A finance video (in the style of Patrick Boyle) examining Anthropic’s expected IPO at a reported ~$2 trillion valuation — roughly the combined value of the ten biggest tech IPOs of all time. The central puzzle: on paper the timing looks perfect (record Nasdaq, booming business surveys), yet IPOs are being pulled, interest rates are at two-decade highs, and the numbers don’t obviously support the price. Using Scott McNealy’s famous 2002 takedown of 10x-revenue pricing, the host shows that even under absurdly generous assumptions, $2 trillion can’t be justified by Anthropic’s cash flows — it must be a bet on invented TAM figures or recursive self-improvement. Meanwhile NVIDIA, the only company in the story actually making money, trades at its cheapest level in a decade. The video tours the circular financing of the AI boom (SoftBank, OpenAI, SB Energy, NVIDIA), the collapsing price of AI outputs, and what a public listing would mean when private paper valuations finally meet a real market price.
Anthropic, whose CEO has just asked the industry to slow down and for government regulation, is expected to go public within months hoping to raise more money faster than any company in history, at a reported ~$2 trillion valuation — the Economist notes that’s about the sum of the ten largest tech IPOs ever. The timing looks ideal: record Nasdaq, the fastest US business output growth in five years. But IPOs are being pulled (Jay Ritter calls it especially surprising at a Nasdaq record), Anthropic’s filing has slipped past its expected late-August release, OpenAI has pushed its listing to next year, and rates are the dark cloud: the 10-year yield hit 5.23%, the highest since 2004, on hot-economy, sticky-inflation, big-deficit dynamics. Renaissance Capital’s Matt Kennedy calls rising rates a double whammy for companies like this. Holtec (nuclear) and SB Energy (data centers) postponed; only three IPOs since Labor Day; five of the ten largest 2025 listings trade below offer. The last Fed hiking cycle into a boom was 1999.
The host recalls Scott McNealy’s 2002 Businessweek explanation of why 10x revenue at the 2000 peak was indefensible — 100% of revenues as dividends for ten years assumes zero COGS, zero expenses, zero taxes, zero R&D — and notes McNealy was still too generous: with 10-year Treasuries at 6.5%, the time value of money stretched the payback to ~17 years. Anthropic at $2 trillion would be ~31x its ~$65 billion run-rate revenue (figures treated cautiously). Under maximal generosity — every dollar of revenue forever, no costs, no taxes, discounted at the risk-free rate — the stream is worth ~$1.27 trillion, roughly three-quarters of a trillion short. The $2 trillion is a bet on growth, and higher rates raise the required growth.
Henry Blodgett made his name (and later a lifetime industry ban) with late-90s TAM logic on Amazon. The FT’s Lex column traces the current inflation: SpaceX/xAI’s $22.7 trillion enterprise-AI TAM in May; a possible $30 trillion in Anthropic’s filing per the WSJ; Morgan Stanley’s $60 trillion estimate last week — half of global annual output — from a bank likely to underwrite the IPO. Four months, $37 trillion of growth, outpacing both Anthropic’s revenue and the economy it’s supposedly carved from. Cautionary precedents: Uber’s $12.3 trillion TAM (99.5% still unaddressed), WeWork’s $3 trillion (followed by bankruptcy). Bigger still: Anthropic’s research arm modeling $10 trillion added to US GDP by 2030 (~$100 trillion of equity value today), and beyond that recursive self-improvement, where value grows so fast that “money and asset prices stop meaning anything at all” — the host jokes about Musk predicting money’s irrelevance while collecting so much of it, and about his own $100 trillion Zimbabwean note.
Alpha Evolve’s matrix-multiplication result was real and impressive, but — as the House of AI channel’s review points out — the striking successes all come with scoreboards, clear right answers. Anthropic’s automated researcher closed 97% of a performance gap, partly by gaming the experiment, and its best idea delivered about half the improvement at production scale, within noise. Anthropic published that honestly. Eighteen months ago Anthropic projected $12 billion 2027 revenue — now the pessimistic number in a story of optimism. Damodaran floated a $1 trillion → $800 billion valuation slip on the Prof G Markets podcast; seven weeks later, the talk is $2 trillion.
SoftBank’s SB Energy seeks ~$50 billion without ever having switched on a data center (it bought a consultancy that built 15), at ~400x EBITDA, needing $170 billion of investment for its promises, with record junk bonds at 9.75%, no dividends until 2029, and a postponed IPO. The host’s satirical “Boyle Compute” IPO (valuation: $25 billion, “to attract value investors”; assets: an excellent deck and a Wi-Fi router once successfully plugged in) illustrates the accounting charm: construction in progress doesn’t depreciate, just like chips not yet switched on (per Jim Chanos). But building is where it goes wrong: Flyvbjerg’s iron law of megaprojects, decade-long grid connections in some states, 71% of Americans opposing nearby data centers, and contract clauses letting customers like OpenAI buy late projects outright. And depreciation starts the day you switch on: SB Energy owns buildings, tenants own chips, and chips obsolesce in a year or two — Paul Kedrosky’s image is financing long-lived infrastructure with phone-lifespans inside, “like taking out a 30-year mortgage on an iPhone.”
