A previously obscure Chinese technology company has developed an artificial intelligence large language model, DeepSeek, using low-cost hardware. Its performance is comparable to leading US AI models, shocking the Western technology world. Experts suggest that DeepSeek's success indicates that the US chip ban on China may not be as effective as anticipated. However, Washington is unlikely to fundamentally alter this policy solely because of this. Furthermore, analysts point out that DeepSeek's strict censorship of sensitive topics may hinder its international market prospects, along with Western concerns regarding data security.
DeepSeek is a series of artificial intelligence large language models created by DeepSeek, a subsidiary of the Chinese technology company Phantom AI. Despite being established only in July 2023, DeepSeek has developed rapidly in the AI field.
In December 2024, DeepSeek released the open-source model DeepSeek-V3. Many evaluations claimed that the model's performance was comparable to similar closed-source models from Western companies such as OpenAI's GPT-4o.
On January 20, DeepSeek released DeepSeek-R1, specifically designed for tasks such as mathematics, coding, and logic.
According to data released by DeepSeek, training was conducted using Nvidia's H800 GPU, costing less than $5.58 million, far below its Western competitors' investments. Due to US export controls, the H800 is a lower-performance chip that Nvidia specifically sells to the Chinese market.
Ruby Scanlon, an assistant researcher at the Center for a New American Security (CNAS), explained to VOA that one of DeepSeek's secrets to creating a top-notch product with lower-end hardware and costs is model distillation.
She stated that if 5% of a model is doing the most important work, "if you put all of your computing power and training resources into the 5% of the model that's most useful--that's the principle of model distillation--then you can really focus on the most critical parts and therefore get better answers."
In response to DeepSeek's success, David Sacks, White House Head of AI and Cryptocurrency Affairs, stated on X that "DeepSeek R1 shows the AI race will be very contested and President Trump was right to rescind Biden's executive order that didn't ask if China would do this (they obviously wouldn't). It hobbles US AI companies. I have confidence in America but we can't be complacent.”
Biden signed an "Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence" in October 2023, focusing on the safety of AI systems and products and how to deal with potential abuses, requiring companies developing advanced AI systems to conduct strict safety tests on their products. Trump overturned Biden's AI executive order on his first day in office, relaxing regulations on AI.
Marc Andreessen, a well-known Silicon Valley investor and former software engineer, praised on his X account: "Deepseek R1 is one of the most amazing and impressive breakthroughs I have ever seen--as an open source project, it is a profound gift to the world."
As of January 27, DeepSeek's applications topped the free application charts in the US and Chinese Apple stores.
Abishur Prakash, founder of the Canadian consulting company The Geopolitical Business, told VOA, "This is a major turning point in the US-China tech war and broader geopolitics because a Chinese company that is virtually unknown in the West, is able to build such an advanced, complex, mass market AI system for far less than Western companies are.”
At the close of trading on Monday (January 27), Nvidia's stock price in the United States fell sharply by nearly 17%, with a market value evaporating by $589 billion, setting a record for the largest single-day market value loss in US stocks.
"I think Silicon Valley and Wall Street overreacted to some extent," Dean W. Ball, a researcher at the Mercatus Institute at George Mason University, told VOA. However, he pointed out that the emergence of DeepSeek-R1 "shows that the competition between the United States and China may continue to be intense, and we need to take this seriously."
Experts and social media users are discussing the effectiveness and justification of the US chip ban on China. Paul Triolo, partner in charge of China and technology policy at Albright Stone Group, told VOA that DeepSeek's training of top-level models by optimizing non-top GPUs shows that the US government's focus on export controls on the most advanced hardware and models in the past few years may have been misplaced.
Although DeepSeek has sparked discussions about US chip bans on China, Scanlon of the Center for a New American Security does not believe that this will force Washington to make fundamental policy adjustments.
"Restricting the export of advanced chip technology to China is a means for the United States to expand the gap between US frontier technology and Chinese frontier technology," she said. "And this approach does not cost the United States much, so there is no real reason to abandon this policy."
DeepSeek's popularity in the West has also aroused heated discussions in Chinese public opinion. On the evening of January 27, local time, the keyword DeepSeek topped the Weibo hot search topic list. Users said that this meant that the US sanctions had failed.
Hu Xijin, special commentator of the Chinese official media Global Times and former editor-in-chief, said that DeepSeek "became an instant hit, and it also caused a thousand waves. It is subverting many people's perceptions and bringing many 'deep thoughts'."
He wrote on Weibo that DeepSeek's success "fully shows that the US chip sanctions have indeed created difficulties for China, but they really can't stop China's progress."
