Alibaba and Baidu Soar on Adoption of Self-Developed Chips for AI Training

TradingKey - Reports have emerged that Chinese tech giants Alibaba and Baidu have started using self-designed chips to train AI models, partially replacing Nvidia's chips. Following this news, Alibaba (BABA) surged by 8% on Thursday, marking the biggest two-week gain, while its Hong Kong shares (9988.HK) rose over 5% on Friday. Baidu (9888.HK) also saw an increase of over 8%, reaching a new high not seen in more than a year.

It has been reported that Alibaba began using its self-developed "Zhenwu" processing unit for training smaller AI models at the start of the year. Meanwhile, Baidu is testing its self-developed "P800 Kunlun" chip specifically for training the new version of its Ernie model. However, for their most advanced AI models, both companies continue to rely on Nvidia chips. Nvidia responded to this development by acknowledging the rising competition and reaffirming its commitment to gaining the trust and support of mainstream developers worldwide.

Analysts regard this as a significant milestone for China's AI sector, where companies previously relied heavily on Nvidia hardware for AI model training. With increasing US restrictions on chip supplies to China, the Chinese government has been encouraging domestic chip research and development. Consequently, local tech companies have intensified their efforts in developing AI chips, achieving notable progress.

Last month, Alibaba announced a more powerful AI chip than its predecessors, drawing market attention. However, this chip targets AI inference, with AI model training still dependent on Nvidia's top-tier hardware. Analysts note that despite rapid advancements, domestic chips still show significant weaknesses in training AI models. Engineers have reported issues such as overheating and crashes during training with chips from domestic companies like Huawei, indicating considerable limitations in China's AI chip capabilities.

Employees who have used Alibaba's self-developed AI chips claim that these chips can compete with Nvidia's H20 GPUs, which are specifically limited in computing power for the Chinese market. Nevertheless, the more advanced version of Zhenwu only slightly outperforms Nvidia's flagship A100 data center GPU, which was released in 2020 and banned from sale in China, highlighting a considerable gap with Nvidia.

Encouraged by the news of domestically developed chips being used for AI model training, China's semiconductor sector also experienced a significant boost. Chip company VeriSilicon, often referred to as the "first semiconductor IP stock," resumed trading on Friday with a surge at the opening, hitting a 20% limit up and approaching a market value of one trillion yuan. Meanwhile, Cambricon Technologies, dubbed "rival to Nvidia," rose nearly 10% within the day, with its stock price once again surpassing that of Kweichow Moutai, claiming the top spot in China's A-shares market.