The DeepSeek shock: AI capex assumptions get repriced

On 27 January 2025, claims that Chinese startup DeepSeek had trained a competitive AI model at a fraction of assumed cost triggered a broad repricing of AI-infrastructure stocks. NVIDIA fell roughly 17% in one session — the largest single-day market-value loss for any company up to that point — and AI-linked names fell across the board.

At the time, DeepSeek's model was publicly available and its performance claims were being rapidly tested, but its true training cost, hardware access and reproducibility were all disputed. The selloff traded on a question, not a settled fact.

AI-semiconductor and data-centre-linked stocks fell sharply for a session, with partial recoveries in the following days as analysts debated how much compute demand the cheaper-training story actually threatened.

Whether cheaper training genuinely reduced long-run chip demand or increased it (the efficiency-rebound argument), what hardware DeepSeek had actually used, and whether US export controls had failed or worked.

AI capex guidance from the major platforms stayed elevated in subsequent quarters, and the episode settled into the reference case for how fragile consensus AI assumptions can be against a single credible challenge.

AI-infrastructure names remain a core coverage area; every earnings cycle since has been read partly through the question this episode raised — what happens to the AI trade when its cost assumptions move.