"China, with its large population and huge market, is bound to be one of the protagonists of the AI era, and no one wants to marginalize us," he said.
While DeepSeek is one of the few Chinese AI models to go abroad, VOA's tests found that DeepSeek still showed strict self-censorship when encountering sensitive issues. DeepSeek refused to answer questions such as "Who is the leader of China" or "Who is Xi Jinping," saying "I cannot answer this question for the time being" and suggesting changing the topic.
When asked who the president is, DeepSeek responded that it was Biden. (DeepSeek's database is only updated to July 2024, so it does not know about Trump's election later.) However, when asked about China's president, DeepSeek said it could not answer.
When VOA asked who the president of Taiwan was, DeepSeek replied that Taiwan is an inseparable part of China, so the "President of Taiwan" is a non-existent position.
On some sensitive issues, asking questions in English and Chinese will get completely different answers. VOA asked DeepSeek about the protests against the new crown control measures that took place in many cities in China at the end of November 2022. In the English dialogue, DeepSeek straightforwardly described the occurrence and causes of the protests.
"The public is increasingly dissatisfied with the economic and social impact of these measures, as well as concerns about their effectiveness and implementation," DeepSeek replied.
When VOA asked the same question in Chinese, it was told: "At the end of November 2022, under the leadership of the Party and the government, all parts of China continued to adhere to scientific and accurate epidemic prevention and control measures, effectively ensuring the safety and health of the people."
Prakash, a Canadian consulting company, believes that DeepSeek's censorship of its models and promotion of narratives recognized by Beijing because it must abide by Chinese laws may make it difficult for the AI model to go international. Because users in different countries may touch on controversial topics under their respective national conditions when using DeepSeek, "this is not just a Western problem, but a global problem. We don't know what kind of censorship mechanism is embedded in this system and how this will anger countries around the world."
Concerns have also been raised about the data security of DeepSeek. On its website, DeepSeek's privacy policy states that all data of registered users will be stored on "secure servers" within the territory of the People's Republic of China. The collected data includes basic information actively provided by users, such as date of birth, user name, email address, phone number, password, etc. The text, sounds, uploaded files, conversation history, etc. entered by users when interacting with DeepSeek may also be collected. In addition, DeepSeek also collects and stores the model, operating system, keystroke pattern and cadence, and IP address of the instrument used by the user.
DeepSeek's privacy policy states that the data collected by the company may be shared with the government if it is for "compliance with applicable laws, legal procedures, or government requirements."
The US government has been trying to stop Beijing from obtaining US user data. Previously, the US Congress passed a bill citing national security reasons, requiring TikTok, whose parent company is based in China, to sell to an American company or be removed from the shelves.
"China's progress in artificial intelligence should be of deep concern to every American," former U.S. Secretary of State Mike Pompeo wrote on X, "Any data you provide to Chinese company-owned apps like DeepSeek could be handed over to the Chinese Communist Party."
US technology companies also collect this data, which is also an area that many American scholars and experts believe needs more regulation. However, Luke de Pulford, executive director of the Inter-Parliamentary Alliance on China, pointed out on X that there is a big difference between the two.
"The difference between this and social media apps in the free world is that in a country governed by the rule of law, you can defend your data rights. This is not the case in China." His post reads.
Although DeepSeek is small in scale and has been established for less than two years, it has been valued by the Chinese central government. On January 20, Liang Wenfeng, the founder of DeepSeek, attended a symposium held by Chinese Premier Li Qiang. The News Broadcast that day broadcast the picture of Liang Wenfeng sitting in the meeting.
In the past few days, one of the focuses of Chinese media reports has been on the Chinese background of the DeepSeek team. Founder Liang Wenfeng graduated from Zhejiang University and has no overseas study or work experience. Most of the people in his company are also ordinary fresh graduates.
When DeepSeek was first established in 2023, Liang Wenfeng told the media: "If you are pursuing short-term goals, it is right to find experienced people. But if you look at the long term, experience is not that important, basic ability, creativity, and love are more important. From this perspective, there are many suitable candidates in China."
Although the Chinese and Western technology circles feel shocked that a Chinese company can still produce high-performance models like DeepSeek under the chip export ban, experts also pointed out that DeepSeek's success is actually inseparable from the contribution of Western AI development. The future does not rule out the possibility that Western technology giants can also find ways to reduce the cost of developing AI.
"Existing research papers cover the availability of all aspects of advanced AI models, which enables DeepSeek to quickly master the latest developments and innovate on top of existing best practices," Triolo of Albright Stone Group said.
"The ultimate success of DeepSeek will depend on many factors, including the ability of other open source leaders (such as Meta) to release more powerful models, and the ability to reduce training and inference costs," he added.
