Gold prices edge lower; heading for weekly losses ahead of U.S.-Russia talks
Investing.com-- DeepSeek has delayed the release of its new artificial intelligence model as training efforts using Huawei chips fell through, the Financial Times reported on Thursday, citing three people familiar with the matter.
The Chinese AI start-up encountered persistent technical issues when training its R2 model using Huawei’s Ascend chips, the report said, prompting it to use Nvidia chips for training and Ascend for inference, the FT reported.
The issues were the main reason why Deepseek’s highly anticipated R2 model launch was delayed from May, the report said.
The FT report highlights the difficulties faced by Chinese AI developers in reducing their reliance on U.S. tech, specifically Nvidia’s AI chips. DeepSeek and its peers were encouraged by Beijing to use Huawei’s Ascend chips earlier, especially as U.S. chip exports to China became a major point of contention this year.
While the U.S. did allow Nvidia to resume sales of its H20 chips in China in July, Beijing flagged security concerns over the use of American-made chips in critical Chinese AI development.
Inference– which involves using a trained AI model to generate responses– usually requires less computing power than training, which involves using large amounts of pre-existing data to improve the capabilities of AI models.
Nvidia’s H20 is still by far the most popular AI chip in China, even as Huawei and Beijing pushed for adoption of the latter’s Ascend line. The U.S. has blacklisted the Ascend line on grounds that Huawei used restricted U.S. tech to develop the chip.
DeepSeek’s R1 model, which shook up the AI sector earlier this year, was developed largely using H20 chips. The chips are also used by other major Chinese AI developers, including ByteDance, Tencent Holdings Ltd (HK:0700), and Alibaba Group (NYSE:BABA).
China earlier this week asked local developers to justify their orders of Nvidia chips.
Chinese media reports said that the R2 model will be released in the coming weeks, as DeepSeek lagged a series of high-profile releases from its competitors.