Who Owns DeepSeek, The Chinese AI Company?

Chinese AI lab DeepSeek made headlines and caused disruption in the AI industry when it launched the R1 model in 2025. It uses a concept called mixture of experts (MoE) to efficiently generate AI responses at a much cheaper cost than its competitors. This resulted in the DeepSeek app becoming very popular in a short span of time. The R1 model launch triggered a global technology stock selloff in 2025 and made a clear statement to Western AI companies: China was by no means behind in the AI race. The company recently launched the V4-Flash model, which is over 90% cheaper to run than peer models such as Kimi K3, GPT-5.6 Sol, and Claude Fable 5. However, it must be noted that the cost-effective DeepSeek models are subject to censorship imposed by local Chinese laws, which may curb the company's future growth. It's one reason why you should stay away even though DeepSeek AI might be the best ChatGPT rival right now.

The company was founded by an electronic engineer from Zhejiang University named Liang Wenfeng, who funded the launch of DeepSeek with profits from his hedge fund, High-Flyer, in 2023. Liang co-founded High-Flyer, which uses math and AI to make investment decisions. In DeepSeek's second funding round, the company secured about $15 billion from large backers such as Tencent Holdings and CATL. 27% of this funding is state-owned, with the China National AI Industry Investment Fund contributing about $150 million. Liang is one of four Chinese tech entrepreneurs who have been labeled by the media as China's "fantastic four," a group that includes DJI's Frank Wang. Wang reportedly offered Liang the opportunity to co-found DJI when it was still in its early stages. DeepSeek is reportedly looking into a potential IPO on Shanghai's Nasdaq-like STAR Market.

With DeepSeek's efficiency, Liang seems to have created a key differentiator in a market so highly contested that it is shaping the fortunes of multitrillion-dollar companies, and is wrapped up in a resource battle that is altering the geopolitical landscape. When the R1 model was released, the global market interpreted DeepSeek's AI compute efficiency as an indictment of resource-heavy AI models built by companies like OpenAI. If powerful AI could be built with less demand on GPU chips produced by companies like NVIDIA, demand for those chips could decline heavily. Accordingly, this triggered a massive sell-off that caused NVIDIA to lose $589 billion in market value in a single day — the largest single-day loss in the history of the U.S. stock market.

Before U.S. export controls took effect in 2022, Liang's hedge fund High-Flyer, among other Chinese entities, stockpiled NVIDIA GPUs for its Fire-Flyer 2 cluster, which DeepSeek later used, in anticipation of further U.S. sanctions that had disrupted the supply of chips to China before. The availability of GPU chips in China is something the U.S. has been trying to curb, to prevent China's relentless pursuit of overtaking America's technological prowess. In response to these restrictions, China has been advancing its homegrown chip manufacturing, to reduce its reliance on Western chips and manufacturing tools.

DeepSeek's efficiency breakthroughs in AI compute could benefit the entire AI industry, which has an opportunity to learn from them to relieve the heavy resource needs of data centers that power most AI models. DeepSeek achieved its industry-leading efficiency through a combination of its MoE architecture, hardware optimization, model distillation, and reinforcement learning, which together lowered model training costs.

Its MoE architecture only uses the necessary portions of the model for a specific task, which drastically reduces the required compute resources. The company also optimized its AI models to work efficiently on less powerful GPUs that are cheaper to acquire and are more readily available. DeepSeek also uses model distillation, where larger and more complex models train smaller models while retaining most of the large model's intelligence. This reduces compute demand. Reinforcement learning methods were also used to automate the training of models, which reduces the reliance on expensive human-labeled data.