Over 1,000 sets of Superpods have already been deployed... Observation: “New product performance is 3 to 4 times that of existing products”
(Shanghai = Yonhap News) Correspondent Cha Byeong-seop = Chinese communications equipment company Huawei announced that it will accelerate the launch of its next-generation AI chip earlier than originally planned, citing the computing field as the core of its future artificial intelligence (AI) strategy.
According to Guachang Daily and Bloomberg News on the 17th (local time), Huawei's rotating chairman Wang Tao announced at the '2026 Huawei Connect Competition' held in Shanghai that day, "The Ascend 960 chip has been prepared ahead of time and has achieved twice the performance (compared to existing products)."
At the same time, the release plan for the Ascend 960, which was originally proposed for the fourth quarter of next year, has been accelerated and explained that the 960DT for AI model training will be released in the first quarter of next year, and the 960PR for inference will be released in the third quarter of next year.
Ascend 970 is scheduled to be released in 2028, and Ascend 980 is scheduled to be released in 2029.
Previously, Huawei revealed its AI chip development blueprint for the next three years at this event last year and said, "We will release (new Ascend products) once a year and double the computing power with each release."
Bloomberg said that this announcement reflects Huawei's confidence and that Huawei has been increasing disclosures related to hardware development for data centers since last year. This is different from the past, when secrecy was maintained regarding production plans.
China views the US sanctions on the high-tech semiconductor sector as a problem strangling the country and places importance on its will to solve the problem through self-reliance and self-strengthening, and Huawei is assessed to be at the forefront of this problem.
This year, Huawei proposed 'Tau's Law' to overcome the limitations of 'Moore's Law', and is also focusing on the AI computing supernode system 'Atlas 950 Superpod', saying it will overcome the limitations of individual chip performance by bundling chips together.
This is a technology that bundles up to thousands of AI chips into one computing pool. China is at a disadvantage in terms of individual chips due to the United States' checks, but the plan is to overcome this through its superiority in terms of the system that bundles the chips.
Chairman Wang Circular said on this day, "More than 1,000 sets of Atlas Superpods have already been deployed, and we have about 370 customers. The Atlas 950 Superpod has already begun to be commercialized on a large scale."
He continued, “(New product) Atlas 960 Superpod will be the first in the industry to use near-package optics (NPO) (a technology that increases data transmission speed and energy efficiency),” and added, “Among these, the Atlas 960 liquid-cooled Superpod is scheduled to be released in the third quarter of next year.”
While the existing Atlas 950 Superpod connects up to 8,192 neural network processing units (NPUs), the Atlas 960 Superpod can connect 15,488 chips, and training and calculation performance are evaluated to have improved by three and four times, respectively, the South China Monetary Post (SCMP) reported.
Chairman Wang Circular believes that the 'unified bus' technology, which makes information sharing between chips more efficient, will be a key link in the next generation of large-scale AI systems.
He said, “The core of Huawei’s AI strategy is computing,” and added, “Through the ‘Superpod Plus Cluster,’ we will lay a solid foundation for computing power in China and create new options for the world.”
Previously, Guo Ping, Chairman of Huawei's highest supervisory body, the Supervisory Board, also said at a recent meeting with new employees that Huawei is focusing on the connectivity and computing area and that "Huawei's goal in this area is to become NVIDIA."
