光网络成为AI机器一部分：NPO架构崛起改写算力互联格局

Markdown Content: As AI clusters grow larger, the connections between processors increasingly determine the performance of the entire system. Near-packaged optics (NPO) offers a practical way to bring optical connectivity closer to the silicon without sacrificing maintainability or supply-chain flexibility. Building faster chips is no longer enough. Those chips must exchange enormous volumes of data quickly and efficiently enough to behave as a single computing system. The network is therefore moving closer to the processor, both physically and architecturally. This shift has intensified the search for an optical architecture capable of delivering the bandwidth and power efficiency AI requires without creating a new set of manufacturing and operational problems. Near-packaged optics, or NPO, is emerging as the most practical answer: a bridge between the limits of conventional pluggable optics and the tightly integrated —but considerably more challenging —world of co-packaged optics. The growth of large language models, multimodal applications and AI agents is driving the construction of ever-larger computing clusters. But adding more accelerators does not automatically produce a proportional increase in usable computing power. Every processor in a cluster must constantly exchange information with thousands of others. If those connections consume too much power, introduce too much latency or cannot deliver sufficient bandwidth, the processors spend more time waiting and less time computing. Traditional electrical connections become increasingly difficult to operate as transmission speeds rise. Signal integrity deteriorates. Reach collapses. The power and complexity required to compensate for the increase. Data presented by Huawei at the sideline event during the NPO Standards Industry Summit in Shenzhen, China, illustrates how severe that physical constraint is becoming. It showed the approximate reach of electrical cables falling from five meters at 50 Gbit/s per lane to three meters at 100G and one meter at 200G. At 400G per lane, the estimated reach falls below 0.6 meters. Pluggable optical modules have been extraordinarily successful for good reasons. They are standardized, replaceable and supported by an extensive global supply chain. Located on the front panel of a switch, they can be installed, upgraded or replaced without disturbing the switching silicon at the center of the system. But their physical separation from that silicon creates an increasingly difficult electrical journey. Data must travel from the ASIC across the printed circuit board to the optical module. As speeds rise, preserving the integrity of that signal can require components such as retimers, adding power consumption, latency and system complexity. Moving the optics closer to the processor shortens that electrical path. The question is how close the industry can move that path without losing the economic and operational advantages that made pluggable optics so successful. Co-packaged optics, or CPO, represents the most integrated approach. It places optical engines alongside the switching silicon within the same package. This can deliver major improvements in bandwidth density and power efficiency, but it also creates difficult problems involving advanced packaging, thermal management, manufacturing yield and field replacement. If an optical component fails in a conventional pluggable architecture, it can generally be replaced independently. In a tightly integrated CPO system, the relationship between the optics and the switching silicon becomes considerably more complicated. The benefits are real, but so are the consequences for deployment, maintenance and cost. That distinction is the heart of the NPO proposition. Shortening the electrical connection can reduce signal loss and the need for power-hungry signal conditioning. Keeping the optical engine separate preserves more of the modularity operators already understand, allowing optical components to be manufactured, tested, serviced and replaced independently of the main processor. Dr. Zhao Wenyu, vice chairman of the International Photonics and Electronics Committee (IPEC) and deputy director of the Technology and Standards Research Institute at the China Academy of Information and Communications Technology, [described](https://www.ipec-std.org/22706.html) NPO as the likely preferred near-term solution for interconnecting large AI computing clusters, or SuperPoDs. He argued that NPO combines energy efficiency with port density and compatibility with the existing pluggable ecosystem, while CPO still faces a relatively closed supply chain and more difficult operational and maintenance requirements. Over the longer term, he expects NPO to coexist with CPO and other optical architectures rather than being displaced by them. For operators building those systems now —not at some theoretical point in the future — the balance is critical. The most tightly integrated technology is not automatically the most deployable. A commercially viable system must combine performance with manufacturability, maintainability, supply-chain resilience and a credible route to scale. The NPO Standards Industry Summit, co-hosted by IPEC and the Optical Internetworking Forum (OIF) on Sept. 10, 2026, brought together representatives from CAICT, Tencent, Alibaba Cloud, Baidu, Huawei, Broadcom, SENKO, Amphenol and Yamaichi. Their presence matters. NPO depends upon coordination between hyperscalers, switch and silicon vendors, optical-module manufacturers, connector suppliers, test companies and the operators that will eventually deploy the systems. A breakthrough in one component does not create an ecosystem. The technical direction is also becoming more concrete. The OIF is developing a 12.8 Tbit/s NPO module project based on 200G-per-lane electrical interfaces. Huawei is a founding participant of that project and helped align the requirements of cloud companies, system vendors, optical-module manufacturers and connector suppliers during its development. Huawei is also working to move NPO beyond standards discussions and into deployable systems. At the China International Optoelectronic Conference (CIOE) 2026, [the company demonstrated](https://www.usatoday.com/press-release/story/42783/ipec-successfully-holds-npo-standards-industry-summit-aiming-to-build-a-robust-international-standards-ecosystem/) dynamic transmission using what it describes as the industry's first high-bandwidth 7.2T NPO product. The design uses an integrated laser source and highly integrated silicon photonics to combine high bandwidth and availability with lower power consumption and latency. Those claims — and the architecture behind the 7.2T product — deserve separate examinations. The immediate significance is that NPO is moving out of presentations and into products, demonstrations and compatibility testing. NPO cannot scale if every supplier implements different electrical, optical, mechanical and management interfaces. Fragmentation would limit interoperability, complicate qualification and recreate the vendor dependencies that standardized pluggable optics helped the industry overcome. Standards are therefore not an administrative detail to be addressed after the engineering is complete. They are part of the NPO value proposition. They will determine whether operators can source interoperable components from multiple suppliers and whether manufacturers can invest against a common set of requirements. AI is redrawing the boundary between computing and connectivity. Electrical links alone cannot carry the full burden as clusters become larger, faster and more power-dense. Optics must move closer to the processor — but they must get there without sacrificing the ability to manufacture, maintain and scale the infrastructure around it. We now know that optics is becoming part of the AI machine. The decisive question is whether the industry can build that machine on an open foundation — or whether the interfaces connecting its processors will become another proprietary battleground.