A chatbot developed by the Chinese artificial intelligence startup DeepSeek has surged to the top of Apple’s App Store charts in the US this week, overtaking OpenAI’s ChatGPT as the most downloaded free app. This innovative AI assistant utilizes DeepSeek’s open-source models, which the company claims can be trained at a significantly lower cost using fewer chips compared to the leading AI models in the industry. As a result, shares for Nvidia, the largest global supplier of advanced AI chips, have dropped over 12 percent in pre-market trading.
The app’s downloads skyrocketed shortly after DeepSeek introduced its new R1 reasoning model on January 20th, designed to tackle complex problems and reportedly perform comparably to OpenAI’s o1 on various benchmarks. The R1 model was built on the V3 LLM released by DeepSeek in December, which the company asserts is comparable to GPT-4o and Anthropic’s Claude 3.5 Sonnet models, costing under $6 million to develop. In stark contrast, OpenAI CEO Sam Altman has disclosed that GPT-4’s training exceeded $100 million.
DeepSeek also claims it required only around 2,000 specialized Nvidia chips to train the V3 model, compared to the 16,000 or more chips needed for leading models, as reported by the New York Times. These unverified claims have prompted developers and investors to reconsider the compute-intensive strategies typically employed by major AI companies. If these assertions hold true, it indicates that DeepSeek’s engineers had to innovate amid trade restrictions aimed at ensuring US dominance in the AI sector.
Companies like Nvidia, Microsoft, OpenAI, and Meta are pouring billions into AI data centers, with $500 billion allocated for the Stargate Project, of which $100 billion is believed to be designated for Nvidia. However, investors and analysts are now questioning whether such investments are prudent, especially as Nvidia, Microsoft, and other companies with significant stakes in preserving the AI status quo are trending downward in pre-market trading.
