Since the emergence of Chinese AI startup DeepSeek, which shocked U.S. tech giants and swept through the global industry, one year has passed, and China's AI has now become a core pillar of the global artificial intelligence ecosystem. With an open-source strategy that emphasizes "high accessibility and low cost," Chinese AI has transformed into a practical technology that developers can "use as easily as extending their arm."

On the 18th local time, the Financial Times (FT) evaluated the Sino-U.S. AI competition, stating that in the long run, "China will be the ultimate winner." Looking back a year ago, when DeepSeek released its inference model R1, it broke the "law of scale" – the iron rule in the AI field that performance increases proportionally with computing power and capital investment. Without using tens of thousands of high-end GPUs, R1 achieved performance comparable to top U.S. models. At that time, the industry reacted strongly; a16z co-founder Marc Andreessen called it "the Sputnik moment for AI," while OpenAI CEO Sam Altman admitted, "they (China) will release models better than ours."

Over the past year since DeepSeek's launch, the evolution of China's AI ecosystem has been rapid. DeepSeek released R1 under an "open-weight" (open-weight) model, allowing global users to freely download, modify, and retrain it. Unlike the "black-box" closed models used by major U.S. tech companies, which charge based on usage, the open-weight model grants developers direct control and the freedom to customize the model. This strategy became a key component of China's AI expansion. Meanwhile, companies such as Alibaba's Qwen (Tongyi Qianwen) and Moonshot's Kimi also followed suit, releasing open-weight models and quickly expanding the boundaries of China's AI ecosystem.

Data from the Stanford HAI Institute shows that 63% of "fine-tuned AI models" (AI models redesigned for specific purposes) on the global open-source platform Hugging Face originate from Chinese models. This means that Chinese AI has become a "standard" in the frontline of global development. In November last year, a joint study by MIT and Hugging Face confirmed this trend: over the past year, the global download share of Chinese models reached 17%, surpassing the U.S. (15.8%), achieving a significant advantage.

This wave is profoundly reshaping the underlying logic of AI hegemony competition. Just two or three years ago, the Sino-U.S. AI rivalry focused on semiconductor export controls, with the mainstream view being that cutting off high-end GPU supply to China would secure U.S. victory. However, Chinese models have significantly improved training efficiency and spread widely despite limited computing power, tearing a huge gap in this premise. Park Seung-jun, director of the Korea Institute for Business Strategy, said, "When communicating with Chinese AI professionals, we often hear sarcastic remarks like 'Thank you, Trump,' reflecting how the Sino-U.S. trade friction has objectively forced Chinese AI technology to evolve and break through."

Even within the core of U.S. AI hegemony, Silicon Valley, Chinese AI has become an unavoidable reality. The rapid narrowing of performance gaps, combined with high flexibility in cost and operations, is prompting many developers to change their stance. When U.S. AI companies still clung to a closed and expensive "premium approach" like Apple's iOS system, China took a similar approach to Google's Android – by extensively opening models, it fostered diverse derivative models and application scenarios, thereby achieving influence breakthroughs in the global development ecosystem through penetration.

The MIT Technology Review noted on the 12th that "increasingly more Silicon Valley applications are quietly rooted in Chinese open-source models; meanwhile, the 'time difference' between the release of leading models from both sides has shrunk from months to weeks, and is now almost synchronized."

At the same time, China is striving to expand its voice in the global AI ecosystem. Especially in non-Western markets where development costs are high and infrastructure is limited – such as Africa, Southeast Asia, and the Middle East – the adoption rate of Chinese models is growing exponentially. On the 13th, Microsoft (MS) President Brad Smith told the Financial Times (FT) that "DeepSeek technology is spreading at an unprecedented speed in emerging markets like Africa, which is a competitive reality that U.S. companies must face globally." Analysts point out that although U.S. giants may still hold an edge in short-term profits, given the strong user stickiness of AI models and tools once selected, over time, the underlying influence of Chinese companies will be incomparable. Professor Baek Seo-jeong (音) from the Department of Global Culture and Trade at Hankyong National University predicted, "In the future, a 'binary structure' may emerge – with the top end dominated by U.S. closed models, while the vast bottom-level ecosystem is built by Chinese open-source models."

In recent times, during the first evaluation of South Korea's Ministry of Science and Information and Communication Technology's "Independent AI Basic Model Project," NAVER Cloud failed due to the use of the Chinese model Qwen. Deputy Prime Minister and Minister of Science and Information and Communication Technology, Bae Kyung-won, expressed his frustration on his personal social media, saying, "DeepSeek-R1 achieved top performance on low-spec chips, not only breaking the monopoly of big players but also shaking up the global industrial landscape. This case raises extremely sharp issues."

Experts unanimously believe that discussions about AI development have moved beyond the initial stage of "whether to use Chinese technology" to a deeper institutional competition – namely, how to define the acceptance boundaries and security management standards of open-weight models. Kim Myung-joo, director of the AI Security Institute, pointed out, "More than 80-90% of the global software industry is already rooted in open-source systems. The immediate priority is not to blindly reject them, but to achieve more refined classification management and access assessment based on business nature and data characteristics."

Source: Central Daily

Original: toutiao.com/article/7598099366796165659/

Disclaimer: This article represents the views of the author.