On July 18, during the "AI Computing Power Technology Architecture Innovation Forum" at the 2026 World Artificial Intelligence Conference (WAIC), Wang Dong, the co-founder and executive president of GPU manufacturer Moore Threads, observed that the evolution of large-scale models has accelerated significantly on a global scale. Leading firms are now capable of iterating versions of their cutting-edge foundational models every two months on average. When it comes to model invocation expenses, China's advanced foundational models exhibit a marked cost advantage over their foreign counterparts with comparable intellectual capabilities, highlighting the superior cost-efficiency of Chinese models. This suggests that, despite constrained computing resources, Chinese model developers have made considerable strides in improving model efficiency, price competitiveness, and training costs.

However, Wang also emphasized that the inference market does not have a 'one-size-fits-all chip'; rather, a blend of 'solutions' is necessary. Given the relatively low barrier to technical application and the highly fragmented nature of application scenarios in the inference market, no single company can dominate all niche areas. There is no single piece of hardware that is absolutely flawless, but through adaptable software-hardware synergy, each model can discover the most appropriate hardware mix to strike the ideal balance between cost and performance. Looking ahead, the market will see the rise of numerous ISP companies, providing MaaS providers or end customers with more cost-effective and adaptable customized inference services.