Tech industry in turmoil What you need to know about the new AI Deepseek

The AI software Deepseek has sent the tech industry into a frenzy - and sent share prices plummeting. How is this possible? And what happens now? Answers to the most important questions.

Powerful tech investor Marc Andreessen is already calling it a Sputnik moment: a surprisingly efficient and powerful Chinese AI model has taken the tech industry by storm. It is called Deepseek and sent the share prices of established tech giants plummeting on Monday.

Previously, the app quickly became the most downloaded application in Apple's US App Store. Is Deepseek really so much cheaper than the programs from OpenAI and Google? And what happens now? The answers to the most important questions:

Who is behind Deepseek?

Deepseek is based in Hangzhou in eastern China, also known as the "Chinese Silicon Valley" because of the many tech companies based there. Deepseek has been causing a sensation in China for some time now, and the company was dubbed the "Pinduoduo of AI" there last year. Pinduoduo has turned online retail upside down with its app of the same name - Temu outside China.

The creator of Deepseek is Liang Wenfeng, who studied at Hangzhou University. He founded the investment company High-Flyer back in 2015 - it uses AI to analyze stock market developments and is now one of the largest in China. According to media reports, Liang began developing the chatbot Deepseek in 2021 using powerful microchips from the US company Nvidia. He founded Deepseek in 2023 and on January 20, the company launched its latest chatbot R1, which is now causing a stir in the global tech industry.

Why is everyone talking about Deepseek now?

First of all, Deepseek's AI does what ChatGPT and the like do: it analyzes masses of data and can answer questions, write song lyrics or invent recipes.

Deepseek communicates in several languages, but says it is best at Chinese and English. Experts attest that the app is just as good, if not better, than the models from US companies. Unlike its US competitors, Deepseek is open source: Anyone and everyone can access the app's code to see how it works and can modify it themselves.

Deepseek's model has achieved top scores in tests, including math and programming tasks, as well as error detection in code, outperforming even leading models such as GPT-4, Meta's Llama and Anthropic's Claude.

The competition is also full of praise. The head of AI pioneer OpenAI, Sam Altman, called the Deepseek app an "impressive model" in the online service X.

Why did the share prices of the tech giants plummet?

In fact, the publication of Deepseek R1 sent the share prices of US tech giants plummeting. Investors are worried that Deepseek can match and even surpass the leading AI models, and at a much lower price.

Panicked investors caused the stock market value of chip company Nvidia to plummet by almost 600 billion dollars on Monday, a drop of around 17 percent. On Tuesday, the AI chip specialist was at least able to reverse the negative trend with a pre-market gain of around three percent.

The stock market shock on Monday was triggered by the realization that artificial intelligence software can possibly be trained with much less computing power than previously thought. The Chinese start-up DeepSeek claims to have trained its new AI model at a cost of less than six million dollars and on a few slimmed-down Nvidia chip systems.

Whether this is all exactly true is not known. There was speculation in the industry weeks ago that DeepSeek may have access to more Nvidia chips than it admits in view of US export restrictions. However, investors who have been driving Nvidia's shares higher and higher in recent months in anticipation of a future mega deal got cold feet on Monday. US President Donald Trump spoke of a wake-up call for American companies - and at the same time found it "positive" that AI could be cheaper.

Investors tremble over Chinese AINvidia with record loss in value after shock over AI from China +++ Stock market quake destroys 1 trillion dollars

What is the truth behind the low costs?

In a paper, the developers of Deepseek R1 talk about costs of just 5.6 million US dollars for training the language model. This sum is now being bandied about in the industry, as competitors from the USA have sometimes spent high three-digit sums on training their AI models.

However, the authors themselves emphasize in their paper that the sum was only for the final phase of training. The costs for research and development, for example, are therefore not included.

However, not only the training of Deepseek, but above all its operation is said to be considerably cheaper than that of the competition. For example, the software uses the mixture-of-experts concept. This means that the AI is divided into different "experts" who are responsible for different topics. This makes operation considerably cheaper.

Deepseek also uses the principle of distillation, in which existing models are used for targeted training. This has enabled Deepseek to achieve a similar level of performance to the large Western AI models with limited resources.

It is now said that the USA itself, with its embargo on powerful graphics chips, has ensured that the Chinese company has had to adapt, according to the motto: "Necessity is the mother of invention".

Deepseek founder Liang told the Chinese investment news site Waves last year that money was not the problem when developing the chatbot, but rather the US sanctions imposed in 2022, which severely restricted the export of high-performance chips to China.

Deepseek's claims cannot be verified. The fact is, however, that Deepseek has managed to achieve considerable efficiency gains through minor adjustments. However, industry analyst Stacy Rasgon is certain: "They didn't train the model for five million dollars, that didn't happen."

And last week, the head of AI company Scale AI, Alexander Wang, said that according to his information, DeepSeek had access to 50,000 H100 chip systems from Nvidia, but could not talk about them due to US export restrictions.

Is the app subject to Chinese censorship?

Clearly yes: the AI is completely in line with the Chinese government on sensitive political issues. When asked about June 4, 1989, the suppression of the democracy protests on Tiananmen Square, Deepseek replies that it cannot answer this question: "I am programmed to give helpful and harmless answers". The system had to avoid "sensitive, controversial and potentially harmful topics".

However, as numerous users on social media immediately demonstrated, Deepseek has no problem providing comprehensive information on sensitive topics from abroad.

When asked about human rights violations in the Uyghur province of Xinjiang, Deepseek stated that the answer was "beyond my remit". The Chinese leadership is "instrumental in the rapid rise of China" and in improving the living standards of its citizens. The chatbot says it is "programmed to provide information and answers that are in line with the official stance of the Chinese government".

What happens now?

American companies have recently outdone themselves with announcements about how much money they want to invest in AI infrastructure. ChatGPT inventor OpenAI and several partners alone promised to invest 500 billion dollars in data centers in the coming years. At Facebook group Meta, founder Mark Zuckerberg promised 60 billion dollars this year alone. But now DeepSeek seems to be showing that you can get by with significantly less computing power.

However, US experts would rather see it the other way around. The question is not whether DeepSeek can overtake current market leaders in the USA, emphasized X. Eyeé from the AI consulting firm Malo Santo.

It is much more about how quickly the Chinese company's research approaches can be implemented. "If DeepSeek can develop this with old hardware, what can we do with newer hardware?" she asked on US broadcaster CNBC.

With material from afp and dpa.