Wang Tao, Huawei Technologies’ rotating chairman, is set to unveil the Ascend 960 artificial-intelligence chips at the company’s annual summit in Shanghai today. The chips are scheduled to become commercially available in 2027, in a move that reinforces Huawei’s efforts to compete with Nvidia in China and global markets amid US export controls.
Plan to run artificial-intelligence models
Huawei aims to enable Ascend chips to run all artificial-intelligence models in China and abroad, Guo Ping, the company’s supervisory board chairman, told new employees at an internal meeting earlier this month.
We are narrowing the gap through innovation in chip-building architecture. Apple was not initially a leader in processor manufacturing, but it delivered an exceptional user experience by redesigning its architecture and improving its software.
DeepSeek plans to use at least 160,000 Huawei Ascend 950DT chips at a data center it is building in Inner Mongolia to run its models. Huawei is also exploring overseas markets, including Malaysia and Egypt.
The Chinese company, whose businesses include telecommunications equipment and smartphones, is leading Beijing’s drive for self-sufficiency in critical technologies, alongside years of US restrictions on China’s chip sector. It is also participating in Beijing-backed efforts to export China’s technology ecosystem overseas.
Chip road map extends to 2028
The Ascend 960 will replace Huawei’s current Ascend 950 family of artificial-intelligence accelerators, which includes the 950DT used by DeepSeek and equipped with faster, higher-capacity memory to handle demanding AI-inference workloads.
The company plans to launch another upgraded version, the Ascend 970, in 2028 as part of a three-year campaign to compete with Nvidia. Under the company’s plan, each new generation of chips will double computing power compared with its predecessor.
Huawei’s artificial-intelligence chips, however, still lag Nvidia’s top products. Leading Chinese laboratories, including DeepSeek, continue to rely heavily on the US company’s chips to train models before running them on a growing number of locally produced semiconductors. Nvidia’s advanced chips also reach China through smuggling in limited quantities, while its most advanced Blackwell processors are not directly available in the Chinese market.
Production constraints and rising prices
Capacity constraints are limiting Huawei’s ability to meet accelerating demand for computing power in a market that Morgan Stanley estimates could reach 646 billion yuan ($96 billion) by 2030. The company recently raised the price of its Ascend 950DT chips 60%, citing component supply shortages in communications with customers.
Huawei is seeking to address its shortcomings through methods including LogicFolding, a technology it recently announced to improve chip technology without access to the best equipment. It has also introduced the SuperPod architecture, which will soon enable up to 100,000 Ascend chips to be linked in a single computing cluster, with faster data transfer through the UnifiedBus protocol.
System-level competitiveness has become more important than ever.
Chan said the actual gap is narrowing through multi-die design, advanced packaging and architectural design for systems spanning computing server racks and optical networks, as well as the co-optimization of software and hardware.
In-house-designed memory
The Ascend 950 chips are Huawei’s first artificial-intelligence accelerators to use memory chips designed in-house. This follows Huawei and other Chinese AI-chipmakers being cut off from the latest AI-memory products made by South Korea’s SK Hynix and Samsung Electronics and the US’ Micron Technology under US export controls.
