DeepSeek delays R2 after training shifts to Nvidia hardware

DeepSeek has postponed the release of its R2 large language model after persistent difficulties training the system on Huawei Ascend chips, forcing the Beijing-backed startup to rely on Nvidia hardware for the training phase while continuing to use Ascend processors for inference. The setback, which pushes the originally planned May launch into the coming weeks, underscores both the technical gaps China faces in its push for domestic AI infrastructure and the practical challenges of replacing established Western technologies in high-performance AI workloads.

Industry insiders say the decision to train with Nvidia chips was driven by stability and software maturity concerns with Huawei’s Ascend line. Despite Huawei engineers being on site to assist, the training runs could not be completed successfully. Training, which teaches a model to recognize patterns from large datasets, proved more demanding on the Ascend platform than anticipated, while inference—the step where a trained model generates responses—appeared to work more reliably on the same hardware. The approach of mixing hardware—training on Nvidia, inference on Ascend—reflects the current reality for some Chinese AI...