Huawei Technologies has shown advanced AI hardware and spoken for the first time of its roadmap for artificial intelligence accelerator chips, as mainland China pushes for increased autonomy in AI technology.
The high-profile announcements came as representatives of China and the US met in Madrid, Spain to discuss trade issues.
Huawei said it had developed what it called the world’s most powerful supernode computing cluster using domestic chipmaking processes, as well as plans for its Ascend AI chip range over the next three years.
Compute clusters
Rotating chairman Eric Xu told the company’s annual Connect Conference in Shanghai that Huawei was using domestic supernode and clustering technology to meet “growing compute needs”.
He said Huawei would use its upcoming Ascend 950DT chips in an AI stack called Atlas 950 SuperPoD, which includes 8,192 cards and occupies 1,000 square metres with 160 cabinets, delivering 8 exaflops of FP8 computing power and 16 exaflops of FP4 power.
The system is planned to become available in the fourth quarter of 2026.
Furthermore, 64 Atlas 950 SuperPoDs can be combined into an Atlas 950 SuperCluster to deliver 524 exaflops of FP8 compute power, Huawei said.
By the end of 2027, Huawei said it is planning the Atlas 960 SuperPoD, powered by up to 15,488 Ascend 960 cards, and the Atlas 960 SuperCluster, aggregating millions of the cards.
Xu said Huawei is planning to follow its current Ascend 910C AI chip, which has been available since the first quarter of 2025, with 950PR and 950DT chips later this year.
Those chips are to incorporate high-bandwidth memory (HBM) technology developed by Huawei itself, he said.
High-bandwidth memory
HBM is one of the technologies restricted from sale to China under recent US export rules, due to its use in AI infrastructure.
The Ascend 960 and 970 are planned for 2027 and 2028 in a timeline running parallel to architectures expected from Nvidia and AMD.
Ren Zhengfei, founder and chief executive of Huawei, told local media earlier this year that although Chinese AI chips are a generation behind those in the US, partly due to limitations on available manufacturing technology, Huawei’s Ascend chips can achieve state-of-the-art computing power using stacking and clustering.
Currently Huawei has been selling its CloudMatrix 384, which is based on the Supernode384 architecture, that clusters 384 Ascend AI chips to deliver 300 petaflops of computing power and is considered to rival Nvidia’s NVL72.
Xu said more than 300 CloudMatrix 384 systems have been deployed to more than 20 clients.
‘Amazing resilience’
Paul Triolo, partner at Washington DC-based Albright Stonebridge Group, said a “healthy dose of scepticism” was necessary in evaluating Huawei’s plans, partly because they depend on the manufacturing capabilities of Chinese chip firm Semiconductor Manufacturing International Corporation (SMIC).
But he cited Huawei’s past successes in producing high-end technologies such as the 5G chips and processors powering its smartphones as evidence that the company has in the past found ways of working around constraints.
“If any company can accomplish the miracle of producing AI hardware that is at least close to Western leaders, it is Huawei,” Triolo said.
He noted that Huawei has demonstrated “amazing resilience” under US efforts to undermine its business model since 2019, when it was placed on a US blacklist.
