Artificial intelligence More efficient AI "Made in China"? What Deepseek can really do

Related Vendors

Fraunhofer-Institut für Produktionstechnik und Automatisierung IPA

FAULHABER Drive Systems

Wuxi InfiMotion Technology Co., Ltd.

Stock market crash, AI shock, China offensive: The news was overwhelming this week. The cause was a new AI model from China. Is the hype justified?

A bombshell hit the AI industry on Monday this week when a Chinese company released its response to ChatGPT with the AI model Deepseek R1. According to reports, the open-source model achieves results at the level of GPT models but is said to operate much more efficiently and was allegedly developed significantly cheaper. Deepseek reportedly invested only 5 to 6 million dollars. Market reactions were swift. Nvidia's stock temporarily lost up to 17 percent, translating to a market value loss of nearly 600 billion dollars, before partially recovering. Investors apparently feared that cheaper open-source models from China could dominate the industry. However, after the shock, the first doubts about the news surfaced. How realistic are Deepseek's claims? Marco Huber, scientific director of digitalization and AI and head of the research department for AI and machine vision at German Fraunhofer IPA, has looked into the cost and computational efficiency.

Huber points out that the 5 to 6 million dollars in development efforts refer to the pure cloud costs for training. "The actual expenditures, including hardware investments and operating costs, are significantly higher," says the AI expert. Accordingly, the Chinese hedge fund High-Flyer Capital acquired 50,000 Nvidia H100 GPUs valued at around 1.5 billion dollars—even though the export of AI chips to China is sanctioned by the US. Deepseek is a subsidiary of Capital. "Therefore, the cost figures should be viewed with caution and might underestimate the actual resource expenditure," says Huber. The strong market reaction is seen by the scientist as exaggerated: "Nvidia, in particular, is much less challenged by Deepseek, as Nvidia primarily provides the hardware and also the software ecosystem for training such models."

Deepseek is subject to Chinese censorship

On the technical side, the main model Deepseek R1 from the Chinese company seems to be capable of competing in the league of GPT-4.0 or Claude 3.5. Huber sees two main reasons for the model's efficiency: it has been optimized for execution on less powerful GPU hardware. Moreover, R1 is a so-called Mixture of Experts (MoE) model, which consists of several smaller specialized models. "The selective activation of models leads to faster response times," explains Huber. The combination of MoE, reinforcement learning, and model distillation is an innovative approach to AI development that could make powerful models more efficient and accessible. Huber adds, "However, the current hype should be critically questioned. The presented cost advantages could be relativized by unconsidered investments, and the actual performance of the model must still prove itself in practice." Therefore, he does not yet see a revolution in AI development.

Criticism of the model quickly emerged in the media and on social networks. Numerous Deepseek users pointed out instances of censorship. The AI model refuses to answer questions about Tibet, the oppression of the Uyghurs by the Chinese government, or the bloody crackdown on the Tiananmen Square protests in 1989.The German Handelsblatt questioned the model about the relationship between China and Taiwan, to which the chatbot responded with official Chinese propaganda: Taiwan belongs to China, and it is hoped that Taiwan will recognize this soon.Additionally, the German Bayerischer Rundfunk examined the privacy policy and found that Deepseek stores all collected data on Chinese servers. This includes chat histories, user inputs, and uploaded files. Experts therefore recommend hosting the open-source model on a private server. Providers like Perplexity already do this, operating the model on EU and US servers.

Data transmission

Chinese satellite beats Elon Musk's Starlink

For companies developing their own Large Language Models (LLMs), such as OpenAI, Amazon, or Meta, the situation is critical. According to Huber, these companies will now take a closer look at the open-source model and draw their own conclusions for the development of their models. A leaked Google document from spring 2023 already revealed that open-source models pose serious competition. "Consequently, factors such as innovation capability, market position, and adaptability of these companies will be crucial," says the AI expert. Deepseek demonstrates that competition in the AI industry has increased.