Huawei is reportedly unveiling an Ascend 960 SuperPoD at Huawei Connect 2026 in Shanghai on September 17, 2026. Huawei says the system combines an upgraded UnifiedBus interconnect with its Hi-ONE optical technology to connect up to 4,000 processors inside one SuperPoD. The announcement also brings near-packaged optics, or NPOs, into the architecture.
That figure is only one part of Huawei’s larger infrastructure strategy. Its 2025 roadmap lists an Atlas 950 SuperPoD with up to 8,192 Ascend 950DT chips and an Atlas 960 SuperPoD with up to 15,488 Ascend 960 chips. Those are separate roadmap configurations, with their own planned specifications and availability windows.
What Huawei reportedly presented on September 17, 2026
The reported 2026 presentation centers on the Ascend 960 SuperPoD, upgraded UnifiedBus and Hi-ONE optical interconnect. Huawei says the combination can link up to 4,000 processors so they operate as a single system rather than as isolated accelerators.
Huawei also introduced near-packaged optics into its SuperPoD architecture for the first time in this announcement. NPO places optical components close to the processing package, an approach intended to support high-bandwidth connections across a large system.
What does connecting up to 4,000 processors mean?
A SuperPoD is Huawei’s term for a group of physical machines that behaves as one logical AI computer. UnifiedBus is the interconnect layer that ties those computing resources together, while Hi-ONE supplies the optical networking element described in the 2026 architecture.
The practical goal is to let AI workloads use a large pool of processors without treating every machine as a separate island. That shifts the focus from the performance of one accelerator to the way thousands of processors share data and computing resources.
Huawei’s 2025 description of UnifiedBus listed 2.1 microseconds of latency, fault detection and protection switching in 100 nanoseconds, and an optical-interconnect range of more than 200 meters. Those are Huawei’s stated system figures.
Atlas 950 and Atlas 960: the roadmap scale
Huawei’s 2025 roadmap gives a clearer view of the planned Atlas systems. The figures below are Huawei’s declared specifications for the full configurations, not independent benchmark results.
| System | Processors or accelerators | FP8 compute | FP4 compute | Memory |
| Atlas 950 SuperPoD | Up to 8,192 Ascend 950DT chips | 8 EFLOPS | 16 EFLOPS | 1,152 TB |
| Atlas 960 SuperPoD | Up to 15,488 Ascend 960 chips | 30 EFLOPS | 60 EFLOPS | 4,460 TB |
The Atlas 950 configuration is planned for the fourth quarter of 2026. Huawei’s roadmap places the Atlas 960 configuration in the fourth quarter of 2027. Huawei also lists 160 cabinets for Atlas 950—128 compute cabinets and 32 communications cabinets—and 220 cabinets for Atlas 960, divided into 176 compute and 44 communications cabinets.
For the Ascend 950DT chip itself, Huawei lists 144 GB of memory and 4 TB/s of memory-access bandwidth. The Ascend 950 series is also described as reaching 2 TB/s of interconnect bandwidth.
Huawei’s chip milestones for 2027
Huawei is reported to be planning the 960DT for the first quarter of 2027 and the Ascend 960PR for the third quarter of 2027. Separately, Huawei’s earlier family-level roadmap places the general Ascend 960 and the Atlas 960 SuperPoD in the fourth quarter of 2027.
The milestones describe different product labels and roadmap levels: two named chip variants, a broader Ascend 960 family entry and a SuperPoD system. The supplied dates do not turn those entries into a single launch event.
Huawei’s bet is system scale, not just individual-chip speed
Huawei’s strategy is to combine Ascend accelerators, Kunpeng processors, optical interconnects, clusters and SuperPoDs into larger AI systems. Guo Ping, chair of Huawei’s supervisory board, said Huawei aims to become the NVIDIA of ICT and computing and described customer infrastructure—not an in-house large language model—as the company’s focus.
That approach matters because restrictions on advanced manufacturing put pressure on Huawei’s ability to compete through chip fabrication alone. A system-level strategy gives the company another lever: connect more processors, coordinate them more tightly and pool their resources across a larger machine.
Huawei’s roadmap compares the Atlas 950 with NVIDIA’s NVL144 and claims higher figures for processor count, compute, memory and interconnect bandwidth. Those comparisons are Huawei’s own stated figures; they do not establish real-world performance parity or superiority in comparable workloads.
The next concrete milestones are the planned 960DT launch in the first quarter of 2027, the planned Ascend 960PR launch in the third quarter, and the roadmap’s fourth-quarter 2027 target for the general Ascend 960 and Atlas 960 SuperPoD.
