On the day following Christmas, a modest Chinese startup named DeepSeek introduced a new artificial intelligence system capable of rivaling the advanced chatbots developed by industry giants like OpenAI and Google.
Achieving such a feat alone would have been a significant milestone. However, the team behind the system, named DeepSeek-V3, claimed to have taken an even more remarkable step. In a research paper detailing their approach, DeepSeek's engineers revealed that they trained their AI using only a fraction of the specialized computer chips that major AI firms typically depend on. These chips are at the heart of an ongoing technological rivalry between the United States and China. As part of its efforts to stay ahead in the global AI race, the U.S. government has been imposing restrictions on the export of high-performance chips—such as those produced by Silicon Valley giant Nvidia—to China and other competing nations.
Yet, the impressive capabilities of the DeepSeek model have sparked concerns about the unintended effects of these trade restrictions. With limited access to cutting-edge hardware, Chinese researchers have turned to a variety of widely accessible tools available online to achieve their goals. Benchmark tests—commonly used by leading U.S. AI firms—show that the DeepSeek chatbot performed at par with existing market solutions, answering questions, tackling logical challenges, and even generating its own computer programs with impressive accuracy.
What sets DeepSeek apart is its cost-effective development process, challenging the assumption that only the largest U.S.-based tech companies can afford to create state-of-the-art AI systems. According to the Chinese engineers, their system required just $6 million in computing power—roughly one-tenth of what tech giant Meta spent on its latest AI project. "There are far more companies that can manage a $6 million investment compared to those that can allocate $100 million or even $1 billion," said Chris V. Nicholson, an AI-focused investor at the venture capital firm Page One Ventures.
Since OpenAI ignited the AI revolution in 2022 with the launch of ChatGPT, many experts and investors believed that competing with industry leaders would require investing hundreds of millions of dollars in specialized chips. Leading AI firms train their models using supercomputers equipped with up to 16,000 or more high-performance chips. In contrast, DeepSeek's engineers claimed they achieved similar results using only around 2,000 Nvidia chips. Facing chip shortages in China, DeepSeek's team had to optimize their training process to remain competitive, explained Jeffrey Ding, an assistant professor at George Washington University specializing in emerging technology and international relations. Earlier this month, the Biden administration introduced stricter regulations to prevent China from acquiring advanced AI chips through indirect channels. These new rules build on previous restrictions aimed at limiting Chinese companies' ability to purchase or manufacture state-of-the-art semiconductors. However, former President Donald Trump has yet to clarify whether he will uphold or repeal these measures.
The U.S. government has imposed these restrictions due to concerns that advanced chips could be utilized for military purposes. In response, some Chinese companies have stockpiled thousands of chips, while others have turned to underground markets to secure the necessary hardware. DeepSeek operates under the umbrella of High Flyer, a quantitative stock trading firm that, by 2021, had invested its profits in acquiring thousands of Nvidia chips to train its earlier AI models. Despite requests for comment, the company remained silent. In China, DeepSeek is known for attracting top-tier talent from prestigious universities by offering lucrative salaries and the freedom to explore their own research interests. Zihan Wang, a computer engineer previously involved with DeepSeek, revealed that the company also hires individuals from non-technical backgrounds to enhance the AI's capabilities in areas such as poetry and answering complex questions from China's notoriously challenging college entrance exams.
Unlike many AI companies, DeepSeek does not develop consumer products. Instead, it focuses exclusively on research, allowing it to operate outside the strict regulatory framework governing consumer-facing AI technologies in China. Meanwhile, top U.S. AI companies continue to push the boundaries of innovation. In December, OpenAI introduced "o3," a sophisticated reasoning system that surpasses existing models, though it remains largely unavailable to the public. Nevertheless, DeepSeek remains a formidable contender, recently unveiling its own advanced reasoning model. Amid the evolving AI landscape, an old concept has regained prominence: open-source software. Like many competitors, DeepSeek has open-sourced its latest AI system, making its underlying code available to businesses and researchers. This approach enables others to build their own products using DeepSeek's technology.
Chinese developers, unlike their counterparts at major domestic tech firms who are limited to internal collaborations, benefit from contributing to open-source projects and working with global talent. "If you work on open source, you collaborate with talent worldwide," said Yineng Zhang, lead software engineer at Baseten in San Francisco, who assists in developing applications based on DeepSeek's system. The open-source AI movement gained traction in 2023 when Meta released its AI model, LLaMA, to the public. Initially, many believed such initiatives would only thrive with the backing of major tech corporations equipped with vast data centers and specialized chips. However, DeepSeek has proven that smaller companies can also drive open-source innovation. Despite this, some industry leaders and policymakers argue against open-sourcing AI technologies, warning that they could facilitate the spread of misinformation or other harmful activities. U.S. lawmakers have even considered measures to curb the practice.
On the other hand, some experts caution that limiting open-source progress in the U.S. could give China a strategic advantage. If Chinese firms lead the open-source AI field, U.S. developers may increasingly rely on their technologies, potentially shifting the epicenter of AI innovation to China. "The center of gravity in the open-source community is shifting toward China," noted Ion Stoica, a computer science professor at UC Berkeley. "This could pose a significant risk for the U.S.," as it may accelerate China's technological advancements. Shortly after taking office, Trump rescinded a Biden-era executive order that sought to restrict open-source AI technologies. Stoica and his students recently developed an AI model called Sky-T1, which performed comparably to OpenAI's o1 model in specific benchmark tests, using only $450 in computing resources. Their achievement was possible thanks to open-source tools provided by Chinese tech giant Alibaba. Although their budget-friendly system cannot yet rival the capabilities of DeepSeek or OpenAI's cutting-edge models, the project demonstrated that even with minimal resources, competitive AI models can be developed.
Reuven Cohen, a technology consultant from Toronto, has been using DeepSeek-V3 since late December. He finds it on par with leading AI systems from OpenAI, Google, and Anthropic—while being significantly more cost-effective. "DeepSeek helps me save money," Cohen said. "This is exactly the kind of technology I need."
