The Chinese startup DeepSeek has unveiled a new AI system, DeepSeek-V3, demonstrating capabilities comparable to leading chatbots from OpenAI and Google. This achievement is significant in itself, but even more notable is DeepSeek’s efficient use of computing resources. According to a report in NewYork Times, DeepSeek’s engineers reported using significantly fewer specialized computer chips, crucial components in AI training, than their American counterparts.These chips are central to the technological competition between the U.S. and China, with the U.S. imposing export restrictions to maintain its AI leadership. However, DeepSeek’s success highlights potential unintended consequences of these restrictions, as Chinese researchers have been driven to innovate with readily available resources.
DeepSeek-V3 reportedly performed on par with existing market offerings in benchmark tests assessing question answering, logic problem-solving, and code generation. Remarkably, DeepSeek achieved this with approximately $6 million in computing power, a fraction of the estimated $60 million spent by Meta on its latest AI technology. This lower cost of entry broadens the field of potential AI developers, as noted by investor Chris V. Nicholson, who pointed out the significant difference between having $6 million versus $100 million or $1 billion available.
While leading AI companies often utilize supercomputers with 16,000 or more chips, DeepSeek reportedly used only around 2,000 Nvidia chips. This efficient use of resources was attributed to optimized training methods necessitated by chip constraints in China.
Despite U.S. efforts to restrict China’s access to advanced chips, including recent regulations aimed at preventing acquisition through third countries, DeepSeek is said to leverage existing resources. The company, backed by quantitative trading firm High Flyer, had previously acquired Nvidia chips and attracted talent with competitive salaries and research opportunities. They even hire individuals without computer science backgrounds to enhance the AI’s understanding of diverse subjects like poetry and performance on challenging exams.
While U.S. companies continue to innovate, with OpenAI’s new “reasoning” system being a prime example, DeepSeek’s rapid progress, including its own impressive reasoning model, demonstrates its competitiveness, as per the report.
A key aspect of this evolving landscape is open-source software. DeepSeek has open-sourced its latest AI system, sharing the code with others. This fosters global collaboration, as noted by Yineng Zhang, a software engineer at Baseten. This open-source approach gained momentum in 2023 with Meta’s release of LLama. DeepSeek’s contribution demonstrates that smaller players can also contribute significantly to the open-source AI ecosystem.