Moore Threads co-founder Wang Dong: The inference market has no “universal chip,” but a combination of solutions

Odaily Planet Daily news: Wang Dong, co-founder and executive CEO of GPU maker Moore Threads, said: “The development of large models is moving very quickly both in China and abroad. Now top vendors complete an iteration of frontier foundational model versions every two months on average. As for the model usage cost, we have found that China’s frontier foundational models have a clear cost advantage over overseas models with comparable intelligence levels, and that China’s models offer better price-performance. This also shows that, under limited compute resources, model companies have done a great deal of work on how to improve model efficiency, price efficiency, and training costs.”

Wang Dong noted that the inference market does not have a “one-size-fits-all chip,” but rather a combination of “solutions.” “The technical application threshold in the inference market is relatively low, and the scenarios are highly fragmented. No single company can monopolize all sub-application scenarios. There is no absolutely perfect single hardware. Through flexible software-hardware co-optimization, each model can find the most suitable hardware combination to achieve the best balance of cost and performance. The market will see many ISP companies emerge, providing more cost-effective and more flexible customized inference services for MaaS providers or end customers.”