Huawei and DeepSeek take a big step forward hand in hand.

On September 18th, at the Huawei Connect Conference (HC Conference) held in Shanghai, Xu Zhijun, the rotating chairman of Huawei, mentioned the national craze triggered by DeepSeek at the beginning of the year as soon as he took the stage.

"From the Spring Festival this year to April 30th, after the coordinated efforts of multiple teams, we finally made the inference capabilities of Ascend 910B/910C meet the basic needs of customers," Xu Zhijun said. After the emergence of DeepSeek, many government agencies and central state - owned enterprises responded and connected to it. As a computing power provider, Huawei also had to follow up and respond.

Since Huawei first released the Ascend 310 chip in 2018 and the Ascend 910 chip in 2019, it has continuously invested in the R & D and innovation of AI basic computing power. Although the model pioneered by DeepSeek has significantly reduced the computing power requirements, Xu Zhijun believes that for the development towards AGI and physical AI, computing power has been and will continue to be the key to artificial intelligence in the future, according to Huawei.

1. Huawei Releases Multiple Chip Products with Plans Set Until 2028

Xu Zhijun announced that for the future, Huawei has planned three series of Ascend chips, including Series 950, 960, and 970.

Among them, the Ascend 950 series includes two chips: 950PR and 950DT. The 950PR will be launched in the first quarter of 2026, and the 950DT will be launched in the fourth quarter of 2026.

The Ascend 960 chip will be launched in the fourth quarter of 2027, and the Ascend 970 chip is expected to be launched in the fourth quarter of 2028.

Release plan of Huawei Ascend chips; Photo taken by the author

Compared with the previous generation, the Ascend 950 has achieved fundamental technological improvements in multiple aspects: it newly supports low - precision data formats such as FP8/MXFP8/HIF8 and MXFP4, with computing power reaching 1 PFLOPS and 2 PFLOPS respectively, greatly improving the efficiency of training and inference; it significantly enhances vector computing power and supports more fine - grained memory access; the interconnection bandwidth is increased by 2.5 times to reach 2TB/s; and it is equipped with self - developed HBM technologies HIBL1.0 and HIZQ2.0.

In the general computing field, Huawei has planned Kunpeng 950 and Kunpeng 960, which will be launched in the fourth quarter of 2026 and the first quarter of 2028 respectively, and will continue to evolve around supporting super - nodes, more cores, and higher performance.

In addition, Huawei officially released the interconnection protocol for super - nodes - Lingqu and opened the technical specification of Lingqu 2.0. Since the research started in 2019, Lingqu 1.0 has started commercial verification. Now, the opening of Lingqu 2.0 aims to invite the industry to develop relevant products and components based on Lingqu and jointly build an open ecosystem for Lingqu.

2. Release of the World's Most Powerful Computing Power Super - Node

Due to complex reasons such as international politics, Xu Zhijun also said bluntly at the press conference that the computing power performance of Huawei's single - chip is inferior to that of NVIDIA. "But Huawei has thirty years of accumulation in connection technology. Huawei's super - node computer can achieve the world's most powerful computing power and meet the huge global demand for AI training and inference."

The super - node (SuperPod) is an important trend in the development of intelligent computing at present. Xu Zhijun believes that a super - node is physically composed of multiple machines, but logically it learns, thinks, and reasons as one machine.

In terms of the specific progress of the super - node business, Huawei released the Atlas 950 SuperPoD and Atlas 960 SuperPoD. Among them, the Atlas 950 super - node based on the Ascend 950 chip supports a scale of 8192 cards, consists of 128 computing cabinets and 32 interconnection cabinets, covers an area of about 1000 square meters, has an FP8 computing power of 8EFlops, an FP4 computing power of 16EFlops, and an interconnection bandwidth of up to 16 PB, which is more than 10 times the current total global Internet bandwidth.

Display of the Atlas 950 SuperPoD released by Huawei; Photo taken by the author

The Ascend 950 super - node will be launched in the fourth quarter of 2026. Xu Zhijun emphasized that the Atlas 950 super - node will be the world's most powerful AI super - node from 2026 to 2028.

The other Atlas 960 super - node supports 15488 cards, consists of 176 computing cabinets and 44 interconnection cabinets. Its computing power, memory, and bandwidth are doubled again on the basis of the Atlas 950. It is planned to be launched in the fourth quarter of 2027.

Xu Zhijun especially mentioned that the value of super - nodes is not limited to traditional business fields such as manufacturing, communication, and computing. It also plays an important role in the recommendation systems widely used in the Internet industry. Huawei can build a hybrid super - node based on the TaiShan 950 and Atlas 950, creating a new architectural direction for the next - generation deep recommendation system.

However, although large - scale super - nodes greatly improve the intelligent computing and general computing capabilities, there are still immature aspects in the interconnection technology.

For example, how to achieve reliable interconnection for a scale of 8192 cards or even 15488 cards is a technical problem that the industry urgently needs to solve. At present, many super - node solutions released by the industry have failed to achieve large - scale deployment. The core bottleneck is not the chip itself, but the immaturity of the interconnection technology, which is specifically reflected in two challenges:

One is how to achieve long - distance and high - reliability. Large - scale super - nodes have many cabinets, and the connection distance between cabinets can be as long as 1000 to 2000 meters. The current electrical interconnection technology has limited distance in high - speed signal transmission and can only support the interconnection of two cabinets at most; while the optical interconnection technology can meet the long - distance connection requirement, it cannot achieve the high - reliability required by a single computer system.

Two is how to achieve ultra - large bandwidth and ultra - low latency. Currently, there is a more than 5 - fold gap between the inter - cabinet card - to - card interconnection bandwidth and the requirements of super - nodes. The best latency can only reach about 3 microseconds, still with a 24% gap from the design goals of Atlas 950/960. When the latency is approaching the physical limit, every 0.1 - microsecond improvement is extremely challenging.

Xu Zhijun elaborated on two solutions.

Huawei's technical accumulation at the super - node level; Photo taken by the author

First, to solve the problem of long - distance and high - reliability, Huawei has introduced high - reliability mechanisms at each layer of the interconnection protocol, including the physical layer, data link layer, network layer, and transport layer; at the same time, it has introduced nanosecond - level fault detection and protection switching in the optical path, making the application unaware when there is a flash - off or failure of the optical module; moreover, Huawei has re - defined and designed optical devices, optical modules, and interconnection chips. These innovations and designs have increased the reliability of optical interconnection by 100 times, with an interconnection distance of more than 200 meters, achieving both the reliability of electricity and the long - distance advantage of optics.

Second, to solve the problem of large - bandwidth and low - latency, Huawei has broken through the multi - port aggregation and high - density packaging technologies, as well as the equal architecture and unified protocol, achieving a TB - level ultra - large bandwidth and an ultra - low latency of 2.1 microseconds.

"It is because of a series of systematic and original technological innovations that we have overcome the super - node interconnection technology, met the interconnection requirements of high - reliability, all - optical interconnection, high - bandwidth, and low - latency, and made large - scale super - nodes possible," Xu Zhijun said.