DeepSeek

DeepSeek is a research laboratory subsidiary of High-Flyer Capital Management, a Chinese quantitative hedge fund. It was founded on July 17, 2023 by Liang Wenfeng, High-Flyer's founder and managing partner. Unlike most frontier AI labs, DeepSeek is explicitly research-first: its stated mission is to "solve AGI," and it operates with no apparent pressure to generate near-term product revenue. The lab is headquartered in Hangzhou, China.

In December 2024, DeepSeek published V3 alongside an official training cost of $5.576 million. That figure applied solely to the formal training run and excluded prior research, ablation experiments, and the infrastructure required to build the compute cluster. Even with those caveats, the number was a fraction of what US frontier labs spend -- GPT-4's training run is estimated at over $100 million. DeepSeek achieved this through a combination of mixture-of-experts architecture, FP8 mixed precision training, and efficient attention mechanisms that dramatically reduced GPU hours.

On January 20, 2025, DeepSeek released R1, a reasoning model competitive with OpenAI's o1. It topped the Apple App Store globally within days, displacing ChatGPT. The market reaction was immediate and severe: Nvidia's stock fell 18% in a single trading session, wiping more than $600 billion in market capitalization. Commentary across the industry called it a "Sputnik moment" -- a signal that American AI dominance could not be assumed. DeepSeek's open-weight approach means the model weights are freely downloadable, enabling self-hosting and independent research without vendor lock-in.