Meituan's LongCat-2.0 shows China can train massive AI models without Nvidia

Editor's take

Meituan has successfully trained its 1.6 trillion parameter LongCat-2.0 large language model utilizing exclusively Chinese-manufactured AI hardware. This development is significant as it demonstrates China's growing capacity to develop and deploy frontier AI models independently of Western semiconductor giants like Nvidia, which currently dominates the high-performance AI chip market. This reduces reliance on foreign supply chains and potentially accelerates China's domestic AI development trajectory, impacting global competition in AI research and deployment.

The next critical area to observe is the performance parity and scalability of these Chinese AI chips compared to Nvidia's H100 and upcoming Blackwell architectures. It will be crucial to see if LongCat-2.0, trained on this indigenous hardware, can achieve comparable inference speeds and training efficiency for real-world applications, such as Meituan's food delivery and local services platform. Continued investment and advancements in Chinese AI chip design and manufacturing will determine the long-term viability of this independent ecosystem.