13x per year since 2023.
  • Prices tend to fall fastest right after new top models ship.
• Dynamic framed as input costs up / output prices down:
  • Chips, power, and labor more expensive
  • Model inference pricing decreasing

Examples of “cheaper than frontier” models

• Typesafe AI model Jev:
  • Claimed up to 440x cheaper than frontier models for some decision tasks.
  • Seed funding: $40M (noted as small relative to $2T scale).
• OpenRouter usage:
  • Weekly usage grew ~25,000% since the start of the prior year.

Usage retention / end-customer economics (signals)

• Yale economics professor Alle Tinsky:
  • 22.5% of Anthropic users still using after a year vs ~13.2% for OpenAI.
• Analyst claim:
  • If labs build tools/software, they may lock in customers and sustain profits.
  • However, routing tools like RAMP suggest customers may multi-home.

---

“Bubble test” mechanics: IPO as the first true market repricing

SoftBank as the risk center

• SoftBank stake in OpenAI last privately valued at $852B.
• SoftBank took a $10B margin loan secured against OpenAI shares.
• An IPO could force real market pricing:
  • If OpenAI’s value declines, collateral shrinks.

Anthropic and OpenAI timing control

• Anthropic IPO described as the “first public test” of a “pure AI lab.”
• OpenAI timing:
  • Sam Altman (Fortune quote): given safety concerns, going public now would be “ill-advised.”
  • OpenAI reportedly in private fundraising at $1.2T valuation—described as “professionals only.”
• Narrator’s implied caution:
  • Pricing that doesn’t “add up” invites later reevaluation when constrained by public markets.

Market-wide IPO supply caution

• Jim Chanos expects record stock issuance.
• Academic research cited:
  • New issuers often underperform for years afterward.
  • Heavy issuance tends to coincide with broader market weakness in subsequent years.

---

Explicit disclosures / cautions

• Early disclaimer: the narrator says they’re not claiming AI is useless.
• The issue is portrayed as price:
  • “A great technology can still be a terrible investment if you pay too much.”
• No explicit “not financial advice” line appears in the subtitles.
• The valuation critique is framed as “show me the numbers.”

---

Tickers / entities / instruments mentioned

Indices / rates

• NASDAQ
• US 10-year Treasury yield (5.23%)

Stocks / companies

• Nvidia
• Micron
• Meta
• Alphabet (Google)
• SoftBank
• OpenAI
• Anthropic
• SpaceX
• WholeTech
• SB Energy
• Uber
• DeepMind (Google DeepMind)

Debt / credit

• Junk bond yields (9.75%)
• $10B margin loan secured against OpenAI shares
• Mentions of “unrealized gains” by companies holding stakes (examples referenced: Amazon and Google)

Other

• RAMP
• Typesafe AI
• Man Group
• Oracle (via historical framing of Sun)

---

Methodologies / valuation frameworks referenced (structural)

• DCF: forecast cash flows → discount to present using interest rates.
• Relative multiples: compare peers’ valuation multiples → apply industry multiple.
• Road show / bookbuilding: gather investor bids → set IPO price using order demand.
• “McNeely test”:
  • Use simplified payback assumptions
  • Apply time value of money to test whether payback implied by “10x revenue” is plausible.
• TAM-based logic:
  • Estimate total potential market revenue
  • Assume capture of market share to justify equity value
  • Critique: TAM assumptions can inflate faster than real firm performance.

---

Key presenters / sources mentioned

• Anthropic CEO (unnamed)
• Sam Altman (OpenAI CEO; via quote)
• Jay Ritter (University of Florida)
• Matt Kennedy (Renaissance Capital)
• Scott McNeely (Sun Microsystems co-founder)
• Jensen Wong (Nvidia CEO)
• Edward Miller (academic referenced in “pessimists vs optimists” framing)
• Lubos Pastor and Pietro Veresi (uncertainty/valuation argument referenced)
• Barberis and Wang (lottery-ticket/overpay argument referenced)
• Jim Chanos
• Loan and Ritter (issuance underperformance research)
• Baker and Wurgler (broad underperformance when issuance is high)
• Alle Tinsky (Yale economics professor)
• Ben Thompson (customer lock-in argument referenced)
• Eric Glimman (RAMP co-CEO)
• Ed Ellson (ProfG Markets podcast host)
• Jorge “Lex” / The Economist Lex
• Financial Times (FT)
• Bloomberg
• Wall Street Journal
• Epoch AI
• DeepMind (AlphaEvolve mention)
• House of L AI (research review channel referenced)
• Brent Flyberg (research cited)
• Paul Kadroski (Man Group research piece)