Jim Chanos — The AI Bubble Is "Much Worse" Than Dot-Com

0:00 Cold open — premium on promises & the Nvidia ceiling

• "In bull markets you put a premium on promises; in bear markets you put a discount on reality" — and "we're clearly in the former." Nvidia is a gatekeeper/price-maker: nothing in hardware that depends on Nvidia should trade richer than Nvidia — "yet many, many companies do."

1:19 Under the surface — dispersion & short-side alpha

• S&P within a stone's throw of all-time highs, but under the surface "a lot of things are acting poorly" and others have "gone parabolic" — a great time for stock dispersion and long-short alpha on the short side.

2:26 Not a short fund — the 40-idea hedged model portfolio

• Chanos & Co. no longer runs third-party money — it runs its own money and advises institutions via a 40-idea model portfolio, designed to be hedged (S&P or custom). The view is "our 40 stocks versus the market," not an outright market call.

3:24 The weirdest tape — semis/DRAM killed while hyperscalers rally

• The SOX down 3%, a DRAM ETF down ~10%, yet hyperscalers up 3%+ and the S&P flat — "on no real news other than flows and sentiment." Large caps moving 5–10% a day is "remarkable" but "highly speculative."

5:16 The third leg of the stool — issuance joins the top

• The signal that had been missing this cycle since 2021 has arrived: record issuance (IPOs, secondaries) + insider selling. Now all three top signs are present — rich valuations, retail speculation, heavy issuance. The analog: first-half 2021, where shorts began working mid-year after GameStop.

7:12 The issuance torrent & off-balance-sheet debt

• SpaceX's $75B IPO + $25B of debt "almost immediately"; Google's $80B→$85B raise; Meta; near-junk raises by neo-clouds and Oracle. "A lot more debt… a lot of it's off balance sheet" — a torrent of credit into a hot sector "with very uncertain returns."

7:56 Uneconomic from day one — worse than 2005–06 housing

• In the housing bubble the assets (homes) held their value; here the physical structures "may be completely uneconomic from day one" — "you just can't get the numbers to pencil out," yet more capital flows in at lower and lower returns.

9:50 Cap-rate fragility — 5–7 caps against a 4.5% ten-year

• "Too many projects get penciled out at mid-single-digit pre-tax rates of return" — office, data centers, storage — everyone doing deals at 5/6/7 caps, layering leverage and mezzanine to promise equity 15%. Rates to 6–7% "blows up asset class after asset class."

10:47 SL Green as the low-cap-rate bellwether

• SL Green trades at a 5-cap on NYC office and "the stock's gone nowhere in 25 years." No position — but a bellwether of "willingness to put really, really low cap rates on assets," which applies to equities too.

11:40 When credit cracks — 5% "with vigor"

• Acceleration toward 5% "with feelings it could go higher" would widen spreads. Lowest-junk CCC spreads are already starting to widen; BB/BBB are not — yet. That's when credit investors rethink buying an office building at a 4.5-cap.

12:25 Much worse than dot-com — SPVs & who spent then vs now

• Meta's Louisiana SPV (Blue Owl/KKR, 20% ownership, 3-year out) is the tell: "no one wants the hard assets." In dot-com only CLECs and fiber failed (~$100B over 5 years); the spenders (BofA, GE, Coca-Cola) were profitable enterprises that just cut orders. Today the spenders outside the hyperscalers are "not necessarily profitable."

16:31 The "sanctity of CapEx" myth — spot prices on 20-year assets

• People treat the CapEx as etched in stone "because marginal returns are so high right now" — but the client pays a 20–25% rate "only for a year or two" while the asset is a 20-year asset. "People are making decisions on long-term projects based on spot prices" — the shale and 19th-century-railroad mistake; a duration mismatch that's "one of the biggest financial crimes… in finance 101."

18:34 Micron & the LTA faith

• DRAM/high-bandwidth memory is the best example: bulls cite LTAs, RPOs and an 80%-gross-margin order book on a company that had negative gross margins three years ago. "History says you should be very cautious about that view."

