State-owned China Telecom trains two major AI large language models entirely using domestic chips as country ramps self-sufficiency
A major Chinese state-owned telecommunications firm said it has trained two large language models (LLMs) entirely using domestic processors, an indication of the country’s progress in developing semiconductors that can be used for advanced AI tasks amidst US sanctions.
TikTok’s parent ByteDance was reportedly also planning to develop an AI model trained primarily with domestic chips from Huawei.
The Institute of AI at China Telecom, one of China’s main state-backed telecoms firms, said the open source TeleChat2-115B and a second unnamed model were trained using tens of thousands of domestically produced chips.
The achievement “indicates that China has truly realised total self-sufficiency in domestic LLM training” and marks a new phase in China’s innovation and self-reliance, the AI institute said in a statement.
AI training
The unnamed model has 1 trillion parameters, while TeleChat2t-115B has more than 100 billion parameters, China Telecom said.
The firm didn’t say who supplied the chips, but previously said it was using Ascend processors from US-sanctioned Huawei Technologies.
ByteDance, the Beijing-based parent of social media app TikTok, is reportedly planning to train an LLM using Huawei Ascend 910B chips, Reuters reported, citing unnamed sources.
The company already uses the chips for less intensive tasks involving pre-trained AI models making predictions, the report said.
The new model will reportedly be less powerful than ByteDance’s existing AI model Doubao, which has become one of the country’s most popular AI tools with 10 million monthly active users.
ByteDance has ordered more than 100,000 of the Ascend chips this year, but has received fewer than 30,000 as of July, preventing it from establishing a timeline for the new model, the report said.
Domestic chips
ByteDance is one of the largest buyers of Huawei’s AI chips, and is also the biggest purchaser of Nvidia’s H20 chip, designed specifically for the China market to comply with US sanctions, Reuters’ report said.
The firm is also Microsoft’s biggest client in Asia for Nvidia processing power available via cloud computing, according to the newspaper.
Huawei has begun offering samples of its new Ascend 910C AI processor to large Chinese server companies for hardware testing and configuration, the South China Morning Post reported last week.
The upgraded version of the 910B is being offered to large internet firms that are also major Nvidia customers, the paper said, citing unnamed sources.
The US began prohibiting the sale of advanced chips to Chinese firms in 2022 and has expanded sanctions to include more chips and chip manufacturing technology from US, Netherlands and Japanese suppliers.