TS Lombard Models AI Crash Scenario: $20 Trillion Wiped Out, but Only Triggering a Mild Recession
nashnova research
TS Lombard analyst Dario Perkins assumes the AI capex boom is unsustainable and models the fallout — roughly $20 trillion in market cap erased, yet the U.S. economy lands in only a mild recession, provided the Fed cuts rates fast enough.
How big is the crash?
Core model output: the S&P 500 drops 30%, tech stocks drop 40%, and ~$20 trillion in market value disappears.
AI-related debt and credit losses hit an estimated $125–190 billion.
Perkins sums it up: "an internet-bubble-scale wealth shock, but only a 1990-scale economic shock."
This means → the stock-market hit could be staggering, but the damage to the real economy stays far milder than 2008.
Where is the money coming from — and going?
U.S. tech companies spend roughly $1 trillion a year on AI infrastructure, but sustaining that pace requires $2.5–3 trillion in annual revenue by 2030 — about 14× current levels.
Most AI revenue growth still comes from capex circulating inside the industry itself. In plain terms = the people selling shovels are mostly being paid by other people selling shovels; genuine outside demand has barely grown.
OpenAI and Anthropic draw roughly 80% of their revenue from just 1% of their customers — most of whom are VC-funded AI startups.
Ramp payment data shows AI token spending has flatlined for two straight months, and Anthropic posted a clear revenue shortfall in Q3.
Six fuses — which one blows first?
Perkins lists six trigger paths and calls the situation "overdetermined" — the question is not whether a crash happens, but which fuse blows first.
Credit risk: AI-related bond issuance is projected to reach $700 billion by 2027, exceeding all eurozone sovereign bond issuance combined; hidden liabilities across the AI ecosystem are estimated at $1.5–2 trillion.
This means → AI expansion now runs on borrowed money, not just equity. If markets turn, the debt chain breaking poses a bigger risk than the stock decline alone.
The report names Meta specifically for using off-balance-sheet structures — five four-year leases that mask what is effectively a twenty-year liability.
Who falls first?
New cloud players such as CoreWeave are most exposed — some already pay roughly 15% financing rates, and CoreWeave's CDS implies about a 63% default probability.
In plain terms = these companies rely on continuous borrowing to meet existing contract payments; if new funding dries up, the payments stop immediately.
Oracle, standalone foundation-model developers, and equipment suppliers follow.
On the chip side, Nvidia has become the bellwether of the entire AI boom. Some buyers double-order to lock in scarce hardware; once demand softens even slightly, the shortage could flip to a glut.
Why only a "mild recession"?
Perkins's conclusion rests on one critical assumption: the Fed cuts rates quickly after the crash, boosting the lower-income consumers previously squeezed by high rates.
This reflects a K-shaped recovery logic — the AI bubble mainly hurts wealthy holders of tech assets, while rate cuts activate bottom-tier spending, partially offsetting the drag.
He states explicitly that whether the Fed acts in time is the single most important variable in this forecast.
What else could accelerate the crash?
Chinese open-source models already account for over half of token traffic on Vercel; Perkins calls this "China Shock 3.0." This means → U.S. AI companies' pricing power is eroding.
Survey data shows roughly 55% of Americans want AI development slowed; only 13% oppose slowing down. A major AI safety incident or regulatory tightening would pile further pressure on the sector.
Depreciation charges from newly built data centers will squeeze profits, and a wave of AI IPOs could open concentrated insider-selling windows.
市场有风险，内容仅供研究参考，不构成投资建议。