The Atlas 960E SuperPoD links up to 4,096 AI processors with near-packaged optics. Huawei claims lower power use and higher reliability, while the first Ascend 960 chips are planned for 2027.
New to this? Read it in simple words
- Huawei showed a new AI system, the Atlas 960E SuperPoD, on September 17. The company says it can link up to 4,096 AI processors.
- Huawei says the design avoids 48,000 normal optical modules. It claims this cuts power use by more than 550 kilowatts.
- Huawei presents it as another choice besides systems built on Nvidia chips. This matters for Chinese organisations that face limits on getting advanced foreign chips.
- These are Huawei’s own figures, and independent results are not public yet. The first Ascend 960 chips are planned for 2027.
- Optical module
- A small part that turns electrical signals into light to send data fast.
- Kilowatt
- A unit used to measure how much electric power something uses.
The design changes how thousands of processors talk
A SuperPoD is a group of machines linked so closely that software can treat them like one large computer. Fast links matter because AI training often spends time moving data between processors instead of doing useful calculations.
Huawei says one Atlas 960E can connect up to 4,096 Ascend processors. It uses Hi-ONE, a near-packaged optics system. This means optical connections sit close to the computing chips, where they can move data with less extra equipment.
The company says the design avoids 48,000 normal optical modules. It claims this cuts power use by more than 550 kilowatts and doubles fault-free operating time. These are Huawei figures and need outside tests.
Huawei claims lower power use and higher reliability. Delivered systems and independent tests must confirm those figures.
The processor roadmap reaches into 2027 and beyond
Huawei says the Ascend 960DT chip should arrive in the first quarter of 2027. A second version, the 960PR, is planned for the third quarter. Later Ascend 970 and 980 chips are company plans for 2028 and 2029.
The Atlas 960E is designed to hold up to one petabyte of high-bandwidth memory. Huawei reports 8 exaflops of FP8 performance. FP8 is a compact number format that can make some AI work faster and use less memory.
A high peak number does not show how every model will run. Software, memory movement, network delays, faults, and the type of AI task can all change real performance and energy use.
Customers need repeatable results, not a scale contest
Huawei presents the system as another path beside Nvidia-based clusters. That matters for Chinese organisations facing limits on access to advanced foreign chips. It also makes software support and supply reliability central questions.
Useful public tests should show training speed, model-serving speed, total power, failed jobs, repair time, and cost. They should compare the same model and workload across different systems instead of using one company’s best case.
The launch is a real hardware step, but several claims remain company claims. Buyers should wait for delivered systems, customer reports, and independent measurements before treating a planned advantage as a proven one.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · September 17, 2026Advancing the Agentic World, Building a Solid Silicon Foundation Huawei
- 2
- 3Research · September 17, 2026Huawei unveils new chip technologies as Chinese firm steps up the AI race with Nvidia Associated Press
We used Huawei’s launch pages for the architecture, product figures, and roadmap. We used Associated Press for wider market and export-control context. Performance, power, and reliability figures remain company claims until customers or independent labs repeat them.