DeepSeek Delays New AI Model Amid Huawei Chip Problems – FT
DeepSeek has postponed the launch of its new artificial intelligence model after training efforts with Huawei chips failed, the Financial Times reported on Thursday, citing three sources familiar with the matter.
The Chinese AI start-up faced persistent technical problems while training its R2 model on Huawei’s Ascend chips. As a result, DeepSeek switched to using Nvidia chips for training and Ascend chips for inference, according to the report. These issues were the main reason for the delay of the R2 model, originally set for release in May.
The FT report underscores the challenges Chinese AI developers face in reducing dependence on U.S. technology, particularly Nvidia’s advanced AI chips. Beijing has pushed firms like DeepSeek to adopt Huawei’s Ascend processors, especially after U.S. chip exports to China became a key geopolitical flashpoint this year.
While the U.S. allowed Nvidia to resume H20 chip sales to China in July, Chinese authorities raised security concerns over using American-made chips in critical AI projects. Inference — the process of running a trained AI model — requires less computing power than training, which involves processing large datasets to improve model performance.
Nvidia’s H20 remains the most widely used AI chip in China, despite Huawei’s efforts to promote the Ascend line. The U.S. has blacklisted Ascend chips, alleging Huawei used restricted U.S. technology in their development.
DeepSeek’s earlier R1 model, which disrupted the AI sector earlier this year, was trained primarily on H20 chips. Other major Chinese AI developers, including ByteDance, Tencent Holdings, and Alibaba Group, also rely heavily on Nvidia hardware.
Earlier this week, Beijing instructed local developers to justify their Nvidia chip purchases. Chinese media reports suggest DeepSeek’s R2 model could launch in the coming weeks, as the company races to catch up with a wave of high-profile AI releases from competitors.
