GTHT released a research report stating that domestic AI computing power is expected to continue advancing, recommending attention to domestic AI computing targets. According to reports, Huawei unveiled its latest Ascend AI chip roadmap, with plans to launch the new Ascend 950PR chip in Q1 2026 and the Ascend 970 in Q4 2028. Additionally, Huawei comprehensively opens its SuperNode technology. First, it opens the LingQu protocol and SuperNode reference architecture, allowing the industry to develop related products or components based on technical specifications. Second, it fully opens SuperNode basic hardware, including NPU modules, air-cooled blades, liquid-cooled blades, AI standard cards, CPU motherboards, and cascade cards in various hardware forms, facilitating incremental development by customers and partners to design various LingQu-based products.
GTHT's main viewpoints are as follows:
Huawei unveiled its latest Ascend AI chip roadmap, planning to launch the new Ascend 950PR chip in Q1 2026 and the Ascend 970 in Q4 2028. ①The roadmap shows that Huawei launched the Ascend 910C in Q1 2025. Subsequently, it will launch the new Ascend 950PR chip in Q1 2026 and the Ascend 950DT in Q4. In Q4 2027, Huawei will launch the Ascend 960 chip, and in Q4 2028, the Ascend 970 chip. ②Starting from the Ascend 950PR, Ascend AI chips will adopt Huawei's self-developed HBM. The Ascend 950 will feature self-developed HBM HiBL1.0, while the Ascend 950DT will upgrade to HBM HiZQ2.0. ③The Ascend 970 offers computing power of 4PFLOPS(FP8)/8PFLOPS(FP4) with interconnect bandwidth of 4TB/s. HBM memory capacity is 288GB with bandwidth of 14.4TB/s. For comparison, NVIDIA's Blackwell Ultra GB300 provides computing power of 15PFLOPS(FP4), equipped with 288GB HBM3e and bandwidth of 8TB/s.
The world's most powerful SuperNode Atlas 950 SuperPoD is expected to launch in Q4 2025. ①As of September 18, 2025, over 300 CloudMatrix 384 SuperNodes have been cumulatively deployed, serving 20+ customers. ②Huawei launched the world's most powerful SuperNode Atlas 950 SuperPoD, expected to go to market in Q4 2025. Additionally, the next-generation product Atlas 960 SuperPoD is expected to launch in Q4 2027. ③Huawei's latest SuperNode products Atlas 950 SuperPoD and Atlas 960 SuperPoD support 8,192 and 15,488 Ascend cards respectively, leading comprehensively in card scale, total computing power, memory capacity, and interconnect bandwidth indicators, making them the world's most powerful computing SuperNodes for years to come. ④Based on SuperNodes, Huawei simultaneously released the world's most powerful SuperNode clusters, namely Atlas 950 SuperCluster and Atlas 960 SuperCluster, with computing scales exceeding 500,000 cards and reaching one million cards respectively, truly deserving the title of world's most powerful computing clusters. ⑤Huawei also released the world's first general computing SuperNode TaiShan950 SuperPoD, developed based on Kunpeng 950, planned for launch in Q1 2026.
Open hardware and open-source software jointly build full-scenario computing foundation. ①Huawei comprehensively opens SuperNode technology. First, it opens the LingQu protocol and SuperNode reference architecture, allowing the industry to develop related products or components based on technical specifications. Second, it fully opens SuperNode basic hardware, including NPU modules, air-cooled blades, liquid-cooled blades, AI standard cards, CPU motherboards, and cascade cards in various hardware forms, facilitating incremental development by customers and partners to design various LingQu-based products. ②Operating system LingQu components will all be open-sourced, with component code gradually merged into multiple upstream operating system open-source communities such as openEuler. Users can integrate partial or complete source code into existing operating systems according to actual needs, independently iterate and maintain versions, or directly merge entire components into existing operating systems.
Risk warnings: Technology iteration may fall short of expectations; ecosystem development may be slower than expected; downstream demand may fall short of expectations.