What can Europe learn from the success of DeepSeek, the new Chinese AI?

The emergence of DeepSeek, the new low-cost Chinese generative AI, has shaken Western technology markets. Nvidia, the leading graphics processor (GPU) company, plummeted 17% in a single day. Such a significant stock market loss (nearly $600 billion) had never been seen before. 

In recent years, Nvidia has become the world's most valuable publicly traded company. Its growth has been driven by the high demand for GPUs required for developing generative AI models—a technological wave that has also boosted sales for microprocessor manufacturers. 

With limited resources, China has managed to develop models that compete with the best US ones

However, it seems that DeepSeek has arrived to change the paradigm. “With limited resources—around 10,000 GPUs and an estimated cost of $5 million—they have managed to develop models that compete with the best US models,” explains Esade profesor Esteve Almirall, an expert in innovation and artificial intelligence. 

By comparison, US firms have been spending anywhere from $100 million to $1 billion to develop their large language models (LLMs). Moreover, DeepSeek’s emergence comes just a week after President Trump's announcement to invest $500 billion in AI infrastructure in the U.S. 

Overall, the DeepSeek-R1 model offers a product as powerful as ChatGPT-o1 but for free, at a much lower cost, and with open-source code, allowing anyone to replicate it. DeepSeek's development team has also published a technical paper explaining how the model works. 

Innovation driven by scarcity

Despite surprising many and shanking markets, Almirall clarifies that “just like with ChatGPT, DeepSeek is not a new phenomenon or something that emerged out of nowhere.” For more than a year, specialized circles have been closely following it, as well as models from Alibaba (Qwen 2.5-VL) and 01.ai (Yi-34B), which have been performing at levels very close to their US counterparts. 

Regulation can also have a positive impact, and China is a good example

DeepSeek's success results from the industrial policy strategy of the Chinese government. “The Chinese government set the goal of leading certain strategic industries, a purpose that materialized with the Made in China 2025 program in 2015 and, more recently, with various AI-related initiatives,” Almirall explains. These policies have yielded results, allowing China to lead in critical sectors such as batteries, electric cars, and now AI. 

Moreover, two notable factors stand out. One is the innovation driven by the restrictions Chinese engineers face. “They have innovated in resource usage, among other things, because they don’t have them,” the professor explains. In recent years, the US has restricted the sale of high-end graphics cards and microchips to its geopolitical rival. As a result, Chinese engineers have been forced to explore new ways to achieve the same results while starting at a disadvantage. 

The role of regulation in DeepSeek

Another important aspect is regulation. “Regulation is often associated with limiting freedom, and only its negative aspects are emphasized. However, regulation can also have a positive impact, and China is a good example of that,” says Almirall. 

China has implemented specific regulations to equip all universities with thousands of GPUs, promote projects to improve research quality, create programs to repatriate key figures, and establish collaborations with major companies like Alibaba, Tencent, or Huawei. 

Europe's regulatory strategy has a limited scope

On the other hand, China's regulation has also been restrictive, as all Chinese AI models must align with the official ideology of the Communist Party. DeepSeek users have noticed that the model avoids discussing sensitive topics for the regime, such as the Tiananmen protests, Taiwan, or the figure of Xi Jinping, although censorship is not always infallible. 

DeepSeek lessons for Europe

According to Almirall, this new chapter in the development of large language models—"with cheaper, more powerful, smaller, and simpler models"—offers Europe an opportunity to play a relevant role. DeepSeek-R1 was developed with about $5 million and only two months of training, involving a young team of fewer than 100 people. “This demonstrates that it is perfectly feasible for Europe,” he suggests. 

However, he also warns that to achieve this, innovation policies that create the necessary ecosystem are required. “In Europe, regulation seems mainly focused on banning and censoring models from other countries, a strategy with limited scope,” Almirall continues. 

Instead, the goal should be to align markets with the common interest. “It’s about designing markets where the incentives of all participants are aligned with shared objectives, avoiding extractive behaviors,” he points out. 

He adds that, in addition to regulation, other instruments such as standards, behavior promotion, projects, resource availability, or direct state participation (through public or public-private companies) can help achieve more efficient markets. 

After all, the earthquake triggered by DeepSeek might bring good news for a Europe that does not want to lag behind in the artificial intelligence race.