China Telecom advances AI with new language models on domestic chips

China Telecom, a prominent state-owned telecommunications giant, has made significant strides in the field of artificial intelligence with the development of two large language models (LLMs), which have been trained entirely on domestic chips. This announcement marks a pivotal moment for China as the country seeks to bolster its technological independence amidst international trade and technology restrictions.

Unveiled by the Institute of AI at China Telecom, the open-source TeleChat2-115B model boasts a robust architecture with over 100 billion parameters. Alongside it is a second, yet unnamed, model that reportedly houses a staggering 1 trillion parameters. Both models have been trained using tens of thousands of locally manufactured chips, highlighting China’s efforts to establish complete self-reliance in LLM training.

This development is particularly notable given the challenging context of strict export regulations imposed by the United States. These regulations have restricted China’s access to high-end graphical processing units (GPUs) like Nvidia’s H100 and A100 models. In response, Chinese technology companies have been turning increasingly to domestic alternatives, aiming to mitigate dependencies on foreign technology suppliers.

While China Telecom has not officially confirmed the suppliers of the chips used in the LLM projects, speculation points towards significant involvement from Huawei. The Shenzhen-based technology conglomerate has been positioning itself as a leading domestic alternative for advanced processors, especially following global restrictions on Nvidia products. Previously, China Telecom indicated its interest in developing LLM technology utilizing Ascend chips, which are developed by Huawei.

Huawei has been proactive in asserting its role in the AI chip market and has started distributing samples of its latest Ascend 910C processor to various Chinese server and telecommunications firms for evaluation. This is part of Huawei’s strategic attempt to attract major Nvidia customers in China, persuading them to consider transitioning to Huawei’s products. Notably, it was reported that ByteDance, the parent company of TikTok, has ordered 100,000 Ascend processors, showcasing confidence in Huawei’s domestic chip offerings.

Moreover, China Telecom is also investigating potential collaborations with Cambricon, an emerging AI chip start-up based in China, further illustrating strategic efforts to diversify its hardware supply chain and reduce reliance on a single supplier.

The ambitious efforts by China Telecom and its partners indicate a significant shift in China’s AI landscape, one that underscores an increasing dedication to technological self-sufficiency. This development may have broad implications for how AI advancements are pursued and developed within China and highlights a concerted response to broader geopolitical challenges impacting technology access and development.

Source: Noah Wire Services