A closer look at DeepSeek’s ownership and operations

OpenAI has accused DeepSea for using its proprietary model without permission, whereas European countries such as Germany and Ireland are investigating into the company. President Trump said it's a wake-up call for U.S. high-tech companies. And I think the U.S. chip ban on advanced semiconductor chips on China did not work. So how did the Chinese AI firms obtain access to advanced AI chips? Didn't the U.S. have an export ban on advanced AI chips such as NVIDIA's 800 and H100 chips to China and Hong Kong? And this ban was in place since October 2022. So how did the Chinese obtain them? So let me play you a video in which a Chinese showed off his shipment of NVIDIA's H100 GPU cards. So this is the second video that he released on the following day, where he unboxed the H200. This is one of the 200 cards that he received. We just opened it this morning. I'll show you the process of unboxing in a minute. In the last video, a friend, a lot of friends, said in the comments that this is illegal. Our thing, our circle is illegal. For the law of the United States, Trump will be very angry at this, right? I've been doing this for two years. Every time it's a lot. He's been on the U.S. blacklist twice. But I think he's made a contribution to the calculation of AI models in China. Otherwise, there would be no such active people, right? Where did these cards come from? What do we use to calculate models? At present, there are only two countries in the world, China and the United States, who have the greatest demand for this. Do you agree? Let's take a quick look at the unboxing video. Some people say he may have a hard time making money from his normal business, but smuggling these advanced AI chips is certainly a lucrative business. The Chinese use various methods to circumvent the U.S. export bans. There are several ways to do this. Smugglers duplicate serial numbers on servers equipped with AMD chips. This makes tracking any forms of enforcement more difficult. Chinese entities also establish or use data centers in other countries where they place bulk orders for these advanced chips, and then redirect a portion of them to China. The truth is Chinese AI companies like DeepSeek have amassed large quantities of chips long before the export ban. The quantity is so impressive that maybe DeepSeek's founder decided to do something about it. According to a Sohu January 27th article titled, DeepSeek founder Liang Wenfeng said, Chinese AI cannot always follow. Someone will inevitably stand at the forefront of the technology. DeepSeek's sibling company, PhantomQuant, which is a company which is a fully automated, quantitative trading financial services firm using machine learning technologies. Now this company has stockpiled tens of thousands of AMD chips as early as 2021. And according to the Chinese media, this was mentioned in a research paper co-authored by the founder Liang Wenfeng. And he said, in deep learning training, we developed the Firefly 2 system, which is equipped with 10,000 A100 GPUs. Now it was around 2021 that this was happening. And PhantomQuant, the company at the time invested 1 billion yuan. And reports also stated that in 2021, fewer than five companies in China possessed over 10,000 GPUs. And Liang's AI driven investment firm was the only non-tech company among them. So due to its large scale stockpiling of AMD chips, PhantomQuant became one of China's earliest major AMD chip buyers. And according to its official website in 2020, the company had idle computing power that reached 15.33 million GPU hours. Now you have to remember when this was happening, DeepSeek wasn't even in existence. So by comparison, in the US, Microsoft announced a custom built supercomputer for open AI featuring 10,000 GPUs. The supercomputer was designed to train open AI's large scale AI models, including GPT-3. Nvidia's A100 card was announced in the same month. I think it's May 2020. So open AI's 10 GPUs at the time were likely Nvidia's V100 or early A100 models. And open AI's total computing power was estimated at 40 to 50 million GPU hours. We don't know the utilization rate. So we don't know the exact idle GPU powers that open AI had in 2020. So the fact that PhantomQuant had 15 million plus idle GPU hours in 2020 means that the company's computing power was on par with open AI's. In 2021, there are only a few organizations in the world possessing 10,000 advanced GPUs or more. But the Chinese media said China already had five organizations or nearly five organizations that each had more than 10,000 advanced GPUs in 2021. So what does this tell you? Liang and whoever were behind him certainly had the vision of getting into the AI competition or war and have been planning for DeepSeek as far back as 2020. The equity structure of DeepSeek is very complex. The company was established on July 17, 2023. Liang Wenfeng owns 1% and the Ningbo Enterprise Management Consulting Limited Partnership owns 99%. And then if you look at the image on the right, as far as who owns this Management Consulting Limited Partnership, Liang owns 50.1% of the consulting firm. But the consulting firm was set up on July 31, 2023, two weeks after DeepSeek was established. Here's my question. The company that owns you was set up after you were born. This is like one of your parents is born after you were born. But behind these companies, there are several layers of ownership. And the majority of shares are controlled by limited partnerships based in Ningbo, Zhejiang, with the primary partner of these limited partnerships being Liang Wenfeng himself. It feels somewhat similar to Jack Ma's Alibaba's equity structure, which was famous for its complexity. And the reason Alibaba's ownership structure was so complex was because Jack Ma was holding shares for many other Chinese officials and their families whose names cannot be revealed. So people wonder, is it for the same reason that DeepSeek's ownership structure is also very complex? So here you have an ownership structure. So let me tell you how to read it because it's in Chinese. The red box is DeepSeek. The purple shaded boxes are various consulting firms and limited partnerships that own companies that own DeepSeek. The individual owners are color-coded in blue. So these are individuals, and they're the same set of people. The cream-colored boxes at the bottom are DeepSeek's subsidiaries. 100% is in Beijing. 0.1% is in Hangzhou. So from this, you can see that DeepSeek's operation is all in Beijing, and Hangzhou is just a shell company. And I calculated Mr. Liang's ownership of DeepSeek. Although, you know, through a complex web of ownership, Liang owns about 92% of the AI company. Even though at a glance, he doesn't own more than 68% of any entities on the chart. There's a podcast called China Talk, which focuses on China and U.S.-China relations. This podcast analyzed how censorship affects Chinese AI chatbots last year. The host tested four of China's top LLMs, or large language models. China Talk found that DeepSeek enforces censorship more strictly than its competitors. And this March 2024 test assessed these models using open-ended questions in both English and Chinese about politics, law, and history. The conclusion? While DeepSeek was the most, quote-unquote, disciplined model among the four. So China Talk concluded, DeepSeek wins the gold medal for following the party's line. On its official website and on Hugging Face, its responses align with the CCP's ideology and socialist values.