23:41 Construction-in-progress accounting — the inflated-EPS engine

• GPUs sitting in warehouses "aren't being depreciated" — parked in "construction in progress," a PP&E subaccount that keeps growing; depreciation is deferred until revenue starts. The mirror image: CapEx creates revenue/profit for Nvidia while the spenders capitalize rather than expense it — which is how you get S&P EPS running 4–5x above its 6% trend (just as in '99–2000).

25:32 Depreciation schedules — 5–6 years vs his 10

• Hyperscalers/neo-clouds depreciate over 5–6 years; Chanos uses 10 "to be conservative" and still "can't pencil out most of the data-center economics." Worse: the clock often doesn't start for ~18 months, and the chips economically/technologically depreciate off-P&L in the meantime.

27:20 The asset-light admissions — Coreweave & Nebius

• Coreweave reportedly seeking derivatives to hedge downside in asset prices; Nebius — the most asset-heavy neo-cloud ($4.40 capital per $1 revenue) — pivoting to an asset-light "Marriott-franchise" model. After two years of "own the physical assets, it's gold," that's "a tremendous admission."

29:32 Legacy data centers dumping portfolios — the maintenance-CapEx tell

• Legacy operators "suddenly" putting portfolios up for sale; one bankrupt in 2023 (Sixtera) re-IPOing as C-Squared; executives leaving. His theory: they've realized maintenance-CapEx will run higher than disclosed, so returns are even lower — and they're rushing to unload.

30:48 xAI/SpaceX & promotional spot deals

• xAI "a money pit," overbuilt, now renting out compute via two "highly promotional" ~$1B spot deals (Anthropic, Google) with ~3-month outs — "stealing from Peter to pay Paul to get this on the tape." xAI was bought for SpaceX stock at a $250B value in February.

32:32 Power is NOT the bottleneck

• Contrarian: "there's one thing we are not short of in this country, it's power." Power is ~5–6% of data-center revenues — the smallest cost component. Transmission/regulatory bottlenecks get solved (Texas is already doing it); the bigger risk is political backlash. "Read the fine print" — the Texas grid charges wholesale.

34:24 Incremental ROIC compressing — Oracle the worst

• The metric he watches for clients: hyperscaler return on incremental invested capital has fallen from ~40% (18 months ago) to ~20% today and is "moving toward 10%." At 10% pre-tax, "the neo clouds are in a whole world of hurt" — and Oracle already has "the worst of those ROIC metrics." Late '26/'27, people "start putting pencil to paper": does the next $1T make sense at $50B of earnings when treasuries pay that?

37:33 Apple's discipline & where the blow-ups are

• The market is rewarding names that stayed on the sidelines — Apple's "heroic run." "They didn't take the risk that might prove unfounded." The concern is the capital-intensive side: "that's where the credit's going, the off-balance-sheet financing… that's where the blow-ups are going to happen."

39:32 Circular financing & the incestuous cap table

• Nvidia investing in its own customers who then raise more equity/debt (and pledge it as collateral) — "small relative to the buildout," but a leverage effect and an echo of Lucent/Nortel vendor financing. The founders "there at creation" all invested in each other and are now competitors suing each other (Apple/OpenAI, a coming OpenAI/Microsoft fight).

41:16 Banks tangential; the unnamed PE short

• The epicenter "isn't the banks" this time — they're tangential and have done well (though JPM >3x tangible book is "a little expensive"). His main financial short is "one of the big private-equity entities" with AI and office/low-cap-rate exposure — "if cap rates go up or rates go up, this company is in big trouble." (Apollo/Blackstone's $35B TPU deal is context, not the short.)

44:16 How it shakes out — everything priced as if it worked

• "Like the internet," there will be huge winners — but right now "everything is being valued as if it has worked or it will work," the '99–2000 problem. Castles-in-the-air narratives (data centers in space) "can't be disproven" — until order books collapse and reality gets its discount.