Huawei Ascend 960 chips will arrive in the first quarter of 2027, not this year. Huawei announced the accelerated timetable at Huawei Connect 2026 in Shanghai on 17 September 2026, moving the 960DT training chip forward by three quarters. The Atlas 960 SuperPoD built from those chips is due in the fourth quarter of 2027.
Ascend 960 and Atlas 960 Are Not the Same Product
Huawei announced a chip, a rack-scale node and a cluster, and coverage has been mixing the three names freely. Huawei Technologies, the Shenzhen networking and devices company, uses Ascend for its AI accelerator chips and Atlas for the systems built out of them.
| Name | What it is | Scale |
|---|---|---|
| Ascend 960DT | Accelerator chip for AI model training | Single chip |
| Ascend 960PR | Accelerator chip tuned for inference | Single chip |
| Atlas 960E | Rack-scale node using near-package optics | Up to 4,096 Ascend cards |
| Atlas 960 SuperPoD | Full supernode system | Up to 15,488 Ascend chips |
| Atlas 960 SuperCluster | Multiple SuperPoDs linked together | Over one million processors |
The figure most often quoted, 4,096 cards delivering up to 8 exaflops of FP8 compute and one petabyte of high-bandwidth memory, belongs to the 4,096-card node rather than to the full SuperPoD. Huawei puts the larger Atlas 960 SuperPoD at 220 cabinets, of which 176 are compute and 44 are communications, occupying about 2,200 square metres and reaching 30 exaflops in FP8 and 60 exaflops in FP4.
When Each Product Actually Arrives
Nothing announced on 17 September is available to buy now. The earliest date on the roadmap is three months into 2027.
- Ascend 960DT: First quarter of 2027, brought forward three quarters from a previously indicated fourth quarter of 2027.
- Ascend 960PR: Third quarter of 2027, one quarter earlier than previously planned.
- Atlas 960 SuperPoD and SuperCluster: Fourth quarter of 2027.
- Ascend 970: 2028.
- Ascend 980: 2029.
As of 18 September 2026, Huawei has published no month within any of those quarters, no pricing for any product on the roadmap, and no list of customers or regions where the systems will be sold. The timing for the smaller Atlas 960E node was reported as the third quarter of 2027 by Wccftech and has not been confirmed elsewhere.
What Changed at Huawei Connect 2026
The news is the schedule, not the existence of the chips, which Huawei had already placed on a public roadmap. Huawei’s rotating chairman David Wang, also rendered as Wang Tao, told the conference that the Ascend 960 has been readied ahead of schedule and delivers twice the performance of the previous generation.
His keynote was titled “Intelligent Future: Building the Silicon-Based Soil for the Intelligent World”, and the conference itself ran from 17 to 19 September 2026 at the Shanghai World Expo Exhibition and Convention Center. Huawei describes the roadmap cadence as one generation per year.
Huawei says the Ascend 960 doubles the computing power, memory bandwidth, memory capacity and number of interconnect ports of the Ascend 950 that precedes it. Huawei has published no independent benchmark results for any of this, a pattern that also ran through the China AI chip cluster plan reported earlier this month.
What Near-Package Optics Changes
The interesting engineering claim is about light, not silicon. Near-package optics, abbreviated NPO, moves the optical components that carry data between chips much closer to the processor package, replacing the pluggable optical transceivers that normally sit at the edge of a rack.
Huawei says a single SuperPoD uses about 5,500 of its Hi-ONE optical modules in place of roughly 48,000 conventional 800G pluggable modules. The company claims that redesign cuts power draw by more than 550 kilowatts per SuperPoD, doubles the mean time between failures and lifts system availability to 99.8 percent.
These are Huawei’s own figures, measured by Huawei, and no third party has verified them. The relevance for buyers outside China is limited in any case, because export controls and Huawei’s own manufacturing capacity, not the design, decide who can obtain the hardware.
What Huawei Did Not Publish
Several things a purchaser would need are still missing, and their absence is the reason to treat the announcement as a roadmap rather than a product launch.
- Benchmarks: No MLPerf or comparable third-party results for any Ascend 960 part.
- Prices: No figure for a chip, a node, a SuperPoD or a cluster.
- Manufacturing: No statement on which foundry builds the chips, at what process node, or at what volume.
- Availability outside China: No named export markets and no comment on sanctions exposure.
Separately, Huawei said it had initiated an interconnect standards project for its UnifiedBus technology. TechTimes reported that the Optical Internetworking Forum approved Huawei’s 12.8 terabit-per-second NPO module standards project in May 2026; that detail comes from a single outlet and Huawei’s own announcement does not spell it out.
Frequently Asked Questions
When Will the Huawei Ascend 960 Be Available?
The Ascend 960DT training chip is scheduled for the first quarter of 2027 and the Ascend 960PR inference chip for the third quarter of 2027. Huawei has not named a month within either quarter.
Can You Buy an Atlas 960 SuperPoD Now?
No. Huawei has set the fourth quarter of 2027 for the Atlas 960 SuperPoD and the Atlas 960 SuperCluster, and has published no price or order process for either.
What Is the Difference Between Ascend 960 and Atlas 960?
Ascend 960 is the accelerator chip. Atlas 960 is the family of systems built from those chips, ranging from a 4,096-card node to a SuperPoD of up to 15,488 chips.
What Is Near-Package Optics?
It places optical data links close to the processor package instead of at the rack edge. Huawei says this lets about 5,500 of its Hi-ONE modules replace roughly 48,000 conventional 800G pluggable optical modules in one SuperPoD.
How Fast Is the Atlas 960 SuperPoD?
Huawei claims 30 exaflops of FP8 compute and 60 exaflops of FP4 across roughly 15,488 chips in 220 cabinets. No independent benchmark has been published to test the claim.
