China Telecom says their new AI model was trained using China-made Huawei AI-chips

China Telecom, one of China’s largest state-backed telecom giants, has announced the development of two advanced large language models (LLMs) using entirely Chinese-made chips. This marks a significant milestone for China’s ambitions to achieve self-reliance in artificial intelligence (AI) amidst tightening restrictions from the United States on advanced semiconductor exports.

The Institute of AI at China Telecom revealed that their open-source TeleChat2-115B and another, unnamed model were trained using tens of thousands of domestically produced chips. This achievement demonstrates China’s strides towards chip independence, a vital move as the country faces increasing challenges in accessing foreign semiconductors, especially AI chips like those produced by NVIDIA.

The statement from China Telecom highlights the importance of this development, which they see as a turning point for innovation and self-sufficiency in AI technology. According to the company, the unnamed AI model was trained with a whopping 1 trillion parameters, a crucial metric in machine learning that indicates the complexity and performance of an AI system. Their TeleChat2-115B model, by comparison, has over 100 billion parameters, showing China’s increasing sophistication in AI model development.

China catching up in AI
With Washington’s export restrictions limiting access to advanced chips, China has been working hard to find alternatives and keep its AI sector competitive globally. While China Telecom has not disclosed the specific suppliers of its domestic chips, the company has previously mentioned developing AI technology using Huawei’s Ascend chips. Huawei, the Shenzhen-based tech giant, has been ramping up its efforts in AI chip development and recently began testing its new Ascend 910C processor with potential Chinese clients.

This collaboration between companies like Huawei and China Telecom illustrates the push to reduce reliance on US-made chips, such as NVIDIA’s powerful A100 and H100, which are critical for training AI models but are now largely inaccessible due to US sanctions. Huawei’s Ascend chips have emerged as a leading alternative to NVIDIA’s hardware, particularly for local AI firms that need to develop cutting-edge technology despite the imposed restrictions.

Quick Reads

View All

OpenAI launches ChatGPT shopping assistant to scour web for best match and deal

Meta accused of burying internal research showing Facebook, Instagram harm mental health

In addition to Huawei, China Telecom has also integrated chips from domestic start-up Cambricon into its AI training infrastructure, specifically within its State Cloud service. The increasing collaboration between these companies signals a growing ecosystem of AI development in China that is aimed at bypassing the constraints imposed by foreign chip suppliers.

US’ chip restrictions and China’s AI push
US sanctions have significantly disrupted China’s access to NVIDIA’s top-tier AI chips, leading to a black market for these components. Some Chinese firms, however, are opting for lower-powered alternatives in order to maintain support and services from NVIDIA. While alternatives like Huawei’s Ascend chips are gaining traction, the broader AI industry in China still faces challenges in meeting the high demands of AI model training and deployment.

China’s progress in AI, as illustrated by China Telecom’s recent achievements, underscores the country’s commitment to fostering innovation and reducing dependency on foreign technology. As the AI race continues, it is clear that China is taking significant steps towards creating a more self-sufficient tech ecosystem, positioning itself to remain competitive in the global AI landscape despite geopolitical obstacles.