Huawei AI Chips: Moving a Step Forward as Company Accelerates Next-Generation Ascend Roadmap and Expands Large-Scale AI Computing Systems. Huawei says the Ascend 960DT is now scheduled for Q1 2027, and the Ascend 960PR is expected for Q3 2027. The company also announced the Atlas 960E SuperPoD, a large AI computing system based on the Ascend 960 chips and near-packaged optics.
Introduction
The global AI industry is increasingly reliant on high-powered computing infrastructure. Not only are companies racing to build better AI models as they get bigger and AI agents demand more computing power, they are also racing to build the chips and systems needed to run them.
Huawei has now announced some new developments in its AI hardware strategy at Huawei Connect 2026 in Shanghai.
The announcement includes a faster roadmap for its next-generation Ascend AI chips and a new Atlas 960E SuperPoD to connect together thousands of AI processors into a larger computing system.
What is Huawei Ascend 960?
The Ascend 960 is Huawei’s next-generation family of AI accelerators.
Huawei says development has progressed faster than originally expected. The company has divided the next generation into 2 versions with different purposes:
- Ascend 960DT — Q1’2027
- Ascend 960PR – Planned launch: Q3 2027
Huawei originally had a later roadmap for these chips, so bringing them forward is an acceleration of the company’s AI hardware plans.
What Is the Atlas 960E SuperPoD?
One of the biggest announcements is the Atlas 960E SuperPoD.
A SuperPoD, rather than a single AI chip, is a larger computing system that connects many processors together. Huawei says the Atlas 960E can scale to 4,096 NPUs with up to 8 EFLOPS of FP8 computing performance and up to 1 petabyte of HBM capacity.
Such architecture is meant for demanding AI training and inference workloads.
Atlas 960E at a Glance
| Feature | Huawei Atlas 960E |
| AI processors | Up to 4,096 NPUs |
| FP8 performance | Up to 8 EFLOPS |
| HBM capacity | Up to 1 petabyte |
| Interconnect | UnifiedBus + NPO |
| Optical technology | Hi-ONE |
| Main purpose | Large-scale AI computing |
Why Near-Packaged Optics Are Important
Huawei is also bringing Near-Packaged Optics (NPO) to its AI infrastructure.
Big artificial intelligence systems require very fast links between processors. Conventional optical interconnects may require large numbers of optical modules, increasing power consumption and system complexity.
Huawei says its Hi-ONE optical-interconnect technology can provide 7.2Tbit/s gearbox capacity per engine.
For the Atlas 960E SuperPoD, Huawei said its approach can reduce power consumption by more than 550 kilowatts compared with the traditional configuration it used for comparison.
Huawei Is Making More Than Just AI Chips
Another major component of the announcement is Huawei’s insistence on connecting multiple processors.
The company claims its UnifiedBus technology can be used across computing, networking, storage and management components.
Huawei has also proposed an AI SuperCluster architecture that can scale up to one million NPUs when multiple systems are combined.
The idea is simple: rather than relying on the performance of any one processor, huge numbers of processors can collaborate on the most demanding AI workloads.
Huawei is accelerating its AI chip roadmap. Why?
The AI chip market has become strategically important as demand for AI computing continues to increase.
Chinese firms face limits on access to some advanced U.S. AI chips and technology, but Huawei is building its own Ascend ecosystem.
Huawei’s domestic AI ecosystem expansion continues with demand for its AI computing equipment in China currently exceeding its production capacity, Reuters reported.
This means that the development of Huawei AI Chips is significant not only in terms of hardware, but also in terms of China’s broader attempt to expand domestic AI computing infrastructure.
What’s Next After Ascend 960?
Huawei already has a longer-term road map.
The company says it plans to launch the Ascend 970 in 2028 and the Ascend 980 in 2029.
Huawei says it plans to improve computing capability, memory bandwidth, memory capacity and interconnect performance while maintaining an approximately one-generation-per-year cycle.
That indicates Huawei is viewing AI hardware as a long-term platform, not a single product launch.
FAQs
What Are Huawei AI Chips?
Huawei AI chips are processors that Huawei builds specifically to run artificial intelligence workloads such as model training and inference. The core of the strategy is the company’s Ascend family.
When is the Ascend 960 release date?
Huawei plans to launch the Ascend 960PR in 2027 and the Ascend 960DT in Q1 2027.
What is Atlas 960E SuperPoD?
The Atlas 960E SuperPoD is a large scale AI computing system built around Ascend 960 processors. Huawei claims it’s scalable to 4,096 NPUs.
Huawei and Nvidia ? Competition ?
Huawei is building its own AI computing ecosystem as an alternative to foreign AI hardware, including Nvidia products. However, different platforms have different capabilities, software ecosystems and availability, so direct comparisons depend on the specific workload and system configuration.
What’s the next AI chip from Huawei after Ascend 960?
Huawei says the Ascend 970 is expected in 2028 and the Ascend 980 is expected in 2029.
Summary
Huawei’s new AI chips announcement suggests it is moving beyond individual AI processors to complete computing systems.
That’s the Ascend 960DT and 960PR chips, designed to power increasingly demanding AI workloads, the Atlas 960E SuperPoD, and Huawei’s UnifiedBus and optical-interconnect technologies.
With demand for AI computing set to rise, Huawei’s accelerated roadmap will be an important development to watch in 2027. The company also has Ascend 970 and 980 in the pipeline for 2028 and 2029, showing that its AI hardware strategy extends far beyond the current generation.