On January 27, 2025, investors reacted sharply to the rapid rise of Chinese artificial-intelligence company DeepSeek. Nvidia’s shares fell 16.9%, the Nasdaq Composite dropped 3.1%, and the S&P 500 lost 1.5%. The Dow Jones Industrial Average, however, rose 0.7%, showing that the damage was concentrated mainly among technology and AI-related stocks rather than the entire global market.
The Low-Profile Founder Behind DeepSeek
DeepSeek was founded in 2023 by Liang Wenfeng, an engineer who previously helped establish the quantitative hedge fund High-Flyer. Liang studied electronics and communications engineering at Zhejiang University before completing a master’s degree in information and communication engineering in 2010. He remained relatively unknown outside China until DeepSeek’s models challenged assumptions about American dominance in advanced AI.
DeepSeek-R1 Changed the Conversation
DeepSeek-R1 attracted attention because its developers reported performance comparable to OpenAI’s o1 model on several reasoning tasks. Its training process used reinforcement learning to encourage problem-solving behaviors such as reflection and self-verification. DeepSeek also released model weights for R1 and several smaller distilled versions, allowing researchers and developers to examine, modify, and deploy them more freely than most leading proprietary models.
The Famous $5.6 Million Figure Was Misunderstood
Reports frequently claimed that DeepSeek created its advanced AI for only $5.6 million. That figure was an estimate for the final training run of the earlier DeepSeek-V3 model, calculated from 2.788 million Nvidia H800 GPU hours at an assumed rate of $2 per hour. It did not include previous experiments, research salaries, data preparation, hardware ownership, or the complete cost of developing R1.
Why Nvidia Lost Hundreds of Billions
DeepSeek raised the possibility that powerful models could be trained and operated more efficiently than investors had expected. Nvidia’s 16.9% share decline erased roughly $593 billion from its market value in one trading session. The fall did not prove that AI chips were becoming unnecessary; it reflected fears that future AI systems might require less hardware spending to achieve competitive performance.
Efficient Architecture Was the Real Breakthrough
DeepSeek-V3 used a mixture-of-experts design containing 671 billion total parameters while activating about 37 billion for each token. It was trained on 14.8 trillion tokens and incorporated techniques designed to reduce communication and computing overhead. These architectural choices helped demonstrate that raw computing scale was not the only path toward stronger AI performance.
Open Models Were Closing the Gap
DeepSeek arrived as open-weight models were already becoming more competitive. Stanford’s 2025 AI Index found that the performance gap between open- and closed-weight models had narrowed substantially on some benchmarks. It also reported that the cost of running a system performing at approximately the GPT-3.5 level fell more than 280-fold between November 2022 and October 2024.
DeepSeek Became a Multibillion-Dollar Company
A Chinese stock-exchange filing reported in July 2026 implied a DeepSeek valuation of approximately 350.88 billion yuan, or $51.82 billion. Because DeepSeek remains privately held, that transaction-implied figure is not a definitive public-market valuation and does not establish Liang’s personal net worth. The company has also begun preparations for a possible Shanghai STAR Market listing.
Featured Image: “deepseek auf dem Smartphone” by ccnull.de Bilddatenbank is licensed under CC BY-NC 2.0
Leave a Reply