The Chinse tech giant plans to use larger interconnected systems to narrow the performance gap with Nvidia and shift LLM pre-training onto its own ecosystem.
By Wu Junyu
Tech giant Huawei will launch two new Ascend 960 AI chips in 2027 as it seeks to build a domestic computing ecosystem capable of supporting the training and inference of China’s increasingly large AI models.
Wang Tao, Huawei’s deputy chairman and rotating chairman, said at Huawei Connect 2026 on Sept. 17 that the company’s AI strategy centers on computing power and hardware monetization.
The Ascend 960DT, designed for AI training, is scheduled to launch in the first quarter of 2027, while the Ascend 960PR, aimed at inference, is expected in the third quarter. Caijing has learned that both chips have been undergoing laboratory testing and validation for several months.
Wang also disclosed a longer-term roadmap under which Ascend chips will move to an annual upgrade cycle. Ascend 970 is due in 2028, followed by Ascend 980 in 2029. Computing performance is expected to double with each generation, while memory capacity, memory bandwidth and chip-to-chip interconnect bandwidth will also increase.
Looking beyond single-chip performance
On paper, the Ascend 960DT and 960PR will outperform Nvidia’s H200 SXM but trail Nvidia’s B300. The H200 began large-scale shipments in the second half of 2024, while B300 shipments started in the second half of 2025, leaving Huawei’s new chips roughly two to three years behind Nvidia in performance.
Huawei is seeking to offset that gap by connecting much larger numbers of chips into a single high-performance “supernode” system.
The previous-generation Ascend 950 series supported supernodes of up to 1,024 cards. The Ascend 960 will increase that to 4,096 cards. The approach links large numbers of AI chips — previously dispersed across multiple servers—into a single system of hundreds or thousands of cards, improving inter-chip interconnect efficiency and reducing performance loss.
Nvidia’s current NVL72 system connects 72 GPUs in a single rack, while its planned NVL576 architecture will link 576 GPUs across eight racks into a high-speed interconnect system.
Huawei’s disclosed simulation data show the potential impact of this approach. In a 100,000-card cluster, a system built around 4,096-card supernodes could achieve a model FLOPs utilization roughly 2.75 times that of a conventional cluster comprising about 12,500 eight-card servers.
Ascend gains ground in China
Nvidia’s most advanced AI chips have had no compliant route into the Chinese market since the second half of 2025, following multiple rounds of U.S. export controls. That has created more room for domestic AI chipmakers.
According to international market research firm IDC, China shipped about 4 million AI accelerator cards in 2025, with domestic chips accounting for 41%, or more than 1.6 million units. Huawei shipped several hundred thousand Ascend cards, the largest volume among domestic AI chipmakers, IDC data from June 2026 showed.
Ascend 910C and the Ascend 950DT and 950PR are among the main domestically developed AI chips shipping in China in 2026. A number of Chinese technology companies, including ByteDance, Tencent, Ant Group and Meituan, have purchased Ascend chips. Some of their large-scale AI models, including ByteDance’s LongCat 2.0, have been trained on Ascend hardware.
Wang said in a group media briefing on Sept. 17 that the Ascend 950 series is already being used at scale for both model training and inference.
Huawei is also working with leading Chinese AI model developers and internet companies to prepare for the Ascend 960, including training, inference and adaptation of AI operators.
Wang said he expects some Chinese open-source AI models to gradually shift to pre-training directly on Ascend 950 and Ascend 960 chips from 2026 onward.
Source:
Caijing Magazine
