DeepSeek R2 Delayed: Huawei Chip Issues Force NVIDIA Use

The release of DeepSeek’s highly anticipated R2 large language model has reportedly been delayed, a setback attributed to persistent performance issues with Huawei’s Ascend processors. This development, as reported by the Financial Times, underscores the significant challenges Beijing faces in its ambitious drive to replace American technology with homegrown solutions.

Following the earlier success of DeepSeek’s R1 model, which garnered considerable attention in the AI community, Chinese government officials reportedly encouraged the startup to utilize domestic Ascend chips for its subsequent R2 development, rather than continuing with NVIDIA processors. However, DeepSeek encountered substantial technical hurdles during the R2 training process when attempting to use the Ascend chips. These difficulties ultimately compelled the company to revert to NVIDIA chips for the intensive training phase, while still planning to use Huawei’s processors for the inference stage—the process of applying the trained model to new data.

This strategic pivot comes amidst a broader landscape of evolving export regulations and domestic pressures within China. Recent adjustments to U.S. export rules have allowed for the shipment of certain NVIDIA and AMD GPUs, specifically the NVIDIA H20 and AMD MI308, to China. These chips are reportedly designed to offer performance roughly comparable to China’s domestically developed GPUs. Concurrently, there are reports that Chinese authorities are increasingly scrutinizing companies, requesting justifications for their continued reliance on American-made GPUs.

Sources cited by the Financial Times indicated that the Huawei processors exhibited several critical deficiencies compared to NVIDIA’s offerings, including stability problems, slower inter-chip connectivity, and inferior accompanying software. In an effort to overcome these obstacles, a team of Huawei technicians reportedly collaborated directly with DeepSeek on-site. Despite these concerted efforts, a successful training run on the Ascend processors could not be achieved.

The situation highlights the complex interplay between national strategic goals, technological capabilities, and market realities. While China is aggressively pursuing self-sufficiency in critical technologies, the DeepSeek case suggests that the performance gap in advanced AI accelerators remains a significant hurdle. The Financial Times report indicates that DeepSeek’s R2 model could still see its release in the coming weeks, presumably after its successful training on NVIDIA hardware.