Nvidia fell roughly 16% in one session — one of the largest single-day market-cap destructions in U.S. equity history — after China's DeepSeek published results showing a frontier-class large language model trained at a fraction of the compute budget previously assumed necessary. The move dragged the broader AI infrastructure complex down sharply, with many names off double digits. Nvidia's most recent fiscal year showed revenue of $215.9B, up 65.5% YoY, with 71.1% gross margins and $4.90 diluted EPS, underscoring how much premium was baked into the stock before the shock.

The DeepSeek development matters because Nvidia's bull thesis rests heavily on the idea that frontier AI requires ever-larger GPU clusters — a compute-intensity assumption that has driven hyperscaler capex guidance and Nvidia's data-center revenue surge. If efficient training methods can deliver competitive models with far fewer H100/H200 hours, the addressable demand pool narrows, and Nvidia's pricing power over the next GPU cycle is less obvious. The companies most directly touched include Nvidia itself, but also the broader AI infrastructure stack: power, networking, memory, and cloud capex beneficiaries.

The bull case is straightforward: even at reduced training compute per model, inference demand scales with deployment, and Nvidia's CUDA ecosystem moat and margins remain exceptional. A 16% single-day drop on a still-unverified efficiency claim from a Chinese lab is arguably an overreaction into a stock that was already pricing in some deceleration. The bear case is that this is the first credible evidence the compute-intensity assumption was overstated, and if hyperscalers quietly trim GPU orders in response, the consensus revenue ramp for FY2027 gets revised lower before the market has fully repriced.

The key watch items are: (1) any hyperscaler capex commentary on upcoming earnings calls, (2) whether DeepSeek's efficiency claims hold up to independent replication, and (3) Nvidia's own next earnings print and forward guidance. Until at least one of those resolves, the range of outcomes is wide enough that the setup is genuinely two-sided rather than a clean fade or a clean buy-the-dip.