A startup most people hadn’t heard of a week ago just vaporized $589 billion from a single company’s market cap. DeepSeek, a Chinese AI lab based in Hangzhou, released its R1 and V3 reasoning models on January 27, 2025, and the shockwave rippled through global markets.
Nvidia bore the brunt of the carnage, suffering what appears to be one of the largest single-day market cap losses in the history of publicly traded companies. The logic was brutal and simple: if a Chinese startup can build AI models that rival OpenAI’s GPT-4o using a fraction of the hardware, maybe the world doesn’t need quite as many $40,000 GPUs as everyone assumed.
The hardware math that broke Wall Street
DeepSeek’s V3 model was trained using just over 2,000 Nvidia H800 GPUs. For context, leading Western AI labs typically throw tens of thousands of cutting-edge chips at training runs of comparable scale. The H800 itself is a downgraded chip that Nvidia specifically designed to comply with US export controls on advanced semiconductors to China.
DeepSeek claims its models match the capabilities of systems built by companies with access to the most advanced chips money can buy. Futures tumbled. Chip stocks across the board sold off. The entire thesis underpinning semiconductor valuations, that AI development requires ever-larger capital expenditures on the most advanced silicon, suddenly looked a lot less certain.
