DeepSeek: Chinese AI model makes Wall Street tremble

The Chinese start-up DeepSeek is only one year old. But its new AI app shot to number 1 in the Apple App Store over the weekend - and surprises with a performance that apparently even surpasses established models from OpenAI, Meta and Google.

Thus, DeepSeek is challenging US AI supremacy virtually out of nowhere. As a result, the shares of the major tech companies are plummeting. The Nasdaq index fell by up to five per cent at times, which is extremely high for one day. Nvidia even lost up to 18 per cent.

Top performance with a fraction of the effort

Since the start of the AI hype at the beginning of 2022, investors have poured a total of 15 trillion (15,000 billion) US dollars into Nasdaq companies. A large part of the high valuations is based on the hope that AI models will become a lucrative business model in the future. The real profits of Google, Microsoft and co. have also risen - but much less strongly than the share prices. Wall Street is correspondingly sensitive to signs of weakness or unexpected competition.

The latter includes DeepSeek. What shocks analysts most about the new AI model: Development, training and operation are apparently massively cheaper than at market-leading companies such as OpenAI. Put simply, it looks as if the Chinese start-up achieves a comparable performance with a fraction of the effort.

According to the company's own figures, the development of the new AI model DeepSeek R1 cost the equivalent of less than six million US dollars. OpenAI reportedly spent around 100 million US dollars on the development of GPT-4o. Training DeepSeek took just 2.78 million GPU hours. Llama 3.1, the comparable model from Meta, required 30.8 million GPU hours.

But that's not all: the US models require the latest GPUs from Nvidia, which founder and CEO Jensen Huang literally has gold-plated in return. As he is not allowed to export them to China, Nvidia only sells older accelerators there. However, thanks to more efficient algorithms, DeepSeek has apparently managed to train and operate a powerful model with these as well - at a significantly lower price. However, the servers are currently running at their limit due to the unexpectedly high traffic.

Doubts about political neutrality

DeepSeek R1 is open source. OpenAI CEO Sam Altman accuses the Chinese start-up on X of copying ChatGPT: "It's easy to copy something that you know works." However, ChatGPT is not an open source model that could be easily copied. Rather, DeepSeek appears to be copying the architecture, training the model with GPT outputs and building on it.

As with other AI models, there are doubts about DeepSeek's political neutrality. For example, the AI blocks enquiries about Tiananmen Square, Taiwan or Tibet. The bot also refuses to give detailed answers to questions about Chinese President Xi Jinping - while it is able to do so for other political figures.

Update 28/01/2025: DeepSeek is now running stably again and sometimes answers questions on topics that are politically sensitive in China in a seemingly random pattern. However, the bot has quite clear views. Here is an example: