A September 17, 2026 report said Huawei was scheduled to present the Ascend 960 in Shanghai, putting the company’s future AI accelerator back in the spotlight. Huawei’s own roadmap places the chip’s planned availability in the fourth quarter of 2027 and pairs it with an even bigger ambition: an Atlas 960 SuperPoD designed to connect up to 15,488 Ascend 960 chips.

The Ascend 960 is an AI accelerator—not a smartphone—and Huawei has presented it as the third chip in a roadmap that also includes the Ascend 950 and Ascend 970. The company’s figures are roadmap targets, so the real test will come when hardware and software are running at scale.

Huawei’s Ascend 960 roadmap

Huawei introduced the Ascend 960 at Huawei Connect 2025 in Shanghai on September 18, 2025. The company plans to make the chip available in the fourth quarter of 2027, after the planned Ascend 950 rollout and before the Ascend 970’s planned fourth-quarter 2028 availability.

Huawei says the Ascend 960 will deliver twice the computing power, memory access bandwidth, memory capacity and number of interconnect ports of the Ascend 950 chips. It also plans support for HiF4, Huawei’s proprietary data format for 4-bit precision.

The company’s stated performance targets are 2 PFLOPS in FP8 and 4 PFLOPS in FP4. PFLOPS measure floating-point operations per second, while FP8 and FP4 describe the numerical precision used in those calculations. Lower-precision formats can be useful for AI workloads because they reduce the amount of data moved through the system, although the usefulness depends on the model and software.

SubjectPlanned timingMetricHuawei-stated valueCondition
Ascend 960Fourth quarter of 2027FP8 computing target2 PFLOPSChip roadmap target
Ascend 960Fourth quarter of 2027FP4 computing target4 PFLOPSChip roadmap target
Atlas 960 SuperPoDFourth quarter of 2027Ascend 960 chipsUp to 15,488Planned maximum configuration
Atlas 960 SuperPoDFourth quarter of 2027FP8 computing target30 EFLOPSSystem target
Atlas 960 SuperPoDFourth quarter of 2027FP4 computing target60 EFLOPSSystem target
Atlas 960 SuperPoDFourth quarter of 2027Memory4,460 TBSystem target
Atlas 960 SuperPoDFourth quarter of 2027Interconnect bandwidth34 PB/sSystem target
Atlas 960 SuperClusterPlanned future systemNPU countMore than 1 millionHuawei plan
Atlas 960 SuperClusterPlanned future systemFP8 computing target2 ZFLOPSHuawei plan
Atlas 960 SuperClusterPlanned future systemFP4 computing target4 ZFLOPSHuawei plan

From one accelerator to a giant Atlas system

Huawei’s official Atlas 960 SuperPoD presentation shows the planned system scale and its headline FP8 figures.

The striking part of Huawei’s announcement is not just the Ascend 960’s chip-level targets. It is the scale of the system built around it.

Huawei plans an Atlas 960 SuperPoD with up to 15,488 Ascend 960 chips spread across 220 cabinets: 176 compute cabinets and 44 communications cabinets. The planned deployment would occupy 2,200 square meters and use Huawei’s UnifiedBus interconnect technology to move data between the accelerators.

The company also describes a future Atlas 960 SuperCluster with more than one million NPUs. Huawei’s targets for that system reach 2 ZFLOPS in FP8 and 4 ZFLOPS in FP4. Those figures describe a planned architecture, not a shipping system available for independent testing today.

Why Huawei is not pursuing a simple chip-for-chip fight

An analysis of Huawei’s roadmap explains the company’s strategy of clustering thousands of Ascend chips instead of relying on chip-for-chip parity with Nvidia.

Huawei’s approach is to combine many accelerators, fast memory and a high-bandwidth interconnect into one larger computing platform. In other words, the company is trying to make the system the product—not just the individual chip.

That strategy matters because a single accelerator’s specification does not determine the performance of a complete AI cluster. Data must move between processors, memory and storage without creating bottlenecks. UnifiedBus is Huawei’s answer at the interconnect layer; the SuperPoD is the physical system intended to put that answer to work.

This is also why the Ascend 960 cannot yet be called better than a specific Nvidia accelerator. Huawei has published its own FP8 and FP4 targets, but the available figures do not form a like-for-like Nvidia comparison under identical workloads and conditions. The meaningful contest will involve complete systems, software, networking, memory behavior and the models being run—not just a single number on a specification sheet.

The limits of the roadmap

Huawei’s plan depends on more than designing a faster accelerator. Large clusters require advanced packaging, high-speed memory, reliable networking and semiconductor manufacturing capacity. The roadmap therefore describes an infrastructure program as much as a chip launch.

Huawei’s official targets also need to be read at their stated maturity: they are planned figures for a future product and system. No Ascend 960 benchmark result under controlled conditions appears alongside the roadmap numbers, and the company has not published a public price in the announcement.

The September 17 report described a scheduled Shanghai presentation, while Huawei’s documented roadmap continues to place commercial availability in the fourth quarter of 2027. That timing remains the concrete milestone attached to the Ascend 960 plan.