1 day Ago By Reuters
HONG KONG, April 1 (Reuters Breakingviews) - Following DeepSeek's release of its cutting-edge and free large language model early this year, Meta's (META.O) , opens new tab chief artificial intelligence scientist Yann LeCun corrected , opens new tab those who surmised China is surpassing the United States in the technology. A correct reading, he said, is "open source models are surpassing proprietary ones".
What's less clear, though, is how long China will flood the world with its free innovations. Tech bosses in the People's Republic including Alibaba's (9988.HK) , opens new tab Eddie Wu, Tencent's (0700.
HK) , opens new tab Pony Ma and Robin Li of Baidu (9888.HK) , opens new tab have embraced open source, a system in which anyone can freely use, study, modify and share artificial intelligence software and code. China appears to endorse the approach: in a meeting in January with Premier Li Qiang, DeepSeek CEO Liang Wenfeng was designated representative of the AI sector.
Open source is not solely a Chinese novelty but the country's AI offerings are more in line with the standard definition. DeepSeek's source code is under a license that allows near-unrestricted use. At the other end of the spectrum, OpenAI, the U.
S.-maker of ChatGPT, keeps the training data and processes of its proprietary models closely guarded, though the company has plans to release a model in which the trained parameters are publicly accessible in the coming months. Even Meta's freely-available Llama models limit some commercial applications but the company agrees that the path to becoming an industry standard is by being open generation after generation.
China is, of course, backing open source because it serves the country well at this moment. Washington's tech restrictions against Beijing mean the country's companies do not have access to the most powerful chips from Nvidia (NVDA.O) , opens new tab to train and deploy AI models.
Using advanced open-source models from companies that do have access is one workaround. Before DeepSeek, most Chinese models, including those developed for military use , were just variations of Meta's Llama. Pooling resources in this way gives China the chance to catch up with the United States faster.
In the past few weeks, Baidu, Alibaba, Tencent and DeepSeek have all rolled out updates or releases to their open-source offerings. Jack Ma-founded Ant, for instance, has developed techniques to train AI models using less powerful domestically-made chips from Huawei with results similar to using Nvidia's processors, Bloomberg reported , opens new tab , citing sources. If widely adopted, such breakthroughs will bring China closer to achieving President Xi Jinping's goal of technological self-sufficiency.
The potential for open source to help laggards catch up is also why the European Union, home to startups like France's Mistral AI, is championing a similar approach: in February, European Commission President Ursula von der Leyen unveiled , opens new tab plans to mobilise 200 billion euros to invest in "cooperative, open innovation" in AI. Generosity with its innovations also wins China global recognition. "Contributing to [open source] earns us respect", DeepSeek's founder Liang said in a rare interview published , opens new tab last year.
That's true for the country too: the availability of free and powerful AI models has bolstered China's soft power beyond the West too. American economist Tyler Cowen recently noted , opens new tab that China has gained an advantage over the United States - "not only in technology, but also in vibes". China has embraced open source in other areas too.
The government has pushed companies to use freely-available RISC-V chip architecture, backed by Huawei, Nvidia and others, instead of licensing technology from UK-based designer Arm (O9Ty.F) , opens new tab , and Intel (INTC.O) , opens new tab and AMD (AMD.
O) , opens new tab in the United States. With the latter, the risk is that China's access to these technologies could be cut off by Washington at some point. There are downsides to being so open, though.
It undercuts companies' ability to generate income and that will hamper future investment. Model-owners like OpenAI usually charge for access to their models and products; they also collect fees from developers that want to integrate chatbots and other products with AI models. Open source means DeepSeek can only rely on the second type of fee for revenue.
That may not matter too much to Liang, who says he is prioritising innovation over profit at the privately-owned company. But for publicly-traded peers like Alibaba, which pledged some $53 billion of investments into AI and cloud computing, poor returns can weigh on its share price and valuation. The $315 billion e-commerce giant has both proprietary and free models, plus a sizable cloud computing business.
Alibaba Chair Joe Tsai explained the bet last week at HSBC's Global Investment Summit: customers will use the free models but buy computing power, data handling, security and "a full stack of software" from Alibaba. That assumes Chinese companies, so far laggards in AI and IT adoption, will up their spending. The bigger threat to open source in China may be from officials in Beijing, who maintain a tight grip on the economy through centralised industrial policies and strict regulations.
Generative AI products and services, for example, have to "adhere to core socialist value" and exclude content that "endangers national security". Current rules are ambiguous when it comes to open-source models, which by nature are decentralised and borderless. As Chinese firms catch up to or surpass Western AI rivals, Beijing may start feeling differently about giving away technology that ultimately has the power to determine which country has an edge in military or cyber warfare.
Some DeepSeek employees are under travel restrictions, The Information reported , opens new tab in March, citing sources. Gregory C. Allen from U.
S.-based think tank Center for Strategic and International Studies observes , opens new tab that DeepSeek's innovative AI training techniques could benefit U.S.
firms more than Chinese ones since the Americans can apply them to far superior computing resources. It is worth noting that China's embrace of open source does not extend to sectors where it is already a global leader, such as electric-vehicle batteries and green energy. In fact, the government banned the export of certain rare earth processing technologies in 2023 and, more recently, has delayed plans by automaker BYD (002594.
SZ) , opens new tab , to build a factory in Mexico on concerns that its technology could leak to the United States, the Financial Times reported , opens new tab last month, citing sources. The supply of cheap and free Chinese AI models may be plentiful but it may not last long. Follow @mak_robyn on X.
For more insights like these, click here , opens new tab to try Breakingviews for free. Editing by Una Galani, Ujjaini Dutta and Streisand Neto Thomson Reuters Robyn Mak joined Reuters Breakingviews in 2013. Previously, she was a Research Associate for the Global Policy Programs at the Asia Society in New York.
She has also worked at the Carnegie Endowment for International Peace in Washington DC and interned at several consulting firms, including the Albright Stonebridge Group. She holds a masters degree in international economics and international relations from the Johns Hopkins School of Advanced International Studies and is a magna cum laude graduate of New York University..
