DeepSeek's ownership and the impact of its efficient AI models

The origins of DeepSeek

DeepSeek was founded by Liang Wenfeng, an electronic engineer who graduated from Zhejiang University. He initially funded the launch of the company in 2023 using profits from his hedge fund, High-Flyer. This fund uses a combination of mathematics and artificial intelligence to guide its investment decisions.

Liang is recognized as one of China's "fantastic four" tech entrepreneurs, a group that also includes Frank Wang of DJI. Wang reportedly offered Liang the chance to co-found DJI during its early stages before Liang started his own venture. This pedigree highlights the high level of expertise behind the company's technical achievements.

The company has since scaled significantly through a second funding round. It secured about $15 billion from major backers, including large corporations like Tencent Holdings and CATL. These investments have allowed DeepSeek to challenge the dominance of Western AI firms in the global market.

State backing and ownership

While DeepSeek has private backing, it also maintains significant ties to the Chinese government. About 27% of the company's funding is state-owned. Specifically, the China National AI Industry Investment Fund contributed approximately $150 million to the venture.

The ownership structure reflects a mix of private capital and national interest. This dual backing allows the company to navigate the complex landscape of Chinese tech development. It also places the company under the oversight of local laws, which include specific censorship requirements.

These regulations may impact the company's future growth and international reach. While the models are highly efficient, they must comply with local mandates. This creates a unique environment where the company competes globally while remaining bound by domestic rules.

How the efficiency works

DeepSeek achieves its cost advantages through a concept called mixture of experts (MoE). This architecture allows the model to only use the necessary portions of its system for a specific task. By doing so, it drastically reduces the amount of compute resources required to generate a response.

The company also employs model distillation to improve performance. In this process, larger and more complex models are used to train smaller models while keeping most of the original intelligence. This technique allows DeepSeek to create smaller models that still perform at a high level without needing massive data centers.

Furthermore, the company uses reinforcement learning to automate the training process. This reduces the need for expensive human-labeled data, which is a common bottleneck for other AI developers. They also optimize their models to run on less powerful and cheaper GPUs that are easier to acquire.

Market and geopolitical impact

The launch of the R1 model in 2025 sent shockwaves through the global technology market. It suggested that powerful AI could be built without the heavy resource demands seen in models from companies like OpenAI. This realization led to a massive sell-off in global technology stocks.

The impact on hardware manufacturers was particularly severe. Because DeepSeek showed that high-end chips might not be strictly necessary for efficiency, NVIDIA saw its market value drop by $589 billion in one day. This marked the most significant single-day decline ever recorded in the U.S. stock market.

The situation is also tied to U.S. export controls on high-end chips. Before these controls took effect in 2022, Liang's hedge fund stockpiled NVIDIA GPUs for a cluster that DeepSeek eventually used. China is now advancing its own chip manufacturing to reduce its reliance on Western tools and hardware.

Looking toward the future

DeepSeek is currently exploring a potential IPO on the STAR Market in Shanghai. This move would represent a major milestone for the company as it seeks to formalize its position in the global economy. It aims to solidify its role as a primary rival to established Western AI platforms.

The company's V4-Flash model is priced to be over 90% cheaper to operate than competitors like Kimi K3, GPT-5.6 Sol, and Claude Fable 5. This price point makes it a significant threat to the current market leaders who rely on more expensive infrastructure.

As the industry watches DeepSeek, the focus remains on whether its efficiency can be replicated by others. If the company's methods become the standard, it could shift the entire landscape of how AI models are trained and deployed worldwide.