Huawei is promoting near-packaged optics (NPO) as the practical next step for AI datacentre networking, positioning it as a cheaper and easier-to-maintain alternative to co-packaged optics (CPO). At the China International Optoelectronic Exposition in Shenzhen, the company demonstrated a 7.2 terabits per second (Tbps) NPO module that will feature in its upcoming SuperPOD AI systems, arguing that while CPO offers lower latency by placing optical engines directly on the switch chip's substrate, its complexity makes it costly and difficult to repair when components fail.

NPO instead keeps the optical engine as a separate package mounted close to the switch's ASIC, and Huawei says integrating DSP functions into the ASIC cuts latency from 100 to 10 nanoseconds while reducing power use by roughly 60 percent. However, industry standardisation is still nascent: the Optical Internetworking Forum, whose members include Cisco, HPE, Arista and Nvidia, only launched an NPO standards project in May, targeting a 6.4Tbps interface, well behind Huawei's 7.2Tbps offering. Huawei executive Dr Man Jiangwei expressed confidence that the industry would reconcile these differing specifications, and hinted that further details, including on the next SuperPOD platform for its Ascend NPUs, would be revealed at Huawei Connect in Shanghai this week.