Huawei AI

Huawei Technologies is a Chinese technology company founded in 1987 whose Ascend line of artificial intelligence chips, CloudMatrix computing systems and Pangu industry models have made it the most significant non-Western competitor in AI infrastructure and a central pillar of China's push for technological self-sufficiency. After United States sanctions cut the company off from advanced American technology in 2019 and 2020, Huawei responded by building an integrated stack of chips, software and cloud services — a pivot that has placed it at the centre of debates about the future of the global AI supply chain.^[1][2]

History

Huawei was founded in 1987 in Shenzhen by Ren Zhengfei and grew from a reseller of telephone switches into one of the world's largest suppliers of telecommunications equipment, held largely through an employee shareholding scheme. In May 2019 the United States added Huawei to its Entity List, restricting its access to American components; further rules in 2020 severed its access to chips manufactured with US technology, ending work with the foundry TSMC and cutting into its smartphone business.^[1]

The company's AI work predates those sanctions. It unveiled the Ascend 310 and Ascend 910 AI processors in 2018 and 2019 and open-sourced the MindSpore deep-learning framework, paired with the Atlas 900 cluster that Huawei claimed as the world's fastest AI training system at its 2019 launch. The Pangu family of large models arrived in the early 2020s, with industry-focused editions used in power grids, mining and manufacturing, and a weather-forecasting model that drew international attention in 2023.^[1][2]

Facing continuing restrictions, Huawei accelerated its self-reliance programme. In the first quarter of 2025 it launched the Ascend 910C processor, and in September 2025 it presented a multi-year roadmap of new AI chips and rack-scale "supernode" systems, positioning itself openly as a full-stack alternative to Nvidia.^[1][2]

Key technologies

The CloudMatrix 384 system, introduced in 2025, is Huawei's flagship AI computing design: it links 384 Ascend 910C chips through an all-optical interconnect, giving the system raw throughput — roughly 300 petaflops of BF16 compute — that exceeds Nvidia's GB200 NVL72 on several measures. Analysis by the research firm SemiAnalysis found that Huawei compensates for less advanced individual chips by scaling systems aggressively and drawing on China's abundant power supply, at the cost of significantly lower performance per watt. Supply remains a constraint: production of the 910C has relied on stockpiled high-bandwidth memory, with domestic fabrication capacity central to future volumes.^[2]

Around the hardware, Huawei sells a full software stack: the CANN computing architecture, the MindSpore framework for training and the Huawei Cloud platform, one of China's largest, which offers model-development services and hosts Ascend-based compute for external customers. The Pangu models run on the same stack and are marketed for industrial deployment rather than general consumer chat.^[1][2]

Applications and impact

Huawei's systems occupy a unique position in the AI industry. For China, they offer an alternative to hardware that US export controls have made difficult to obtain, and Chinese model developers and state-linked projects have increasingly adopted Ascend-based compute. Internationally, CloudMatrix demonstrated that systems-level engineering — fast interconnects, dense racks, abundant power — can partially offset chip-level disadvantages, a lesson watched closely by governments weighing dependence on Nvidia. At the same time, Huawei remains on the US Entity List, and Washington has continued to debate and tighten export rules in response to China's progress.^[1][2]

>See Also

• AI Chip Export Controls
• China AI
• DeepSeek
• ByteDance AI
• National AI Office Malaysia