Lu Shaowei Column: Don't Need to Be the Strongest to Be Useful? Kimi K3 Brings a New Chapter to the US-China AI Competition

Last week, Chinese AI startup Moonshot AI unveiled its new large language model, Kimi K3, sending shockwaves through the global AI and chip industries. The model's release has led to a drop in global chip stocks, narrowed the US-China AI gap to just 3 months, and sparked debates on open-source vs. closed models, regulation, safety, and accessibility. Moonshot AI, founded in 2023, joins other Chinese AI companies like DeepSeek in offering models that are more cost-effective and energy-efficient while delivering near-top-tier performance. Kimi K3, for instance, is priced at about one-third of Anthropic's Fable model, making it an attractive option for many users. The immediate impact of Kimi K3's release was felt on Wall Street, with global AI and chip-related stocks taking a hit. The model's ability to autonomously design chips in just 48 hours also poses a threat to the two dominant players in the electronic design automation (EDA) field, causing their stock prices to drop by 7-9%. While analysts have downplayed the idea of Kimi K3 completely replacing EDA tools, its potential to handle mature processes and its expected evolution make it a significant disruptor in the industry. The broader implications of Kimi K3's release lie in the US-China AI competition, where cost and pricing play a crucial role. Consumers, whether individuals or businesses, often prioritize price over having the most advanced product. Historical examples, such as the victory of Microsoft's Windows over Apple's Macintosh in the PC era, VHS over Betamax in the video format war, and QWERTY over Dvorak in keyboard layouts, demonstrate that the best product is not always the winner. Despite the US's OpenAI and Anthropic models still leading in performance, the price gap and the open-source nature of Chinese AI models make the outcome of the competition uncertain. China's AI companies adopt an open-source and open-weight model, reducing costs and prices, rapidly expanding their market, and popularizing technology. This strategy was emphasized by Chinese leader Xi Jinping at the World Artificial Intelligence Conference in Shanghai last week, setting the tone for China's AI path. Chinese AI models are also gaining popularity overseas, with the top six most popular models on the OpenRouter platform being Chinese, accounting for about 63% of US enterprise token usage, up from less than 10% a year ago. On the Hugging Face platform, Chinese models have surpassed US models in monthly and overall download volumes, handling nearly 30% of AI requests. This shift in market competition, from pursuing the strongest model to valuing cost, autonomy, and customization in open-source models, presents an opportunity for Chinese AI. The release of Kimi K3 also raises the question of how much China has freed itself from US chip restrictions. Moonshot AI has not disclosed the chip models used to train K3, leading to speculation that Chinese AI companies, unable to obtain advanced US AI chips (such as those from NVIDIA or AMD) due to strict regulations, may have used domestically produced chips. If K3's training, from AI chips (Huawei) to memory, was done using Chinese products and the results have been verified, this would not be good news for US AI chip companies or South Korean memory manufacturers. US top AI company executives have issued warnings about Chinese models, and China is planning to expand and tighten export controls on AI technology. Regardless of the 'sovereign AI' issue, having more models, companies, and even countries competing in the AI field is a good thing. In the current and foreseeable future, the global AI landscape will likely be dominated by the US and China, with other countries struggling to keep up. The final outcome is more likely to be a scenario of 'each leading in their own way and coexisting' rather than a dramatic 'winner takes all' situation, similar to the Apple and Android duopoly in the smartphone world.