DeepSeek, a subsidiary of the High-Flyer quantum hedge fund, was founded by Liang Wenfeng and two classmates from Zhejiang University in February 2016. The company was funded entirely by the fund, avoiding external venture capital and deadlines for revenue. The firm built its own clusters, including Fire-Flyer I with 1,100 GPUs for ¥200 million in 2019 and Fire-Flyer II with approximately 10,000 A100s for ¥1 billion in 2021.
The company released several models over several years, including DeepSeek Coder and DeepSeek LLM 67B in November 2023, V2 in May 2024, and V3 in December 2024. The technical approach was driven by hardware limitations, specifically the use of H800 GPUs with reduced interconnect bandwidth (approximately 400 GB/s NVLink vs 900 GB/s). To overcome this, they implemented Multi-head Latent Attention (MLA) to compress the KV cache, DeepSeekMoE to balance expert load, and FP8 training from the start.
The final run of V3 required 2.788 million GPU-hours of H800, costing $5.576 million at $2 per GPU-hour. SemiAnalysis estimated total group investments at approximately $1.6 billion. The R1-Zero model was trained using 648 H800s for 198 hours, while R1 used 648 H800s for about 80 hours, and the SFT dataset assembly took approximately 5,000 GPU-hours, totaling $294 thousand for the RL stage.
On January 20, 2025, DeepSeek released R1 under the MIT license. By January 26, the DeepSeek app topped the US App Store, surpassing ChatGPT. On January 27, Nvidia's shares fell by 17%, wiping out $589 billion in market capitalization in one day, the largest single-day drop in US stock market history. The Nasdaq lost about 3%, and the total market dip was approximately $1 trillion, with Jensen Huang losing $21 billion of his fortune.
DeepSeek's business strategy focused on open-source distribution and aggressive pricing. The model's weights were released under MIT, leading to 10.9 million downloads on Hugging Face by September 2025. The cost for input tokens is $0.55 per million, and output tokens are $2.19 per million, which is 27 times lower than o1's prices ($15 and $60 respectively).
