After years of strategic silence, Huawei has detailed its roadmap for computing chips and systems, for the first time positioning itself so openly as a direct competitor to global AI market leader Nvidia.
The plan is ambitious and involves not only new generations of chips, but also self-sufficiency in key areas such as HBM memory.
The Chinese tech giant, operating under the conditions imposed by US sanctions, has unveiled a comprehensive development strategy to make it independent of foreign suppliers and technology. At the heart of this strategy is the Ascend line of AI accelerators.
The currently available Ascend 910C chip will get three successors in the next three years: the Ascend 950, 960 and 970. The first of these, the 950 model, is expected to debut as early as the first quarter of next year in two variants.
The 950PR version will be optimised for inference tasks and recommender systems, while the 950DT will focus on decoding inference and language model training.
Each successive generation is expected to bring a significant leap in performance – the Ascend 960 will offer twice the processing power and memory capacity of its predecessor.
Manufacturing remains a key challenge. Relegated to working with the market leader, Taiwan’s TSMC, Huawei has not officially named its partners. However, industry analysts agree that manufacturing will be handled by China’s largest chipmaker, SMIC, raising questions about production scale and energy efficiency in the most advanced technology processes.
One of the biggest surprises is Huawei’s announcement of its own proprietary high-bandwidth memory (HBM) technology, challenging the dominance of companies such as SK Hynix and Samsung. The Ascend 950PR chip will be integrated with HiBL 1.0 memory (128 GB, 1.6 TB/s), while the Ascend 950DT will receive an even more powerful HiZQ2.0 module (144 GB, 4 TB/s).
Huawei is also betting on the power of scaling. Instead of competing at the single chip level, the company is building powerful cluster systems. The announced Atlas 950 SuperPod supercomputer, comprising 8,192 Ascend 950DT chips, is expected to offer 6.7 times more computing power and 15 times more memory capacity than the NVL144 system that Nvidia plans to launch in 2026.
Further plans include an even more powerful Atlas 960 SuperPoD, using nearly 15,500 accelerators.
The strategy is complemented by the development of the Kunpeng series of general purpose processors. New generations, the Kunpeng 950 and 960, will arrive in 2026 and 2028 respectively, strengthening the company’s hardware ecosystem.
The revealed roadmap is a clear signal that Huawei does not intend to be just a passive observer of the AI revolution, but wants to become one of its key architects, building on its own independent technologies.
