Investors in tech companies do not know where to look for it since the introduction of Deepseek, an innovative AI model from China. Last week the latest version of Deepseek was released. In some tasks, this chatbot is just as good as the American market leader OpenAi (from Chatgpt), but achieves that level of artificial intelligence for a fraction of the costs and with significantly less computing power.
This performance causes doubts about the potential revenue model of many American AI companies, but also about the effectiveness of American export restrictions on AI chips and advanced chip machines. Despite – or perhaps thanks to – those restrictions, Chinese researchers have a good one workaround found to develop AI models.
Investors responded immediately to the international stock markets: the price of NVIDIA, market leader in AI chips and with a market value of around $ 3,500 billion, fell by 12 percent on the market for opening the stock market. Tech giants Google and Meta and other chip companies also fell in value; The Dutch chip machine manufacturer ASML started the day with 9 percent price loss, ASM International and Besi fell by 11 percent.
The American OpenAi released an AI model in December that can ‘reason’-an improvement in the usual approach to chatbots that produce answers by predicting the next word. A step -by -step, more logical approach helps to solve mathematical and scientific problems, or program computer code. That requires a lot of calculation work from very expensive chips from Nvidia.
That can be more efficient, Deepseek proves. According to the Chinese start-up itself, the chatbot cost only $ 6 million. Deepseek uses similar technology – so -called language models that recognize patterns in large quantities of data – but in a different way. It cuts down and distributes it over several ‘expert systems’. This approach saves computing power, according to experts. Deepseek uses open source models, such as Llama from Meta. According to Yann LecunDie Meta’s AI branch does not show Deepseek that China is better in AI than the US, but that open research and open sources lead to faster innovation than closed models, such as those of OpenAi. The Deepseek software is also freely accessible.
Most popular downloads
Deepseek, founded in 2023, is a start-up that is led by researcher Liang Wenfeng. Chinese media see the success of Deepseek as a big PR success. In an interview with the pro-Chinese The China Academy Liang already stated that “Silicon Valley was surprised that a Chinese company plays their game, not like a follower but as a innovator.”
OpenAi, one of the pioneers in artificial intelligence, is now appreciated at $ 157 billion thanks to the fame of Chatgpt. But the lead with regard to competitors is shrinking, and deepseeks progress accelerates that process. The Chinese app is now at the top of the list of most popular downloads.
But Deepseek remains a Chinese product. If you ask the Deepseek-Chatbot or China a dictatorship, or to come up with a way to bypass the Chinese firewall, the answer is: “Let’s talk about something else.” This shows the announcement that Deepseeks models are aimed at the Chinese market, “as a result of which English/Dutch -language documentation can be limited”.
Export restrictions
DeepSeek used a AI chips cluster from Nvidia for his research that were purchased by Hedgefonds Highflyer-an investment vehicle from the founder himself-even before America tightened the export restrictions. Chinese companies have access to significantly less powerful AI chips, because the US is restricting the delivery of Nvidia and AMD.
Chinese chip companies cannot build the most advanced AI chips themselves, because they do not have access to the most modern chip machines from ASML.
The US tries to keep their computing power and AI data centers on their own soil as much as possible. The newly departed Biden government set up a restriction last month for the amount of chips that can be sold abroad.
Last week President Trump announced a major investment in American data centers. That so -called project Stargate is decorated around OpenAI, with the help of companies such as Oracle and investors Softbank (Japan) and MGX (United Arab Emirates). It is uncertain whether the conglomerate can ever cough up the astronomical amount – somewhere $ 100 and $ 500. Despite the large amounts and the big words, Stargate is primarily a political statement of Trump, a smartly timed repackaging of previous announcements for large AI data centers. It will become more difficult to collect the required investments now that it is clear that the competition for OpenAi not only comes from the US – from companies such as Google, Meta or Anthropic – but also from China, and for significantly less money.