Shen Yunxui Column: After the 'Kimi Moment' – Will the AI Cold War Shift from Chip War to Model War?

Over the past two years, the focus of the US-China tech war has been on chips. The United States has restricted exports of high-end NVIDIA GPUs, aiming to slow China's development of large AI models. Meanwhile, China has actively built its own chip supply chain while seeking various ways to improve computing efficiency. It was once believed that whoever owned the most GPUs would dominate the AI era. But what may now be more noteworthy is not chips, but AI models themselves.

Recently, The Wall Street Journal reported that US AI companies such as OpenAI and Anthropic are lobbying the US government to impose stricter restrictions on Chinese AI models, citing national security risks. At the same time, several Chinese AI companies have accelerated the release of open-weight versions in recent months, aiming to rapidly build a global developer ecosystem.

Against this backdrop, Moonshot's newly launched Kimi K3 has drawn global attention from the AI community. Some developers have begun using the term 'Kimi Moment' to describe what might become, after DeepSeek earlier this year, another moment when Chinese AI shocks the world. According to public benchmarks, Kimi K3 performs close to, and in some cases surpasses, OpenAI's GPT-5.5 and Anthropic's latest models. With extremely low inference costs, many commentators believe the technological gap between US and Chinese AI models is rapidly narrowing.

However, another voice quickly emerged. Recently, an analysis document circulated online questioned whether some of Kimi K3's benchmark results might be due to training methods specifically optimized for test items, failing to reflect the model's capabilities across all real-world applications—thus being 'unreliable' or 'inconsistent'.

What's worth considering is not whether Kimi K3 surpasses GPT-5.5, nor whether the leaked document is ultimately true or false, but rather the larger trend this entire incident reflects: the global AI competition's focus is gradually shifting from 'who owns the most GPUs' to 'who can launch the most attractive model fastest' and 'who can build the largest developer ecosystem'.

From a national security perspective, the US concerns are well-founded. Large language models have become critical tools for infrastructure in government, military, energy, finance, and healthcare. If a model's training methods, update mechanisms, and data sources are entirely controlled by foreign companies, nations naturally worry about potential risks. China has similarly begun restricting the export of certain AI technologies and model-related technologies, reflecting Beijing's view of AI as a strategic asset, not a mere commodity.

In fact, what made the 'DeepSeek Moment' and today's emerging 'Kimi Moment'震撼 the world is cost. If US models lead by ten percent in performance, but Chinese models can deliver comparable capabilities at 10%, or even 1%, of the cost, then what global enterprises truly compare is no longer just technology, but economics. If Chinese models offer similar capabilities at near-zero or much lower costs, global developers and enterprises may naturally shift to other platforms. This resembles how Chinese electric vehicles have recently challenged the automotive industries of Europe, the US, and Japan: winning by cost, capable of reshaping the entire industry's pricing structure.

Even if the US wants to restrict Chinese AI, how effective can such restrictions be? If the target is a website like ChatGPT or Claude, restricting official websites, APIs, or government procurement is relatively easy to enforce. But if Chinese models adopt an open-weight strategy, the situation is entirely different. Any developer worldwide can download and run the model on their own GPU, private server, local NAS, or internal enterprise systems. Unlike cloud services, it doesn't require connecting to official servers, nor does it rely on centralized authorization like traditional software. If the US attempts to restrict businesses, research institutions, or individuals from downloading Chinese open-weight models, enforcement difficulty would greatly increase, potentially sparking further controversy over scientific freedom, open-source technology, and global technology circulation.

The core of future international political competition will no longer be controlling chips, but controlling the 'algorithms' themselves. Whoever can get more countries, companies, and developers to adopt their model as foundational infrastructure will gain a strategic advantage more important than chips in the next global order reshaping. This war might be humanity's last 'traditional' war—because the next one may already be fully dominated by AI.

*The author is a Hong Kong-based international relations scholar and currently an Associate Professor at the Taiwan-Hong Kong International Research Center, National Sun Yat-sen University.