Stocks Sink in Broad AI Rout Sparked by China's DeepSeek
2 min read
The story
U.S. markets sold off sharply, led by the Nasdaq, after Chinese AI lab DeepSeek published results showing it trained a large language model rivaling top U.S. models at dramatically lower cost and using fewer high-end GPUs. Nvidia bore the brunt of the reaction, falling roughly 16% in one of its largest single-day drawdowns in years — wiping out hundreds of billions in market cap. The move rippled across the entire AI infrastructure complex, with data center suppliers, power plays, and networking names all posting double-digit losses.
The core fear is straightforward: if AI training and inference can be made dramatically more compute-efficient, the insatiable demand for Nvidia GPUs that has driven its revenue up 65%+ year-over-year may be overstated. Nvidia's fiscal-year revenue of $215.9B and 71% gross margins reflect a supply-constrained, pricing-power environment — one that a true efficiency revolution could erode. Names across the AI capex chain — from server ODMs to power infrastructure — faced the same re-rating logic.
The honest bull case rests on Jevons' Paradox: cheaper, more efficient AI typically expands total adoption and aggregate compute demand rather than shrinking it. Nvidia's $4.90 diluted EPS and industry-leading margins suggest the business is structurally profitable even in a more competitive landscape, and hyperscalers have not publicly signaled capex reductions.
The bear case is that DeepSeek is a proof-of-concept for a structural de-rating of GPU intensity per AI workload — and if even a portion of the demand thesis was priced on inefficiency, the multiple compression has further to run. With Nvidia having traded at a significant premium to the broader market heading into this event, the valuation buffer is limited if consensus EPS estimates get trimmed.
The immediate watch points are hyperscaler capex guidance in upcoming earnings calls (Microsoft, Meta, Alphabet, Amazon), any Nvidia management response or investor day commentary, and whether DeepSeek's efficiency claims hold up to independent scrutiny. The next Nvidia earnings print will be the real verdict on whether orders and ASPs are bending.
The case — both sides
Nvidia's $215.9B revenue base growing at 65% YoY with 71% gross margins reflects a structurally dominant moat, and historical precedent (cheaper compute driving more compute demand) suggests DeepSeek-style efficiency gains expand the total addressable market rather than cannibalize GPU orders.
If DeepSeek's approach proves reproducible at scale, it de-rates the GPU-intensity assumption embedded in Nvidia's forward estimates, and with the stock having traded at a steep premium multiple heading into the event, even a modest EPS trim could justify further multiple compression from here.
The house read
Leans bearNVDA and the AI infrastructure complex face the question of whether DeepSeek's efficiency breakthrough is a genuine demand destroyer or a Jevons-paradox accelerant that ultimately expands total GPU consumption.
Wrong ifHyperscalers confirm unchanged or accelerated GPU capex on upcoming earnings calls — Meta, Microsoft, Alphabet, Amazon — which would signal Jevons paradox is playing out and trigger a sharp reversal in the entire AI infra complex.
Published read · research, not advice