Report Interpretation

Overview

This update examines Huawei’s launch of a Near-Packaged Optics solution for the Ascend 960 SuperPoD AI infrastructure. Morgan Stanley views the launch as evidence of rapid optical-interconnect evolution and as supportive of China developing a competitive position in NPO.

Core views

At Huawei Connect 2026 in Shanghai on September 17, Huawei unveiled a third-generation optical-interconnect architecture based on Near-Packaged Optics (NPO) for the Ascend 960 SuperPoD AI infrastructure. Morgan Stanley argues that this adoption supports its view that China is likely to establish a competitive edge in the new product category. The report frames the announcement as a rapid technology progression: Huawei moved from discrete optics in Atlas 910C, to Linear Pluggable Optics (LPO) in Atlas 950, and then to NPO in Atlas 960. This progression is central to the report’s conclusion that Huawei is advancing optical-interconnect integration quickly within its AI infrastructure. Huawei’s NPO optical engine is specified at 7.2 Tbps across 36 channels, each operating at 200 Gbps per lane. Morgan Stanley compares this with 6.4 Tbps, 32-lane interfaces under development by groups such as the OIF. The report also cites a 10 ns communication latency and lower power consumption, positioning the NPO implementation as relevant to the performance and efficiency requirements of large-scale AI systems. At the system level, the Ascend 960 SuperPoD can scale to 4,096 NPU cards through UnifiedBus, delivering up to 8 Exaflops of FP8 computing power and 1 PB of unified HBM. Huawei also launched the OPEN NPO initiative and an MSA with China Mobile Research Institute, JD Cloud and more than 20 domestic partners. Morgan Stanley further notes that other Chinese companies had announced NPO initiatives during the prior week’s CIOE event, reinforcing the broader domestic momentum around the technology.

Analysis framework

Morgan Stanley assesses Huawei’s announcement by tracing its optical-interconnect technology progression, comparing the announced interface specifications with interfaces under development by the OIF, and linking the technical specifications to the scale and efficiency needs of the Ascend 960 SuperPoD AI system.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Huawei

  Primary subject; introduced NPO-based optical interconnect architecture for the Ascend 960 SuperPoD.

  Strengths

  Rapid progression from discrete optics to LPO and NPO; 7.2 Tbps engine, 10 ns latency, lower power consumption, and large-scale system specifications.

  Comparison

  The 7.2 Tbps, 36-channel interface exceeds the 6.4 Tbps, 32-lane interfaces under development by groups such as the OIF.

Key data

• NPO optical-engine bandwidth7.2 TbpsAcross 36 channels at 200 Gbps per lane.
• Comparison interface bandwidth6.4 TbpsAcross 32 lanes under development by groups such as the OIF.
• Communication latency10 nsReported for Huawei’s NPO optical engine.
• SuperPoD scale4,096 NPU cardsConnected via UnifiedBus.
• FP8 computing powerUp to 8 ExaflopsReported capability of the Ascend 960 SuperPoD.
• Unified HBM1 PBReported capability of the Ascend 960 SuperPoD.
• Domestic partnersMore than 20Participating with China Mobile Research Institute and JD Cloud in the OPEN NPO initiative and MSA.

Impact & implications

The report considers Huawei’s NPO adoption and partner initiative supportive of China’s ability to build a competitive position in next-generation optical interconnects for large-scale AI infrastructure.