Chinese AI labs are turning constraint into cost advantage

Exponential View published on 13 May 2026 a field report by Azeem Azhar and Hannah Petrovic from visits to Chinese AI and robotics labs in Beijing, Hangzhou and Shanghai. The authors spoke with representatives of 14 labs and companies, including DeepSeek, MoonshotAI, MiniMax, Z.ai, ByteDance, Alibaba, Ant Group, Xiaomi, Unitree and ModelScope. The main finding is paradoxical. US chip export controls make Chinese development harder in the short term, but they also force Chinese labs to become highly efficient. The authors estimate that Chinese labs obtain four to seven times more intelligence per unit of compute than simple scaling models would suggest. According to Exponential View, Chinese open-source models are only six to eight months behind the US frontier, even though available compute capacity is roughly two to three years behind. For companies, the most important effect is pricing. Cheaper model operation can change the economics of enterprise AI applications. Exponential View says DeepSeek charges $0.43 per million input tokens and $0.87 per million output tokens. Z.ai prices its GLM-5 model at one dollar per million input tokens and, according to public documents, reports a 50 percent gross margin. MiniMax has margins of around 70 percent in enterprise services. The business conclusion is uncomfortable for US and European providers. If cheap model serving becomes a key competitive advantage, Chinese firms may push prices down across the market. European companies, however, must assess not only price but also data protection, supplier dependence, geopolitical risk and regulatory compliance.

Chinese AI labs are turning constraint into cost advantage

The full article is available to users registered at Hard Skills