Few modern technological advances have been more impactful than China’s DeepSeek AI. Since the release of the AI-powered chatbot’s new model, in January of 2025, DeepSeek has rattled global markets, eliciting responses from various governments and international organizations. 

The frenzy for DeepSeek is understandable. Not long after its release, DeepSeek-R1 broke records, passing the already established ChatGPT to become the most downloaded free iOS app in the U.S. Just 18 days after its release, the AI-powered chatbot had 16 million downloads, far surpassing ChatGPT’s 9 million downloads within the same time frame. 

DeepSeek’s viral success has led to disruptions and chain reactions in international markets. Semiconductor companies, like American tech giants Nvidia and Broadcom, experienced monumental falls in the stock market. Other tech companies like Microsoft and Google’s parent company Alphabet also demonstrated the same trend. Even President Donald Trump acknowledged the impact of DeepSeek, calling it a “wake-up call” for AI companies in the United States. 

DeepSeek has indeed been a wake-up call for AI companies around the world, introducing China’s fast development and capacity for AI despite limitations. In recent years, rising geopolitical tensions between the United States and China have led to competition in the AI market. In fact, many have pointed out the parallels between the “AI Cold War” between the U.S. and China and the Cold War between the U.S. and the former Soviet Union. The strained relationship between the United States and China has led to trade regulations in key industries including semiconductors and artificial intelligence. In 2022, the Biden administration imposed strict trade regulations that limited the chips that China could buy from the U.S., preventing Chinese companies like DeepSeek from having the most advanced chips created by companies like Nvidia. These limitations meant that the company had to find alternatives to close the technological gap and, judging from the results, it seems that they have succeeded. 

But how exactly was DeepSeek able to achieve this success? To fully understand the topic, we have to go back to the beginnings of the creation of the company. 

The Beginnings of DeepSeek

In May of 2023, DeepSeek was founded by Zhejiang University alumnus Liang Wenfeng. Prior to DeepSeek, Liang founded the Chinese hedge fund High-Flyer which helped lay the foundation for DeepSeek’s success. Back in 2016, High-Flyer differentiated itself from other hedge funds through its use of AI models to determine stock positions, and in 2017 it hired a talented team of researchers that focused on AI. Many of these technicians from High-Flyer went on to work for DeepSeek. High-Flyer was in a unique position to develop AI because back in 2021 the company had bought thousands of Nvidia graphic processors before the regulations on China. These processors played a critical role in helping DeepSeek overcome the chip limitations when building their model. 

Other than the insight to buy technology, High-Flyer has played a pivotal role in funding and staffing DeepSeek’s rise. The ample funds (a $13.79 billion portfolio) from High-Flyer have allowed the hedge fund to become the primary investor of DeepSeek. This self-funded system has allowed DeepSeek to focus on developing technology without disruptions from outside investors and shareholders.

In addition to the available talent, funding, and technology, the innovative techniques that DeepSeek used were crucial in helping to train the chat box. Techniques like reinforcement learning, reward engineering, supervised fine-tuning, and distillation cut down on inefficiencies and helped DeepSeek overcome the use of lower-quality CPUs (Central Processing Units). One particularly important technique used was distillation, which is the use of preexisting larger models to train smaller models.

“They were able to train their models on other, or slightly less capable GPUs (Graphics Processing Units). They were not state-of-the-art GPUs, but they made some interesting innovations in the architecture of the machine learning models that allowed them to train models even on less powerful hardware,” said Umar Iqbal, an Assistant Professor at the Washington University in St. Louis, in an interview with me. 

DeepSeek’s Development and Impacts

The approaches that DeepSeek has taken have clearly been effective as DeepSeek has been consistently putting out new models. In November of 2023, not long after the creation of the company, DeepSeek released its first-ever model, DeepSeek Coder, that specialized in tasks related to coding. Just a month later, in December of 2023, DeepSeek released its second model, DeepSeek LLM, which was the first version of the company’s general-purpose model.

