Gromen’s claim (09/2026): in the prior five US capex booms of the last ~200 years, it paid to buy gold once the boom was 2–3 years old, because gold outperformed the boom sector over the rest of the boom/bust cycle. “Is it different this time?” — the AI capex boom is now ~3 years old (AI circular financing, The Mag-7 free-cash-flow bet).
The stat has a hole: for four of any five candidate booms (canals, railroads, electrification, 1920s autos — the fifth being 1990s telecom), the dollar price of gold was fixed by the gold standard. “Gold outperformed the boom sector” then means only “the boom sector fell,” which is the trivially true half of every boom/bust. The one free-float specimen, 1997–2003, does fit (Nasdaq −78%, gold flat-to-up from its 1999 low), n=1. What survives is the reference class itself: capex booms overbuild, the builders’ returns collapse, and the surviving gains go to the users of the cheap capacity (rail freight, dark fiber). Hold that as the bear case for the chip and data-center builders, not as a gold signal — gold’s own case runs through Real rates and gold and the Debasement trade, which Gromen sells.
The sell side joins the reference class: “JPMorgan is worried about an autumn downturn, says AI stocks show similarities to 2000 tech peak” (CNBC, 08/21/2026, via Barchart https://x.com/Barchart/status/2091671768440152250). Headline only; the comparison’s content is what Buffett Indicator and The Mag-7 free-cash-flow bet already carry.