NVIDIA recently showcased its next-generation CPO (co-packaged optics) optoelectronic integration scheme, preparing to move the light engine even closer to the compute chip. At SEMICON West 2026 in San Francisco last month, NVIDIA researcher Nikola Nedovic introduced the new technology, called Optics-on-Interposer (OOI), charting a fresh evolutionary direction for interconnect architectures in AI clusters.
In technical terms, compared with existing CPO schemes, OOI moves the light engine responsible for electro-optical conversion from the package substrate onto the silicon interposer where the compute chip resides, further shortening the transmission distance of high-speed electrical signals to reduce power consumption and increase interconnect bandwidth density. Tracing the evolution of optical interconnects — from pluggable optical modules, to CPO designs that place the light engine on the same package substrate as the switch or compute chip, and now OOI pushing the light engine onto the silicon interposer — each step compresses the length of “electrical” traces and amplifies the density advantage of “optics.” In ten-thousand-GPU-class AI clusters, interconnect power and signal integrity have become scaling bottlenecks, and every bit of distance shaved between light engine and compute chip translates directly into power and bandwidth gains.
From an industry-landscape perspective, competition in optical communications is shifting from the “module level” to the “packaging level”: as the physical position of electro-optical conversion becomes the decisive variable for bandwidth density and power, the capability boundaries of traditional optical module vendors, OSAT players and wafer foundries are being redrawn, with foundry giants such as TSMC and network chip vendors such as Broadcom all doubling down on CPO and silicon photonics. By pulling the competition up to the interposer level with OOI, NVIDIA is in effect defining the next generation of “electro-optical co-packaging” standards for AI infrastructure. Going forward, judging a supply-chain player’s AI credentials will depend not only on how many chips it builds, but on how close it can package light and electricity.
For the edge and industrial hardware territory where Junyi Microelectronics operates, CPO today mainly serves data-center AI clusters, but the technology spillover path is clearly discernible: high-density electro-optical interconnects will gradually reach edge servers, 5G edge computing nodes and industrial backhaul networks. Junyi’s 5G-EDGE-3399Pro edge computing host is itself a product of the “compute + connectivity” fusion, and the edge computing nodes built on the AI-BOX-3588 will also benefit from the backhaul-bandwidth and power optimizations that electro-optical integration brings. Junyi Microelectronics will continue tracking how this technology lands at the edge, reserving room for evolution in the next-generation architecture of its edge hardware.
Source: https://k.sina.com.cn/article_7857201856_1d45362c001908t9qk.html (November 1, 2026)