Experts suggest this could make AI technology more accessible, particularly within the European Union.
Last week, the Chinese company DeepSeek released a free language model that it claims is far cheaper to operate than competing AI systems.
The announcement caused stock prices to drop for major American tech companies involved in AI development, including Nvidia, Meta, Microsoft, Tesla and Apple.
'DeepSeek has potential to reshape global AI market': Polish expert
According to Piotr Sankowski, a professor at the University of Warsaw, American AI models such as those developed by OpenAI and Grok are closed-source, meaning they cannot be run on private computers.
In contrast, DeepSeek is an open-source model, allowing companies and researchers to modify and customize it for their needs.
Similar open models exist, such as MetaAI’s LLama, but they are not available within the European Union.
"DeepSeek has the potential to reshape the global AI market, including in the EU," Sankowski said in comments to Polish state news agency PAP. "The drastic reduction in costs will make AI solutions more widely accessible."
'Efficient memory management'
Another Polish expert, Prof. Tomasz Trzciński, echoed this view, emphasizing that open-source AI could accelerate adoption across various industries, particularly small and medium-sized enterprises.
DeepSeek’s efficiency stems from its smaller scale compared to AI models from OpenAI, Google and Meta.
"This is due to a range of innovations, including reduced numerical precision and efficient memory management," Trzciński explained. These advancements significantly lower both training and operational costs.
China claims that DeepSeek is about 25 times cheaper than competing models. However, technology journalist Sylwia Czubkowska cautioned that DeepSeek has received significant financial backing.
The High-Flyer fund, led by founder Liang Wenfeng, financed two supercomputers—Firefly I and Firefly II—at a combined cost of USD 174 million to develop the model, she said. High-Flyer has also been acquiring Nvidia chips since 2015, according to Czubkowska.
Trzciński noted that while China reports a massive cost advantage, the figures only reflect the final training phase and do not account for earlier experimental costs. Despite this, he acknowledged that DeepSeek delivers significant savings compared to other models.
Last Wednesday, OpenAI released a statement claiming it had evidence that DeepSeek was trained on its ChatGPT model, a practice that would violate OpenAI’s policies.
Trzciński commented that this claim is plausible, explaining that AI models can be "distilled"—a process where one AI system is trained using responses from another to enhance performance. However, proving this is challenging due to the complexity of training processes, he said.
That same day, the American cybersecurity firm Wiz discovered an unsecured database linked to DeepSeek containing chat histories. The database disappeared from the internet an hour after being reported.
Following this incident, Euroconsumers, a coalition of European consumer protection organizations, filed a complaint with Italy’s data protection authority, questioning whether DeepSeek complies with the European Union’s General Data Protection Regulation (GDPR).
In response, Italy blocked access to DeepSeek, and both the Italian and Irish data protection authorities have demanded explanations from the company regarding its data processing practices.
The US National Security Council is also reviewing DeepSeek for potential national security risks, the White House confirmed.
Legal expert warns of privacy risks
Legal expert Paula Skrzypecka from Creativa Legal warned that DeepSeek poses privacy risks.
"The service operates in China, where there are no guarantees of privacy protection or safeguards against government surveillance," she told the PAP news agency.
However, she noted that using DeepSeek locally on private servers would provide more security.
Sankowski said that privacy concerns apply to most AI services and that only paid versions of models guarantee that user data is not repurposed.
He emphasized that since DeepSeek is open-source, it can be run on private infrastructure, ensuring full control over data processing.
Built-in censorship
Another issue surrounding DeepSeek is its built-in censorship. Users have reported that the model avoids answering questions about sensitive topics, such as Chinese government actions and the status of Taiwan.
"Ironically, China has developed a more open AI model than the United States, but it still carries ideological constraints," Trzciński observed. "If you ask about events like the Tiananmen Square protests, DeepSeek systematically avoids the topic."
However, Sankowski pointed out that American AI models also impose restrictions on certain types of content, such as jokes on controversial topics.
"Both Chinese and American AI models apply constraints; the difference lies in the nature of the censorship," he said.\
'Poland must invest more in AI'
Sankowski stressed the need for Poland to strengthen its position in the global AI race.
"Poland must invest more in AI development and find its own niche; our strongest advantage is our highly skilled IT professionals, who are among the best in the world," he said.
(rt/gs)
Source: PAP