Chinese AI labs innovate amid hardware constraints, driving cost-effective smart economics.

ChainCatcher report: Research firm Delphi Digital has published findings indicating that hardware constraints have pushed Chinese labs toward cheaper models and more efficient tech stacks. Export controls limit access to advanced chips; domestic accelerators can handle inference, but cutting-edge training remains difficult. Chinese labs have innovated in system efficiency, model architecture, training data, and reinforcement learning: DeepSeek developed a method to avoid GPU idle time by optimizing data movement, while ByteDance and Moonshot released similar versions within months. Architecturally, training has shifted toward cheaper numerical formats compatible with domestic chips, sparse architectures reduce per-token computation, and attention mechanisms have been redesigned three times to control long-context memory costs. On the data side, denser prediction targets and improved optimizers enhance per-token learning signals. In reinforcement learning, DeepSeek’s GRPO has become the default approach, with ByteDance, Alibaba, and MiniMax each launching successor systems within a year. Chinese models are now the most cost-efficient frontier-adjacent systems. On OpenRouter, the share of weekly token consumption by Chinese-developed models rose from under 1.2% at the end of 2024 to the majority by 2026. The Chinese AI market experienced a price war from 2024 to early 2026, with multiple labs carrying implicit revenue multiples of approximately 47–117x, compared to Anthropic and OpenAI at 15–21x. Huawei’s Ascend chips are currently viable for inference; DeepSeek reportedly attempted training R2 on Ascend but encountered technical issues and reverted to NVIDIA. Domestic accelerator supply still falls short of projected demand.