When Chinese artificial intelligence start-up DeepSeek unveiled an updated foundational model late in August, investors in Nvidia were stunned. Shares of the US chip giant slid, as market watchers grappled with news that the two-year-old start-up, which has developed models rivalling the world’s best, was shifting towards supporting domestically produced chips.
Adding to their concerns, DeepSeek was not alone in its endeavour to empower China’s AI ambitions without relying on US technology.
Last week, Huawei Technologies – the Chinese firm at the forefront of the nation’s tech self-sufficiency drive – showcased the latest offerings from its Ascend chip series and unveiled hardware designed to deliver world-class computing power without using Nvidia’s processors.
Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.
It marked the first time that the company had divulged details of its chip road map since it was blacklisted by the US in 2019 over national security concerns. As confidence in home-grown technology mounts, Beijing has urged the country’s tech giants to cease buying chips that Nvidia tailored for China, which were designed to comply with US export restrictions.
Together, Huawei and DeepSeek have emerged as symbols of China’s resilience, showing how US trade curbs have sparked a wave of innovation in the domestic AI industry that allowed Beijing to gain ground in its tech rivalry with Washington.
An AI sign at the annual Huawei Connect event in Shanghai. Photo: Reuters
With US restrictions acting as a push factor and China’s pursuit of self-sufficiency serving as a pull factor, “DeepSeek will look for alternatives for chips”, said Gary Ng, director and senior economist for Asia Pacific at Natixis Corporate and Investment Bank. “As the national champion, Huawei will play a major part.”
At the centre of attention is Huawei’s ambitious blueprint for its coming Ascend chips, which deputy chairman Eric Xu Zhijun described as “the foundation of Huawei’s AI computing strategy.”
Future generations of chips would double in compute power every year while supporting more data formats, enhancing usability and increasing bandwidth, the company said.
The Ascend 950PR, designed for pre-fill and recommendation tasks, is set to launch in the first quarter of next year, followed by the Ascend 950DT, optimised for decoding and training, in the fourth quarter. The Ascend 960 and 970 processors are expected to be released in the fourth quarters of 2027 and 2028, respectively.
Alongside the new chips, Huawei also introduced “the world’s most powerful” supernode computing clusters – the Atlas 950 and Atlas 960 SuperPoDs, as well as the Atlas 950 and Atlas 960 SuperClusters – which can aggregate up to a million Ascend neural processing units. NPUs are designed to accelerate AI tasks.
Those innovations were expected to help Huawei “circumvent the limitations in China’s chip manufacturing process”, Xu said.
“Huawei’s announcement shows that its AI accelerators are inching closer to Nvidia’s capabilities by leaning into its networking roots,” said Kevin Xu, founder of US-based Interconnected Capital.
These systems gave Huawei a lead among publicly available products in terms of card capacity, said Charlie Zheng, chief economist at Samoyed Cloud Technology Group, although Huawei still lagged “one or two generations behind” Nvidia in single-chip performance.
That performance gap could possibly be addressed by chip stacking, a method mentioned in remarks from Huawei founder and CEO Ren Zhengfei. In an interview with a state newspaper in June, Ren indicated that Ascend chips could reach state-of-the-art performance through techniques like stacking and clustering, despite being “one generation behind” US counterparts.
While stacking presents its own challenges – such as in power consumption, cabling and reliability – according to Zheng, Huawei’s advancing technology is drawing attention across the country.
Crucially, the Ascend chips are expected to enhance AI training for DeepSeek, which is pivoting to using more domestic chips for its AI models.
In a cryptic one-line social media post, DeepSeek said at the time that the UE8M0 FP8 format was specifically tailored “for next-generation home-grown chips to be released soon”. The announcement spurred speculation about the potential suppliers of those chips.
“The architecture is specifically designed to accommodate the hardware logic of Chinese chips, which enables a model to run smoothly on these hardware,” said Su Lian Jye, chief analyst at research firm Omdia.
While Huawei and DeepSeek have not disclosed whether Ascend chips have been or will be used in developing V3.1 and other DeepSeek models, many analysts and industry insiders anticipated that Huawei would be one of DeepSeek’s hardware suppliers.
Huawei’s coming computing clusters were expected to be integral to DeepSeek’s model training in the future, said Liu Jie, an engineer at a Shanghai-based graphics processing unit developer. He added that Huawei already had experience applying Ascend chips to the inference process of existing DeepSeek models, which would also help in training new models.
Huawei’s Xu highlighted the use of Ascend chips in running DeepSeek’s R1 reasoning model. “Between DeepSeek-R1’s January launch and April 30, our teams collaborated closely to ensure that the inference capabilities of our Ascend 910B and 910C chips met customer needs,” he said.
He underscored DeepSeek’s impact on Huawei’s chip initiatives: “After DeepSeek went open source, our customers began reaching out with various issues regarding Ascend, while expressing their hopes for the future.”
Liu added that DeepSeek was likely to adopt a combination of domestic chips to optimise training efficiency and minimise power consumption. Besides Huawei, Cambricon Technologies, MetaX and Moore Threads were potential suppliers, as their products supported DeepSeek’s FP8 format, according to Omdia’s Su.
Another notable player in this realm is Beijing-based Cambricon. Founded in 2016, the AI chipmaker has emerged as a stand-out in China’s AI industry, surging nearly 500 per cent over the past year to become one of the priciest stocks in China’s onshore market.
Encouraged by DeepSeek’s announcement and the chipmaker’s soaring revenue, investors are betting that Cambricon’s AI processors will replace Nvidia products in the country’s data centres.
In August, Cambricon reported a staggering 4,348 per cent year-on-year increase in first-half revenue to 2.88 billion yuan (US$404 million), a record since going public in 2020. The jump was fuelled by “continued market expansion and active support for the implementation of AI applications”, according to the company.
Huawei’s booth during the World Artificial Intelligence Conference in Shanghai in July. Photo: Reuters
China’s largest tech companies are also ramping up their efforts to adopt more domestic chips in their AI infrastructure.
Tencent Holdings said recently that its cloud computing unit had “fully adapted to mainstream domestic chips”, although it did not specify any particular brand.
Alibaba Group Holding’s semiconductor design unit T-Head developed a new application-specific AI chip, which was said to be on a par with Nvidia’s H20 processors, according to a report by state broadcaster China Central Television last week. Alibaba owns the Post.
Still, as China strives for full self-sufficiency in AI computing, challenges remain, particularly concerning software ecosystem efficiency and production capacity.
A key question looms: will China’s foundries such as Semiconductor Manufacturing International Corporation, which are also under US tech sanctions, be able to mass-produce Huawei’s Ascend chips at yields comparable to Taiwan Semiconductor Manufacturing Company, the world’s largest contract chip manufacturer, which ceased collaboration with Huawei in 2020?
“China’s 7-nanometre production lines are estimated to supply about 600,000 to 700,000 AI chips in 2026,” said Samoyed’s Zheng. “If [Huawei’s] million-card cluster comes to fruition, it will consume one-sixth of the annual production capacity, placing immense pressure on the yield and production capacity of domestic fabrication plants.”
“It is also an open question whether Huawei’s ‘supernode’ architecture can work seamlessly so that many chips can function smoothly as one computer, which is important for large-scale AI training,” Interconnected Capital’s Xu said.
Regardless of the challenges, Huawei is determined to forge ahead. “Our chip technology is currently one or two generations behind [Nvidia], and we cannot predict how many generations behind we’ll be in the future,” Xu said in a recent interview. “We have no choice but to find another way out.”
