AI infrastructure is scaling rapidly as demand for compute capacity continues to rise. Early single-rack deployments have now evolved into complete clusters and even ‘AI factories’, where the entire data centre becomes a unified, interconnected supercomputer optimised for extraordinarily large AI workloads.
While power requirements have often been seen as the biggest constraint for large-scale AI infrastructure, another bottleneck cannot be ignored: interconnects. Alongside increasing cluster sizes, inter-rack distances are growing too, with reach extending from two to 100 metres. At the same time, there is an ever-increasing need for speedy connections.
With these two interconnect challenges – speed and distance – in direct opposition, one thing is certain: traditional copper cables are no longer an option. Instead, data centres are shifting to next-generation optical interconnect technologies that rely on light to transmit data without latency.
But here lie two possible optical interconnect directions: NPO or CPO?
Why NPO is the practical approach for industry
The core difference between NPO and CPO is architectural. Near-packaged optics (NPO) integrate with the main chip on the PCB, meaning optical components sit in close proximity to the switch on a separate, independent substrate. Whereas co-packaged optics (CPO) integrate directly at the substrate level, promising ultra-short trace lengths.
While CPO may sound theoretically superior, other important distinctions point towards NPO as the more practical solution today. In terms of supply chain, NPO manufacturing is closely based on pluggable optical modules, while CPO relies on both pluggable optics and semiconductors. While each supply chain is mature in its own right, integrating the two for co-packaged optics depends on a supplier collaboration that is significantly less established.
NPO also wins from both scale-up and maintenance perspectives. Ramp-up speed is significantly faster for NPO than CPO, enabling a smoother transition to mass manufacturing as adoption picks up. Finally, since NPO is pluggable, it facilitates modular, field-replaceable units. This highlights a key limitation for tightly integrated CPO networks, where a single optical failure can be disastrous for the entire assembly.
For these main reasons, Huawei views NPO technology as the most practical and mature interconnection option for scaling AI, but the company is keen to confront the challenges that lie ahead for this route.
Huawei’s contribution to global NPO standards
In AI networks, standards and specifications are essential for helping components made by various manufacturers work together seamlessly. During a recent roundtable at the NPO Standards Industry Summit, Huawei’s Director of the Advanced Opto-Electronics Laboratory, Dr. Man Jiangwei, spoke about the fragmented dynamics of NPO standards organisations.
In this disjointed ecosystem, Huawei says upstream players work on certain standards while downstream players try to progress others, from CPx and IPEC NPO to ODCC UPO, GCC-OAII, and more. With the market mostly characterised by smaller companies, the result is a somewhat weak, siloed ecosystem that, in its current form, could struggle to compete against the more consolidated CPO community.
To unlock NPO’s full potential, Huawei realised that change was needed through open industry discussions, cross-border collaboration, and global standardisation.
Enter the Optical Internetworking Forum, or OIF. Since 1998, OIF has been instrumental in creating vital interoperability specifications for optical and electrical networking components, making it the obvious choice for Huawei’s ambitious project to unite the optical interconnect ecosystem under one standard.
Based in California, US, OIF has membership across industry-leading cloud providers, chip and system vendors, and optoelectronic manufacturers, amongst them Google, AWS, Nvidia, Cisco, Broadcom, Microsoft, and Coherent. At a session in May, the discussion was re-initiated, with players engaged across a wide range of geographies and markets.
“We’re trying to get people on the same page so that we can get more support to push on the NPO standard project initiation,” Dr Man explains. “Huawei is playing a key role in promoting this process. We had over one week of meetings, and we kept clarifying our positions, answering their questions, responding to the challenges, and proactively interacting. This way, we can have a thorough understanding of the technical evolution.
It shows that the new NPO standard is in line with the core interests of the vast number of vendors.”
Throughout this challenging process, the intention of the NPO standard received support from more than 40 vendors across the entire industry value chain, from North America, Japan, and China. Following this, standards documents were released for 12.8 Tbps and 6.4 Tbps (terabits per second) solutions, defining the mechanical module specifications, electrical interconnect and pin map, liquid cooling, laser source provisions, power supply requirements, and optical connector considerations.
Dr Man says NPO standardisation is now “marching forward”, with players across the industry actively contributing to the project with OIF. “I firmly believe that with the guidance of an organisation like OIF, the entire industry will become more unified and open,” he says.
Dr. Zhao Wenyu, Vice Chairman of IPEC and Deputy Director of the Technology and Standards Research Institute of China Academy of Information and Communications Technology (CAICT), also echoed Dr. Man’s views regarding the significance of global standards for NPO, “We call upon stakeholders across industry, academia, research institutions, and user groups to join hands in developing NPO standards and driving the NPO ecosystem to evolve from ‘usable’ to ‘easy-to-use at scale’”.
The next step: Huawei launches 7.2 Tbps module
At the China International Optoelectronic Exposition (CIOE), Huawei demonstrated the industry’s first high-bandwidth 7.2T NPO product and successfully conducted a demonstration of dynamic transmission at the IPEC exhibition booth. This has taken the NPO technology from theoretical discussion to industry verification. Dr. Man highlighted the key technologies such as built-in light sources and highly-integrated SiPh chips in the 7.2T NPO module, which features high bandwidth, high reliability, and high availability’ while achieving ‘low latency, low power consumption, and low cost’.
Dr. Man went on to introduce Huawei’s solid foundation and deep industry experience in optical interconnects. Huawei has introduced optical interconnects for the first time in the Atlas 910C SuperPoD, and for the Atlas 950 SuperPod, Huawei brought in Linear Pluggable Optics (LPO). Now, Huawei has pioneered the 7.2T NPO for the third-generation product and claims that it will be deployed in its future SuperPoD series.
Asked when the company will announce the mass production of 7.2T NPO, Dr. Man didn’t answer directly but pointed to the upcoming Huawei Connect on September 17 in Shanghai, where Huawei will share its vision and roadmap for NPO innovation.
