US-sanctioned Chinese chip start-up Moore Threads unveils GPU-based computing centre, AI graphics card

In a fresh round of expanded US export controls, Moore Threads and Biren Technology, two of China’s top AI chip start-ups, were added to Washington’s so-called Entity List in October over concerns that they were hurting US national security and policy interests.

01:36

AI chip maker ordered by US government to halt exports to China

AI chip maker ordered by US government to halt exports to China

The US sanctions have made it challenging for the two companies to find wafer foundries that are willing to manufacture their chips, putting the firms in a similar predicament that Huawei Technologies has faced since late 2020.

The export restrictions also affect Moore Threads and Biren’s access to US electronic design automation software.

Weeks after it landed on the US trade blacklist, Moore Threads announced a round of lay-offs. At the time, Zhang, who worked with leading GPU company Nvidia for 15 years before starting his own venture, said the job cuts were a tough but necessary decision to consolidate company resources to focus on GPU development.

Moore Threads had been a darling of investors amid China’s chip self-sufficiency drive.

The company had raised a total of US$525.7 million as of December last year, including 1.5 billion yuan (US$205.4 million) from a Series B funding round led by Hexie Health Insurance and Hechuang Digital Private Equity Fund Management that gave it a pre-market valuation of 28.95 billion yuan, according to data from private equity market tracker PitchBook.

Moore Threads did not elaborate on which foundry produced its latest graphics cards. The company did not respond to a request for comment on Wednesday.

However, it has said that the KUAE computing facility only took 30 days to build and is compatible with Nvidia’s CUDA, the parallel computing platform and programming model developed by the US firm for general computing using its GPUs.

Moore Threads said its centre can finish training a 130-billion-parameter model within 56 days.

Under updated tech export controls announced by the US Commerce Department in October, Nvidia can no longer sell to Chinese companies its A800 and H800 chips, which had been designed to comply with Washington’s earlier rules.

Why China is still several moves behind in the AI chess game started by ChatGPT

Analysts have said the US move could undercut China’s AI drive.

Owing to the export curbs and surging interest in developing large language models, Nvidia GPUs have become highly sought after by Big Tech companies in China.

Baidu, ByteDance, Tencent Holdings and Alibaba Group Holding, owner of the South China Morning Post, have each spent billions of dollars stocking up on the chips, and the firms are expected to buy a total of 125,000 Nvidia H100 GPUs – 20 per cent of all H100 shipments in 2023 – according to data from research firm Omdia.

Even smaller non-tech Chinese firms, including a landscape designer and a powdered flavour maker, have touted their use of Nvidia processors.

Moore Threads on Tuesday said it had teamed up with 15 partners – including 360 Security Technology, Baidu’s PaddlePaddle, JD.com’s Yanxi, NetEase, Tsinghua University and Fudan University – to establish a large model alliance.