TAIPEI, TAIWAN — Since breaking onto the global stage in late January, China’s homegrown AI model DeepSeek has changed the competitive landscape with the United States and fueled Beijing’s ambition to promote its AI models internationally, especially in developing countries.
“We have helped developing countries enhance capacity building, advocating that AI technologies should be open-sourced and there should be greater accessibility to AI services so that the benefits of AI can be shared by all countries,” China’s Foreign Ministry spokesperson Guo Jiakun told a news conference on February 10.
Chinese Vice Premier Zhang Guoqing told the Paris AI Summit on February 11 that China is “willing to work with other countries to safeguard security and share achievements in the field of artificial intelligence to build a community with a shared future for mankind.”
But experts warn that China’s low-price strategy and the Chinese Communist Party narratives baked into its model pose long-term challenges to the global AI market and to global information.
China tells its side of AI story
China is trying to reset the narrative around AI to convey the message that it is different from Western countries, and not elitist or self-serving, said Alex Colville, a Taipei-based researcher at the China Media Project.
As Chinese President Xi Jinping put it, China wants to ensure AI is “not just a game for rich countries and rich people,” effectively positioning China as a leader of developing countries while portraying the West as an obstacle to progress, he said.
Such a narrative enables China to claim it is challenging U.S. dominance of AI, and helping Chinese companies expand overseas through open-source and low-cost AI models that offer more flexibility than their Western competitors, Colville said.
He cited the examples of Saudi Arabia, which signed a joint venture agreement with Chinese facial recognition company SenseTime in September 2022 to create an AI lab locally, and Egypt, which signed an agreement with Chinese telecom giant Huawei in May 2024 to create the “world’s largest” Arabic language AI model.
China has already worked with dozens of Global South countries through its “Belt and Road Initiative” over the years, so it’s not surprising if these countries are now more inclined to cooperate with Chinese companies like DeepSeek in AI development.
In India, which has 22 official languages, DeepSeek could improve communication between citizens — a significant advantage, especially since some Indian politicians already use AI translation software to translate their speeches and expand voter reach.
However, the risk remains that Indians may gradually become more immersed in Beijing’s world view and its values.
The disruptive pricing of Chinese AI
Mohit Agrawal, Director at Counterpoint Research in the Netherlands, said DeepSeek should be viewed “beyond a geopolitical lens.”
“Its emergence won’t force companies to choose sides between the U.S. and China,” he told Voice of America by email.
“However, DeepSeek’s success — despite existing constraints — will undoubtedly boost confidence in the Chinese AI ecosystem,” he said.
Companies adopting the DeepSeek AI model will benefit from lower inference costs, open-source flexibility, and competitive performance, according to Agrawal.
There are also disadvantages, including data security concerns, content moderation issues, and the risk of bans in multiple countries.
While AI may bring benefits, these don’t necessarily translate into sustainable revenue streams, he said, adding that DeepSeek’s low-cost approach doesn’t only challenge Western companies, but also Chinese tech giants like Alibaba, Baidu, and Huawei.
Baidu has announced that its AI chatbot “ERNIE Bot” will be completely free from April 1, including a new deep search function.
Fraser Howie, U.K.-based co-author of “Red Capitalism,” warned that there is always a price to be paid, even for supposedly low-cost AI.
“This is the classic Chinese model of basically undercutting everybody on price,” Howie said. “In some ways, it should be good. It does become a much more accessible and cheaper option.”
AI is expensive, and someone always foots the bill, whether they are U.S. AI companies investing in R&D for the models DeepSeek uses, or users paying in non-monetary ways like handing over their data, Howie said.
DeepSeek claims to have cost just US$6 million, but that figure only covers the final training phase, and doesn’t include the cost of chip manufacturing and earlier training.
“Someone’s paying a price. Because this isn’t free,” he said.
Calls grow for U.S. to boost chip export controls
Dario Amodei, CEO of U.S. AI startup Anthropic, said many of the claims about how much of a threat it poses to U.S. dominance in the sector are exaggerated.
In an article published in January, Amodei argued that DeepSeek does not “do for US$6 million what cost U.S. AI companies billions.”
A fairer statement would be that “DeepSeek produced a model close to the performance of US models 7-10 months older, for a good deal less cost (but not anywhere near the ratios people have suggested),” he wrote.
DeepSeek’s limitation remains computing power, the inevitable result of China’s inability to obtain advanced AI chips, he said.
Amodei also predicted that AI could achieve the level of “countries of geniuses in a datacenter” within two years.
His article called on the U.S. government to strengthen chip export controls to China, arguing that if China can’t lay its hands on millions of chips, “we’ll (at least temporarily) live in a unipolar world, where only the U.S. and its allies have these models.”
If China does manage to catch up and produce millions of advanced AI chips domestically, the world could develop along bipolar lines in future, with both the U.S. and China having powerful AI models driving rapid scientific and technological advancement, Amodei wrote.
However, China could direct more talent, capital, and focus to military applications of the technology, which could help it take a commanding lead globally, not just for AI but for everything, he said.
Democratic version of DeepSeek
Rather than choosing a side between the U.S. and China, some analysts suggest that global AI development may be gradually moving towards an era of “sovereign AI,” where countries rush to develop their own AI technologies to reduce reliance on U.S.-based technology.
Liang-bin Hsueh, co-founder of Taiwanese AI startup MeetAndy, started working on a traditional Chinese character version of DeepSeek R1 in early February.
His project “Freedom Gunpla R1” aims to crowdfund NT$35 million (US$1.07 million) to cover GPU chips, electricity costs, and the fine-tuning of the DeepSeek model.
If retrained, DeepSeek R1 could be developed into a sovereign AI model that aligns with democratic values and suits Taiwan’s needs, facilitating the development of consumer-friendly applications, robots, and automated services, Hsueh told VOA, acknowledging that his project may only have a 30% chance of success.
“Taiwan definitely learns from the U.S. because of our shared values, identity, and target markets,” Hsueh said.
“But in software, we talk about globalization while implementing localization... In AI terms, we call it sovereign AI. With DeepSeek, I see an opportunity because it’s cheap and effective, so people will use it,” he said.
According to Hsueh, if Taiwan were to develop sovereign AI independently, the costs would be astronomical and the model would be hampered by the island’s limited access to training data, and using DeepSeek R1 gets around those issues.
Democratic vs. authoritarian AI
The risk of adapting Chinese AI like DeepSeek is that “these models become personalized propaganda tools, parroting CCP narratives on China-related topics to users,” China Media Project’s Alex Colville said.
There are concerns that programmers might not be able to remove all Chinese propaganda or censorship from the massive dataset, Colville said in an email to VOA, citing DeepSeek’s censorship of politically sensitive information about the status of Taiwan or the Tiananmen massacre.
He said his research on “uncensored” Chinese AI models still revealed implicit CCP propaganda narratives or false information. Even new models built on Chinese platforms (like Alibaba’s Tongyi Qianwen and DeepSeek), while incorporating diverse information from outside China’s firewall, still present CCP propaganda alongside it as equally valid.
“Developers and coders sometimes don’t know what they should be looking for if they want to remove Chinese propaganda from models they are using to create apps and services, beyond removing code that triggers the model to delete or censor an answer,” he said.
“This risks impacting an ordinary user’s ability to tell the difference between what is and is not (Chinese) propaganda,” Colville said, calling for more efficient ways to train LLMs (large language models) to identify these false narratives.
To read the original story in Chinese, click here.
