Huawei Aims to Dethrone Nvidia with Ambitious AI Chip Roadmap
Shenzhen, Friday, 19 September 2025.
Huawei is stepping up its game in the AI chip market. The company just revealed its roadmap for the Ascend AI chips, directly challenging Nvidia’s dominance, especially in China. This move occurs as U.S. regulations limit Nvidia’s access to the Chinese market, opening doors for Huawei. Huawei plans to release the Ascend 950PR in early 2026, followed by subsequent chips through 2028. Huawei’s new SuperPoD technology can link 15,000 graphics cards. This poses a significant threat to Nvidia’s revenue in the region.
Huawei’s ascend ai chip roadmap
Huawei is set to release a series of Ascend AI chips over the next few years [1]. The Ascend 950PR is slated for release in the first quarter of 2026 [1][3][4]. The Ascend 950DT will follow in the fourth quarter of 2026 [1][3]. The Ascend 960 is planned for the fourth quarter of 2027 [1][3]. The Ascend 970 is expected in the fourth quarter of 2028 [1][3]. These chips signify Huawei’s commitment to advancing its AI capabilities [4].
superpod interconnect technology
Huawei’s SuperPoD interconnect technology can link as many as 15,000 graphics cards, incorporating Huawei’s Ascend AI chips, to increase computing power [2]. This technology rivals Nvidia’s NVLink infrastructure, which facilitates high-speed communication between AI chips [2]. Huawei’s rotating chair, Eric Xu, stated that the Atlas 950 SuperPoD would surpass its counterparts in all major metrics for the next few years [1]. The Atlas 950 SuperPoD has a computing scale of 8192 cards and is expected to launch in the fourth quarter of this year [3].
competitive implications for nvidia
The U.S. ban on Nvidia hardware for Chinese tech companies creates a significant opportunity for Huawei [1][2]. Huawei’s ability to cluster its AI chips together provides users with more computing power for training and scaling AI systems [2]. While Huawei’s chips may be less powerful individually, the SuperPoD technology allows Huawei to compete effectively with Nvidia by offering large-scale computing solutions [2]. China’s push for domestic AI chip alternatives strengthens Huawei’s position [7].
technical specifications and comparisons
The Ascend 950PR and 950DT will feature an upgraded SIMD/SIMT architecture [3]. They will achieve computing power of 1 PFLOPS (FP8) and 2 PFLOPS (FP4), respectively [3]. These chips support various data formats and offer a 2 TB/s interconnect bandwidth [3]. The Ascend 950PR will have 128 GB of memory and 1.6 TB/s memory bandwidth. The Ascend 950DT will feature 144 GB of memory and 4 TB/s memory bandwidth [3]. Huawei’s Ascend 910C has a computing power of 800 TFLOPS (FP16), while Nvidia’s Blackwell B300 has approximately 3840 TFLOPS under the same standard [3].
market analyst perspective
Citic Securities anticipates rapid iteration of domestic computing power represented by Ascend [7]. They suggest focusing on investment opportunities within the Huawei Ascend supply chain [7]. The development of the Atlas 960 SuperCluster, supporting a 15488 Ascend card configuration, highlights Huawei’s focus on large-scale computing solutions [3][7]. Huawei’s strategic emphasis on ‘super nodes + clusters’ aims to meet the increasing demand for computing power within China, despite limitations in chip manufacturing [3].