Cambricon aims to triple chip output to replace Nvidia in China

[BEIJING] Cambricon Technologies plans to more than triple its production of AI chips in 2026, aiming to wrest market share from Huawei Technologies in China and fill a void left by Nvidia’s forced exit.

The Beijing-based company is preparing to deliver half a million artificial intelligence accelerators in 2026, people familiar with the matter said. That includes as many as 300,000 units of its most advanced Siyuan 590 and 690 chips, the people said, asking to remain anonymous discussing private targets.

The company will rely primarily on Semiconductor Manufacturing International Corp’s (SMIC) latest production process, known as “N+2” 7-nanometre, the people said.

The ramp-up at Cambricon underscores the rapid ascent of Chinese chipmakers after Beijing began actively discouraging the use of Nvidia’s product this year, part of a longer-term effort to wean the country off US technology.

Huawei is also preparing to double the output of its most advanced artificial intelligence chips over the next year. And up-and-comer Moore Threads Technology debuts on Friday (Dec 5) in Shanghai, showcasing its own ambitions to carve out a slice of the market.

Nvidia boss Jensen Huang said in November his company is effectively blocked from China, which would spur the rise of more domestic competition from the likes of Huawei. And while the Trump administration is considering a plan to allow the sale of its H200 cards, there’s no guarantee Beijing won’t also hinder its adoption.

Few companies have benefited as visibly from that situation as Cambricon, which reported a 14-fold surge in its revenue in the September quarter – and a nine-fold leap in market value since 2021.

It is now on track to win new orders from some of China’s biggest AI spenders including Alibaba Group Holding in coming years, the people said. The chip designer already counts ByteDance as a primary customer, which accounts for more than 50 per cent of all Cambricon’s orders right now, the people said.

Alibaba, ByteDance, Cambricon and SMIC did not respond to emailed requests for comment.

Whether Cambricon will hit those targets depends in large part on not just the pace of AI development, but also its ability to secure capacity at SMIC at a time Huawei and other rivals are also vying to place orders with China’s most advanced chipmaker.

For context, Cambricon will build just 142,000 AI chips this year, Goldman Sachs estimates.

SMIC’s own technology may prove an obstacle. When it comes to Cambricon’s top-of-the-line 590 and 690 chips, the company is for now managing yields of just 20 per cent, the people said.

That means about four out of five silicon dies – the basic components of a full chipset – are considered flawed and unusable. The top global contract chipmaker Taiwan Semiconductor Manufacturing Co now has an estimated yield of at least 60 per cent with its latest two-nanometre process, which is three generations or seven years ahead of SMIC’s technology, according to some analysts.

Another potential bottleneck is the supply of the high-bandwidth memory chips required to make AI accelerators. That technology remains a challenge for Chinese companies, which is why Huawei’s latest 910C AI accelerators still rely on memory chips from SK Hynix and Samsung Electronics

“Though ramping up production at Huawei, Baidu and Cambricon has led to a sizable increase in the supply of domestic AI accelerators in China, the availability of chips remains insufficient to meet surging demand. This imbalance is exacerbated by low production yields on advanced 7-nm chips from SMIC, which likely will remain a key bottleneck for the foreseeable future,” said Robert Lea, analyst at Bloomberg Intelligence.

“Despite these challenges, the direction remains clear. Geopolitical tensions, combined with the ongoing uncertainty regarding the availability of Nvidia’s chips, will continue to drive China’s AI firms towards domestic chip suppliers. The divergent strategies of China and the US will ultimately lead to a bifurcated global AI ecosystem.

Still, Cambricon’s industry prominence suggests it is got negotiating leverage with partners like SMIC.

Founded in 2016 by computer scientist Chen Tianshi, Cambricon is regarded as one of the biggest beneficiaries of Beijing’s drive for self-reliance in AI and chipmaking – an effort that could draw US$98 billion in combined government and corporate spending this year alone.

Huawei and Cambricon are just the largest in a rapidly expanding industry that includes fast-rising rivals such as Hygon Information Technology and startups like Moore Threads and MetaX.

In the long run, the supply of domestic AI chips will be critical for Chinese players from DeepSeek to Alibaba hoping to compete with the likes of OpenAI and Alphabet’s Google.

President Xi Jinping and the Communist Party have criticised US efforts to cut off the supply of American AI chips into China, and have explicitly said they want to build up domestic capabilities to minimise their vulnerabilities to Washington’s political priorities. BLOOMBERG