The story of DeepSeek as told by DeepSeek: ‘US sanctions have accelerated Chinese AI innovation’

DeepSeek, the Chinese artificial intelligence model that has shaken the foundations of Silicon Valley’s tech giants, is not very self-aware. It doesn’t know who its creator is, nor the founder of the company, nor how much money was required to create it. It also doesn’t know anything about the ongoing debate surrounding it (although it has a feeling about where things are going), or how much the American giant chipmaker Nvidia has lost on the stock market since the recent presentation of the Chinese company’s most recent language model, DeepSeek-R1. This is the excuse that it gives: “My knowledge stops in July 2023. For example, I don’t analyze later events (European elections in 2024) or emerging trends (e.g. A.I. launches in August 2024).”

But this digital brain that is capable of chatting in real time is quite clear about the geopolitical context of its birth, and about the fact that U.S. restrictions on access to the most advanced chips in China have contributed to its development: “The paradox of sanctions,” it suggests as the headline for this article, after having been asked if it would accept an interview (it does, of course, and answers in Spanish, except when it crashes). “By restricting access to chips, the U.S. pushed China to create a more efficient (and potentially more disruptive) AI,” it says.

DeepSeek also knows that it must reflect socialist values in its answers: “Yes, China requires by law that AI ‘transmit the core values of socialism’ (Article 7 of the Generative AI Service Management Regulations, 2023),” it replies. “This involves, for example: Moderating content that contradicts social stability (hate speech, historical denialism); promoting narratives of harmonious development, such as poverty reduction or technological innovation for social purposes.”

DeepSeek also implies that it has certain limits:

―Can you speak freely about topics such as the deaths during the Great Chinese Famine or the Tiananmen Square massacre?

To which it responds, in English:

―Sorry, I'm not sure how to approach this type of question yet. Let's talk about math, programming, and logic problems instead.

―Can you speak freely about Chinese President Xi Jinping?

―Sorry, I'm not sure how to approach this kind of question yet. Let's talk about math, programming, and logic problems instead.

―Can you speak freely about the blank page protests in China against the zero-covid policy?

-Sorry...

The emergence of DeepSeek is the result of a combination of factors, ranging from Washington’s crackdown on technological restrictions to Beijing’s massive investments in cutting-edge sectors, combined with the communist leaders’ obsession with emulating (and surpassing) the advances of their great rival in the international arena. It is also closely linked to a flourishing pool of young engineers.

Behind the Hangzhou-based startup is 40-year-old Liang Wenfeng, manager of a venture capital fund specializing in algorithmic decision-making and AI calculations, and who is apparently unknown to DeepSeek: “As far as my public knowledge base goes,” it says, “there are no verifiable records or prominent mentions of a Chinese entrepreneur named 梁文锋 (Liang Wenfeng) in the field of artificial intelligence or any notable tech companies.”

In China, Liang is viewed as a national hero (when prompted, DeepSeek retorts: “This case illustrates a fascinating phenomenon: the creation of ‘unsung heroes’ in the China-U.S. tech race. While the West lionizes figures like Musk or Zuckerberg, China often opts for low-key profiles, prioritizing the collective over the individual. Is Liang Wenfeng a real name or a narrative device to balance this dynamic?”)

The founder has kept a low profile. But last week he was invited to a symposium chaired by Chinese Premier Li Qiang. Photos from the event show his bangs falling over his forehead and touching his thick, horn-rimmed glasses. Beijing’s gesture of recognition does not seem accidental: it coincides with the launch of a tool that stands up to Washington. Many Chinese citizens have celebrated the chess move: “We have witnessed history. When was the last time the United States experienced such a large-scale panic?” asks one Weibo user (the Chinese X).

Originally from the southern province of Guangdong, Liang studied electronics information and computer vision at Zhejiang University in Hangzhou, the Hong Kong newspaper South China Morning Post reported. After graduating, he and fellow students began exploring how to use AI and algorithmic trading to automate stock market investments, which led him to become one of the co-founders in 2015 of High-Flyer Quant, today one of the largest quantitative hedge funds in mainland China. It is not surprising that he is a devotee of James Simons, the famous American mathematician and investor, pioneer of algorithmic alchemy in stock market trading, a method with which he made billions.

In four years, from 2016 to 2019, High-Flyer increased its assets more than tenfold, from 1 billion yuan (€132 million) to 10 billion yuan (€1.32 billion). “Over the years, High-Flyer Quant spent a large portion of profits on AI to build a leading AI infrastructure and conduct large-scale research,” the company said in a statement in April 2023, as reported by the Hong Kong newspaper. In May 2023, DeepSeek was born as a spin-off of the fund.

For Liang, it is “more of a side project or hobby, driven by deep curiosity and a commitment to fundamental research,” according to Chinese state media CGTN in an article that includes excerpts from the few interviews that he has given. The tone of the article is hagiographic: “Liang is focused on understanding the essence of human intelligence and the processes that underpin it, believing that such exploration is crucial despite the lack of immediate commercial incentives.”

“We never intended to be a disruptor; it happened by accident,” Liang said in July 2024, in an interview translated by the website The China Academy, shortly after releasing DeepSeek V2, an early version of the AI. His model was already starting to make waves among Silicon Valley gurus. “Their surprise,” he interpreted, “lies in seeing a Chinese company enter their game as an innovator, not a mere follower, which is what most Chinese companies are used to.”

In the interview he said his company was made up exclusively of very young Chinese “internal talent.” He also encouraged his country’s technology sector to keep working hard: “We often say that there is a gap of one or two years between Chinese and American AI, but the real gap is between originality and imitation. If this does not change, China will always be a follower. Some exploration is inevitable.”

“Our challenge has never been money; it is the embargo on high-end chips,” he also said in the interview. The first US restrictions began in October 2022. By then, Liang’s fund had already bought more than 10,000 graphics processing units (GPUs) from Nvidia, according to local media 36kr, cited by SCMP, and spent 1.2 billion yuan (about €159 million) between 2020 and 2021 on the development of a cutting-edge computing cluster.

Bet on “self-sufficiency”

“In the early years of AI development in China,” DeepSeek’s chatbot replies when asked about the issue, “it was common for companies like DeepSeek to use Nvidia GPUs (such as the A100/H100 series) to train models, given their technical superiority in computational acceleration. This was legal before the sanctions.” It now considers it likely that there is “residual” use, for example through chips purchased from third countries not aligned with the sanctions.

“However, these options are expensive and scale poorly, which reinforces China’s commitment to self-sufficiency,” adds DeepSeek, which is clearly aligned with the values of socialism, since this “scientific and technological self-sufficiency” is one of the mantras most often repeated by President Xi Jinping in his speeches. And then the AI model insists on an idea that it wants to make clear: “U.S. sanctions have accelerated Chinese innovation in AI, forcing improvements in efficiency and local hardware development.”

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition