Wang Dong, co-founder and executive president of GPU manufacturer Moore Thread, said, “The development of large models is very rapid at home and abroad. Currently, leading manufacturers complete the cutting-edge basic model version iteration every two months, and in terms of model call costs, we found that the Chinese cutting-edge basic model has a clear cost advantage over foreign models with the same level of intelligence, and the Chinese model is more cost-effective. This also shows that model companies have done a lot of work on how to improve model efficiency, price efficiency, and training costs when computing power is limited.” Wang Dong pointed out that there is no “universal chip” in the inference market, but a combination of “solutions”. “The technology application threshold in the inference market is relatively low, and the scenario is highly fragmented, and no single company can monopolize all application segments; there is no single hardware that is absolutely perfect. Through flexible software and hardware collaboration, every model can find the most suitable hardware combination for it to achieve the best balance between cost and performance. A large number of ISP companies will emerge in the market to provide more cost-effective and flexible customized inference services for MaaS providers or end customers.”

Wang Dong, co-founder and executive president of GPU manufacturer Moore Thread, said, “The development of large models is very rapid at home and abroad. Currently, leading manufacturers complete the cutting-edge basic model version iteration every two months, and in terms of model call costs, we found that the Chinese cutting-edge basic model has a clear cost advantage over foreign models with the same level of intelligence, and the Chinese model is more cost-effective. This also shows that model companies have done a lot of work on how to improve model efficiency, price efficiency, and training costs when computing power is limited.” Wang Dong pointed out that there is no “universal chip” in the inference market, but a combination of “solutions”. “The technology application threshold in the inference market is relatively low, and the scenario is highly fragmented, and no single company can monopolize all application segments; there is no single hardware that is absolutely perfect. Through flexible software and hardware collaboration, every model can find the most suitable hardware combination for it to achieve the best balance between cost and performance. A large number of ISP companies will emerge in the market to provide more cost-effective and flexible customized inference services for MaaS providers or end customers.”