The rollout of DeepSeek’s highly anticipated R2 AI model has been delayed, as the Chinese startup ran into significant technical difficulties training the model using Huawei Ascend chips, according to reports from the Financial Times and Reuters.
What Went Wrong
- Authorities had encouraged DeepSeek to adopt Huawei’s Ascend chips instead of U.S. technology to support China’s goal of reducing reliance on foreign hardware.
- However, DeepSeek encountered persistent issues: training runs with the Ascend chips failed despite the intervention of Huawei engineers who were dispatched to assist
- As a result, DeepSeek resorted to using Nvidia chips for training, with Huawei Ascend chips retained only for inference tasks.
Broader Implications
- This setback underscores the performance and compatibility limitations of Chinese-made chips in handling complex AI model training—especially when compared to U.S. alternatives like Nvidia.
- The delay, which pushed the R2 model launch past its original May timeline, also included setbacks from extended data labeling efforts.
- Internally, DeepSeek’s founder Liang Wenfeng has expressed frustration with the model’s slow progress and is advocating for extra time to ensure the R2 maintains the company’s technological edge
What It Means for China’s Tech Ambitions
- DeepSeek’s difficulty reflects a larger challenge: despite heavy investments in domestic chipmaking, China’s hardware still falls short in critical AI tasks.Financial Times
- The episode illustrates the complexity of achieving technological independence and may push Chinese firms to continue relying on foreign hardware for high-demand AI workloads
