What happens when the AI bubble pops?

Fundamentals

A share price is earnings times a multiple, so a crash is some mix of lower earnings and a lower multiple:

P₁/P₀ = (1 − ε·d·k) × (1 − m·resettable share) × (1 − c)

d is a miss against expectations.
ε is an assumed exposed-revenue share times an assumed earnings amplification. The resettable share is the forward P/E above an assumed normal, not an identified AI premium or observed historical norm.
m resets that share
c assumes a general-market decline.
k is the assumed share of the earnings hit priced as lasting, so the multiple applies to normalized rather than trough earnings.
P/E also reflects growth and accounting choices; this identity is not the Campbell–Shiller present-value decomposition.

Several of the largest companies have forward P/Es at or below their assumed normal, so they have no resettable share and the earnings miss carries their whole decline. At the default k = 100% the multiple is held fixed as earnings fall, although cyclical companies' P/Es often rise when earnings are depressed; for low-P/E cyclicals that may overstate the fall. Lowering k lets the P/E on trough earnings rise. Its value is a chosen assumption, not an estimate of how cyclical each business is.

CoreWeave, Nebius and Cerebras instead use enterprise value: stress revenue and the valuation multiple, then subtract dated net debt. Cash, debt and share counts are held fixed; refinancing, cash burn and dilution remain outside the calculation.

Factor stress

For every stock we regress two years of weekly returns on two factors:

rᵢ = α + β_mkt·r_SPY + β_semi·SEMI + ε

The second factor is SMH beyond its market exposure, which includes semiconductor influences unrelated to AI. Fundamental scenario returns are weighted using dated IVV holdings (an S&P 500 proxy) and actual SMH holdings, then pushed through the measured betas. Unmodeled holdings receive only the assumed general-market decline. Both calculations share assumptions; neither is an independent check on the other.

Blend, range and history

We mix the two estimates using the weight you pick. The average is taken in log terms, which handles large drops better than a plain average. The illustrative band shows how much the answer could move. It adds up three things: the random part of each stock that the factors don't explain, how far the two methods disagree, and how much the result changes when company assumptions change. For a group of stocks, the random parts move together the way they have in the past.

The band is a rough guide, not a tested probability range. It leaves out uncertainty in the factor drops and in the measured betas. The ends of each assumption's range are our choices, not statistical limits.

Each company is matched to a dot-com company that played a similar role. For those still trading, we use their 1999–2003 fall from peak, based on month-end prices. Four have since disappeared, so we use a rough fixed loss for them. The severe preset was first set to match the S&P 500's 2000–02 fall. That means it was fitted to that history, not tested against it.

The economy

Output gap = (domestic investment loss + phased wealth effect) × (1 − cushion)

Dated global cash investment is annualized as an explicit scenario baseline, then bridged to assumed AI spending, US installation and domestic value added. These are distinct steps.

FRED household equity assets anchor the wealth base; an editable portfolio exposure maps it to the scenario. The research estimate of 3.2¢ of annual spending per lost dollar at two years phases in over an assumed 24 months, applied to each monthly wealth loss. A Federal Reserve estimate for equity wealth alone is about 1¢. The wealth effect is most of the fully phased output gap.

Output effects are divided by current-dollar GDP as a first-order volume approximation with prices held fixed. Okun's coefficient gives illustrative unemployment pressure, not predicted layoffs or a dated unemployment peak.