Chinese artificial-intelligence start-up DeepSeek has postponed the release of its R2 large language model after running into technical problems training the system on Huawei’s Ascend 910B chips, the Financial Times reported. DeepSeek had aimed to debut the Mixture-of-Experts model in the second half of August, part of a broader effort encouraged by Beijing to rely on domestic semiconductors instead of US-made Nvidia processors. Sources told the FT that the Ascend hardware, while approaching roughly 91 per cent of the performance of Nvidia’s A100 GPUs, proved difficult to scale for full-scale training. The company is now retraining R2 on Nvidia hardware and intends to use Huawei chips only for inference, according to the report. The delay highlights the continuing dependence of China’s most advanced AI projects on foreign chip technology despite government efforts to build a self-sufficient supply chain.