Huawei brings near-packaged optics to AI with Atlas 960E superpod

Huawei's Atlas 960E superpod addresses a critical constraint in AI at scale: data movement. The company's NPO approach, using its Hi-ONE engine, aims to shorten electrical paths before optical conversion. This differs from Nvidia and Broadcom's co-packaged optics (CPO), which integrate optical engines directly into the chip package. Huawei argues NPO offers easier manufacturing and servicing, citing its pluggable components.

This move positions Huawei to address the growing demand for high-performance AI infrastructure within China. As US export controls limit access to advanced Western AI chips, Huawei's ability to deliver integrated solutions becomes crucial for domestic AI development. The Atlas 960E, with its claimed 550 kilowatt power reduction, offers a compelling alternative for large-scale AI training in the region.

The test for Huawei is whether its NPO proposal gains traction with the Optical Internetworking Forum (OIF). Industry standardization would broaden NPO's adoption and potentially establish it as a viable evolutionary step toward CPO, as suggested by OIF officials. Without interoperability, NPO risks remaining a proprietary solution, limiting its market reach beyond Huawei's own ecosystem.