From the 'Three No's' Principle to Billion-Dollar IPO Rumors: Liang Wenfeng's Three-Year Battle with DeepSeek

Chinese AI startup DeepSeek, founded by Liang Wenfeng, has gone from adhering to the principle of "three no's" (no financing, no profit, no IPO) to rumors of a multi-billion dollar IPO. In 2023, DeepSeek's team consisted mainly of researchers without commercial experience, and the company did not attract external funding. However, after the success of the DeepSeek-V3 and DeepSeek-R1 models, which cost only $5.6 million to train (with inference costs up to $6 million), the company changed course. In 2025, rumors emerged about an IPO in Hong Kong and the construction of a data center with 100,000 GPUs. DeepSeek also faced competition from Alibaba Cloud (Qwen2.5-Max model), Moonshot AI (Kimi K2), and Baidu (Ernie 4.5 and Ernie 5.0), but managed to attract attention with its efficient model. At the end of 2024, DeepSeek released DeepSeek-R1, which became popular among the developer community. The company also adopted the Mixture-of-Experts (MoE) architecture and Multi-head Latent Attention (MLA), and developed its own DeepSeekMoE and DeepSeekMLA mechanisms. In early 2025, in response to the price war, DeepSeek reduced API costs by 75% for DeepSeek-R1 and 50% for DeepSeek-V3, and introduced the DeepSeek-R2 model. At the same time, there were rumors of raising $2 billion in funding from Alibaba Group and plans for an IPO worth up to $10 billion.