As demand for inference computing rises, Chinese companies are fighting through bottlenecks by upgrading technologies.
Lanxin Computing in south China's Shenzhen City has developed chips that can run large models on all-in-one AI computers.
"In inference computing, the key factor now is token-generation efficiency. With the rapid growth of AI agents, demand is surging across government services, healthcare, finance and other sectors. Our products are currently in short supply," said Xie Shengluo, co-founder of Lanxin Computing.
Electricity supply for chips is also facing pressure. In high-power applications at computing centers, conventional silicon-based solutions are increasingly challenged by energy losses and heat management.
Wan Yuxi, director of Shenzhen Pinghu Laboratory, leads a team that has developed a gallium nitride crystal film on an 8-inch silicon carbide substrate. It can substantially improve both efficiency and power-delivery capacity when used to charge GPUs.
"Previously, the efficiency of the silicon-based gallium nitride devices we used was only around 85 percent. Devices made using our new platform can raise that to around 95 percent. What does that mean? It means by 2035, we estimate that AIDCs (AI computing data centers) in China alone could require as much electricity as the output of five Three Gorges hydropower stations. After a technology upgrade, the energy losses alone could be reduced by an amount equivalent to the electricity generated by half a Three Gorges station," said Wan.
Chinese companies upgrade inference computing technologies to meet AI demands