SoftBank borrows at junk rates (over $11 billion of new bonds at 9.75%) to fund its next OpenAI payment; OpenAI expects to burn ~$280 billion by 2030; OpenAI leases SB Energy’s Ohio campus for 20 years (the revenue underpinning SB Energy’s valuation), invests in SB Energy, holds warrants at $80 billion; NVIDIA bought $1.5 billion of SB Energy shares at a 10% discount with $1.5 billion more at IPO, and guarantees up to $105 billion for the campus (no liability booked until 2028) in exchange for 20 years of exclusive NVIDIA hardware. The same pattern everywhere: SpaceX rents compute to Anthropic at $1.25 billion a month; 85% of Amazon’s latest net income was unrealized gains on its Anthropic/OpenAI stakes, 87% of Google’s from SpaceX/Anthropic stakes. “The industry has essentially agreed to buy each other’s products, guarantee each other’s debt, and mark up each other’s valuations.”
NVIDIA’s sales went from ~$27 billion four years ago to an estimated $410 billion this year with net income nearly doubling — yet it trades under 17x forward profits, its cheapest in a decade, half last year’s price per profit dollar. Jensen Huang calls it “the world’s first and only growth-of-value stock,” “incredibly misunderstood.” Four explanations: (1) cyclical-at-peak treatment, with gross margins expected to slip below 72% and Micron (+180%) squeezing, plus Meta and Alphabet building in-house silicon; (2) NVIDIA’s revenue is everyone else’s spending, and TCW’s Eli Horton notes the multiple only makes sense if AI spending slows — which is roughly what Anthropic’s CEO asked for, knocking the chip index ~6% (a slowdown helps lab margins and hurts the shovel seller); (3) uncertainty pricing — a lab might go bust or capture a $60 trillion market, and Pastor-Verraes/Barberis-Huang explain why uncertainty and lottery-ticket hope add value; (4) price discovery: NVIDIA is priced publicly with short sellers, Anthropic privately with no ability to bet against it, and Miller (1977) showed optimists set prices when pessimists can’t. Conclusion: Anthropic at $2 trillion fails the McNealy test; NVIDIA, at 17x real and growing profits, passes it — and is the one being marked down.
Why don’t the labs slow down? A pause would erode the premium-pricing lead, and competitors aren’t pausing: Chinese open-weight models (DeepSeek, Moonshot) are near frontier performance and taking share; Anthropic and OpenAI both released models 40–50% cheaper on the same day — good for customers, awkward before an IPO. Epoch AI estimates the cost of a given performance level has fallen ~30%—13-fold—per year since 2023, perhaps faster than any transformative technology ever, with prices falling fastest right after each new top model. Inputs (chips, power, even electricians) inflate while outputs deflate. Examples: startup TypeSafe AI’s “JEV” model for developers claims up to 440x cheaper than frontier for simple tasks on $40 million of seed funding (0.002% of $2 trillion); Ramp cut its AI bill 40% with routers that match tasks to models (“you don’t need a Ferrari to go pick up your groceries”). McNealy’s old line — open source is “free like a puppy is free” — is the labs’ best argument against cheap open models, though cheap software on cheap hardware is what destroyed Sun, which ended up sold to Oracle for a fraction of peak value.
There is a real bull case: OpenRouter usage up ~25,000% since the start of last year; 22.5% of Anthropic’s users retained at a year vs ~13.2% for OpenAI (per Yale’s Aleh Tsyvinski); Damodaran seeing Anthropic as possibly the only lab with real end-customer revenue; Ben Thompson’s argument that model-plus-tooling could lock in profits — though Ramp’s routers suggest customers are working to avoid that lock-in. The labs may find a highly profitable model — in the meantime, they’d like to sell you shares. None of this makes AI a fad: the internet changed the world long before Google arrived, after Yahoo, Lycos, and AltaVista; AI could end up like email, used daily and barely monetized — great for customers, not for anyone who paid $2 trillion.
The IPO will price the circular chain for the first time. SoftBank’s $10 billion margin loan is secured against OpenAI shares last valued privately at $852 billion; a listing would monetize that paper but also risk repricing it. OpenAI controls its timing — Altman calls now “ill-advised” for safety reasons while raising privately at $1.2 trillion, “a price for professionals only.” Damodaran’s metaphor: the dot-com correction arrived as many small trees falling until half the forest was gone — and the pulled IPOs, junk yields, and delayed filings may be those trees. His advice to AI companies: “show me the numbers, not just the words,” which only a prospectus forces. Chanos expects record issuance this year, and the academic record (Loughran & Ritter; Baker & Wurgler) says heavy issuance predicts weak returns. Investor Mike Paul’s FT quote lands as the closing warning: the lab CEOs say slow down and take care — the market says the profit motive is overwhelming; “we may look back and wonder why we didn’t take them at their word.” Take them at their word on the price too: at $2 trillion, with even the most generous assumptions, the math doesn’t get you there — and if it goes wrong, someone will ask McNealy’s 2002 question: “What were you thinking?”