China Telecom trains AI model with Chinese-made chips

China Telecom, a state-owned carrier, has developed two LLMs that were entirely trained on domestically produced processors.

This breakthrough demonstrates China’s progress toward chip independence, particularly in the field of AI.

The company’s AI Institute announced that its open-source model, TeleChat2-115B, along with another undisclosed model, was trained on tens of thousands of Chinese chips. This is especially significant given the US’s increasing restrictions on China’s access to cutting-edge chips, such as Nvidia’s latest AI processors.

According to the AI Institute, this achievement reflects China’s complete self-sufficiency in training its own LLMs, which power tools like OpenAI’s ChatGPT. They also regard it as the beginning of a new chapter in the country’s pursuit of AI technology innovation and independence.

China Telecom revealed that the unnamed model features one trillion parameters. In layman’s terms, parameters are the variables that an AI model utilises to learn during training; the more parameters the model has, the smarter and more powerful it is. TeleChat2-115B, the other model, has over 100 billion parameters.

Chinese companies are working hard to stay competitive in AI, especially given Washington’s export restrictions. While China Telecom didn’t specify exactly who provided the chips, the South China Morning Post reported that it has previously mentioned Huawei’s Ascend chips as part of its LLM development.

Huawei, which is under US sanctions, has lately begun testing a new AI processor with potential clients in China. Companies are seeking alternatives to Nvidia’s high-end hardware, and Huawei appears to be stepping in to fill the gap.

According to sources, Huawei has already offered samples of its Ascend 910C processor to major Chinese server companies for testing. The chip is also being made available to Chinese internet companies, many of which have long relied on Nvidia’s technology.

For local AI companies, Huawei’s Ascend chips and training solutions are emerging as viable alternatives to Nvidia’s offerings. In addition to Huawei, China Telecom is offering processors from Cambricon, a Chinese start-up, through its State Cloud service.

These collaborations are likely to help China minimise its reliance on foreign chips such as Nvidia’s GPUs, which have long been the preferred option for training AI models.

With US export regulations limiting access to advanced Nvidia GPUs such as the A100 and H100, several Chinese AI firms have resorted to black markets. However, many opt to use lower-powered alternatives to ensure that they can continue to enjoy Nvidia’s official support and services.

See also: US examines security risks posed by China Telecom and China Mobile’s operations

Want to learn more about cybersecurity and the cloud from industry leaders? Check out Cyber Security & Cloud Expo taking place in Amsterdam, California, and London. Explore other upcoming enterprise technology events and webinars powered by TechForge here.

Tags: AI, China, Huawei, NVIDIA