In 2024, researchers at DeepSeek continued to create new and improved models at an astonishing pace. In May of 2024, DeepSeek-V2 was released and saw an improvement in performance and training costs. Two months later, in July 2024, the company put out another coding-focused model with DeepSeek-Coder-V2. The model was an improvement from DeepSeek Coder, having 128,000 tokens and 236 billion parameters. At the end of 2024, DeepSeek continued to add to its AI collection, with DeepSeek-V3 in December of 2024. The model had widened to 671 billion parameters and was able to accomplish more advanced tasks than previous models, showcasing better reasoning skills and strong performance in coding and mathematics. 

However, it was not until the release of DeepSeek-R1 in January 2025 that DeepSeek exploded in global popularity. The model seemed to be exactly what the market was missing. DeepSeek claimed to have functions on par with those of big companies like ChatGPT, with only a fraction of the cost. According to DeepSeek, their V3 model had cost only 5.6 million dollars to train, a huge difference from the estimated 100 million dollars that models like Open AI’s ChatGPT needed. 

The revelation of the alleged cost of creating DeepSeek sent the global markets into a frenzy. Before DeepSeek, it had been a widely accepted idea that the creation of large language models was for the most part reserved for larger companies that had more funds available. However, DeepSeek’s case tore that narrative apart by demonstrating that creating a model like DeepSeek or ChatGPT didn’t have to be very costly. This new notion is extremely important as it will lead to more innovation and the creation of newer technologies.

“I think there will be more innovation. If let’s say, it takes a lot of money to create a model, there will be fewer models that exist in the world. But if it becomes cheaper to create a model with less capable hardware, perhaps with less data, you can imagine more models coming in,” said Iqbal. 

In fact, many firms have already been inspired to develop AI because of DeepSeek. Sources report that, since the success of DeepSeek, many Chinese companies have increased orders for the H20 chip in hopes of creating an AI model of their own. Research and investment in AI has also risen. For example, Alibaba-backed firm Zhipu recently secured over $138 million in funding for its new AI developments, and other smaller companies have come to join the tech race. 

Uses and Concerns 

The widespread development and use of Deepseek, from industries to individuals, has brought about changes. For example, in China DeepSeek has been integrated into companies and organizations everywhere. Dozens of automobile manufacturers in China have added DeepSeek into their vehicles and many hospitals and pharmaceutical companies have employed DeepSeek into patient care. In the southern city of Meizhou, the local government even used DeepSeek to help answer the government helpline.

Outside of China,  many businesses, including Western companies, have also been contemplating the use of DeepSeek for cheaper artificial intelligence. Even companies that don’t plan on using DeepSeek due to security concerns think that DeepSeek can help them by driving down AI prices. For Professor Iqbal, DeepSeek presents an opportunity for research by driving down the costs. 

“In terms of me, and maybe more broadly, it also applies to other researchers. A lot of researchers use the APIs (Application Programming Interfaces) provided by state-of-the-art large language models, which are very expensive to use. So with APIs becoming cheaper, it will be easier to conduct large-scale experiments,” said Iqbal. 

However, there are also concerns regarding the use of DeepSeek and other AI models despite the opportunities they provide. 

“Language models take user input, and that information contains a lot of information, which would also contain sensitive information. If users include their personal information in their prompts, it would mean that the personal information goes to the other side, and a lot of privacy is automatically leaked,” Iqbal said. 

He continued, by explaining the risks. “The other thing is that the user prompts, even if they don’t contain specific personally identifiable information, could also be used to infer things—basically what could be the user’s interest or personal interest—and again, poses serious privacy risks to the users,” Iqbal said.

Due to these political and security concerns, countries like South Korea and Australia have banned DeepSeek on government devices. Other countries like the United States are predicted to follow in their footsteps. 

Unfortunately, in the current age of artificial intelligence , these security risks are unavoidable and will continue to be a concern as AI grows. The most that users can do is try to exercise their rights to privacy on AI, such as disabling functions on AI models to prevent data from being used for training.

The success of DeepSeek signals the development of technology and the ushering of a powerful AI wave. As AI continues to develop, we can only hope that regulations are put in place to protect users as they explore the digital world.

In fact, many firms have already been inspired to develop AI because of DeepSeek. Sources report that, since the success of DeepSeek, many Chinese companies have increased orders for the H20 chip in hopes of creating an AI model of their own. Research and investment in AI has also risen. For example, Alibaba-backed firm Zhipu recently secured over $138 million in funding for its new AI developments, and other smaller companies have come to join the tech race.