Huawei Unveils Ascend 960, Sets Annual AI Chip Cadence: what are the key facts?

The Ascend 960DT R&D was accelerated by three quarters, achieving 2 PFLOPS of FP8 compute per chip.

Huawei Unveils Ascend 960, Sets Annual AI Chip Cadence: what are the key facts?

Huawei launched a 4,096-card supernode, improving cluster MFU by 2.75x compared to traditional eight-card servers.

Huawei Unveils Ascend 960, Sets Annual AI Chip Cadence: what are the key facts?

It is the industry's first system to adopt NPO optical interconnect technology, reducing power consumption by over 550kW.

Huawei Unveils Ascend 960, Sets Annual AI Chip Cadence: what are the key facts?

The CANN ecosystem has more than 5,200 monthly active developers, with external developers accounting for 61%.

Huawei Unveils Ascend 960, Sets Annual AI Chip Cadence: what are the key facts?

Huawei plans to maintain an annual release cadence, launching the Ascend 970 in 2028 and the 980 in 2029.

Physix Frontier · News Briefing Card (enbrief · Sep 20, 2026)

Huawei Unveils Ascend 960, Sets Annual AI Chip Cadence

KEY FACTS

The Ascend 960DT R&D was accelerated by three quarters, achieving 2 PFLOPS of FP8 compute per chip.
Huawei launched a 4,096-card supernode, improving cluster MFU by 2.75x compared to traditional eight-card servers.
It is the industry's first system to adopt NPO optical interconnect technology, reducing power consumption by over 550kW.
The CANN ecosystem has more than 5,200 monthly active developers, with external developers accounting for 61%.
Huawei plans to maintain an annual release cadence, launching the Ascend 970 in 2028 and the 980 in 2029.

KEY DATA

2 PFLOPSFP8 Compute Per Chip

4,096 NPUsSupernode Scale

2.75xMFU Improvement Factor

5,200+CANN Monthly Active Developers

PHYSIX OBSERVATION

Huawei is shifting from single-chip competition to exporting systems engineering capabilities. Amidst advanced process node restrictions, it compensates for individual chip limitations through NPO optical interconnects and supernode architectures, maintaining compute supply elasticity via high-frequency 'one-generation-per-year' iterations. This move aims to build an autonomous and controllable AI infrastructure foundation, providing stable compute support for domestic large model training that does not rely on overseas supply chains, through open-source software ecosystems and hardware scaling synergy.

Source: enbrief original report ↗