[Relevant companies involved in this article] Huawei, HiSilicon, Amkor, Tianshui HT Technology, Broadcom, ASE.
Development of NPO Product Technology Roadmap
Before discussing NPO, let us review the AI clusters within data centers over the past few years, which primarily relied on copper cable interconnects. As rates and bandwidth steadily advance to higher levels, data centers initiated the transition of "optics replacing copper" in 2025. 200G-800G pluggable optical modules took the stage, handling all interconnects from within servers to between racks entirely through optical modules. As the single-lane rate comprehensively advances from 112G to 224G, the physical layer limitations of traditional optical modules in handling electrical and optical signals have been fully exposed:
1. Excessive insertion loss: At a 224G lane rate, when electrical signals are transmitted over long-distance PCB traces and connectors, the insertion loss exceeds 20dB, resulting in extremely severe signal attenuation.
2. Power consumption and DSP bottlenecks: To compensate for signal attenuation, traditional optical modules have to incorporate highly power-consuming Digital Signal Processing (DSP) chips for signal shaping and equalization, leading to a persistently high energy consumption proportion across the entire optical communication link. To address this, the industry developed LPO (Linear Pluggable Optics), which removes the DSP chip from traditional pluggable optical modules and uses linear analog circuits for direct driving. However, this significantly increases the bit error rate and shortens the effective transmission distance, making it a fleeting trend that failed to make a substantial impact.
It is precisely against this backdrop that the industry urgently needs to push optical interconnect interfaces closer to computing chips. Although the universally recognized ultimate evolutionary direction is Co-packaged Optics (CPO), which directly packages the optical engine and ASIC on the same substrate, bringing massive performance improvements, CPO currently faces insurmountable engineering gaps in manufacturing yield, thermal management, and especially on-site maintainability. Once the optical engine inside the CPO fails, the entire highly valuable ASIC master chip faces the risk of being scrapped.
Huawei's release of the NPO technology roadmap at the Huawei Connect Conference balances performance with engineering feasibility, making it the most promising technology route for the industry on the path toward the ultimate solution. By placing the optical engine at an extremely close distance of about 5cm from the ASIC, the Hi-ONE module successfully compresses the electrical signal transmission path by several times, significantly reducing physical loss. First, looking at the specifications of Hi-ONE, which are top-tier in the industry, its single-engine transmission capacity reaches an industry-leading 7.2T, integrating 36 high-speed optical channels with a single-lane rate of 200G. It is also currently the only mass-produced NPO product in the industry that achieves a built-in light source. Compared to traditional solutions using External Laser Sources (ELS) introduced through polarization-maintaining fibers, the built-in light source greatly reduces complex fiber routing and coupling losses, while also lowering the mechanical complexity of system integration.
Most crucially, based on the core figures revealed at the launch event, in the Ascend 960 Super Node, Huawei directly replaced the 48,000 800G optical modules required by the original traditional solution with 5,500 Hi-ONE modules. The economic and energy efficiency impact of this replacement is staggering: the overall power consumption of the system is directly reduced by over 550 kilowatts, the Mean Time Between Failures (MTBF) is doubled, and system availability reaches 99.8%. Insights from industry veterans indicate that the release of the Ascend 960 Super Node is not merely a hardware iteration; it marks the strategic leap of domestic AI computing infrastructure from "single-chip computing power competition" to "cluster interconnect efficiency and system-level availability competition." It provides high stability and low power consumption engineering feasibility at the physical level for 10,000-card and 100,000-card clusters of 10-trillion-parameter models.
In-Depth Analysis of Technology and Patent Core
An in-depth analysis of the technical core of the first NPO product, Hi-ONE, reveals extensive research in multidisciplinary cross-innovation involving optics, mechanics, electronics, electromagnetics, and thermal management. Regarding the two major industry challenges of chip design and packaging manufacturing, Huawei has indeed demonstrated strong system-level problem-solving capabilities. In such a high-density 7.2T optical engine, thermal density is a persistent headache for the industry. Optical chips (especially lasers and modulators) are extremely sensitive to temperature, and even minor temperature fluctuations can lead to wavelength drift and increased bit error rates. Hi-ONE features a precise balanced design of "optics, mechanics, electronics, electromagnetics, and thermal management," optimizing the layout of micro heat sinks and flow channels to ensure that the optical engine maintains extremely low thermal resistance and thermal crosstalk under high-power operation, guaranteeing long-term operational stability.
Based on the revealed information and combined with previously published Huawei patents, Hi-ONE adopts a technology route that combines self-developed compound optical chips with silicon photonic chips. The silicon photonics part uses CMOS technology to achieve extremely high integration, while the compound optical chips excel in luminous efficiency and high-frequency modulation. The selection of these two technology routes is highly astute, and the heterogeneous integration achieved through advanced packaging processes yields excellent results, realizing extreme bandwidth density within a single engine.
Meanwhile, the physical distance between the optical engine and the ASIC chip is shortened to within 5 centimeters, causing the transmission loss of high-frequency electrical signals to drop exponentially. The routing of a large amount of underlying PCB has also been balanced, further improving signal integrity margins under high bandwidth. Compared to the CPO solution that co-packages the optical engine with the expensive ASIC, the NPO form factor of Hi-ONE supports independent testing and independent replacement. According to industry insiders, this "decoupled but not distant" design significantly improves the mass production yield of the entire board, and also offers high post-maintenance feasibility for computing unit manufacturers.
Industry Chain Development and Commercialization Prospects
Standing at the node of the fourth quarter of 2026, the successive launch and scaled commercialization of NPO concerns not only the success or failure of Huawei's products but also drives the reconstruction of the entire industry chain involving optical communications, semiconductor packaging, precision manufacturing, and cloud service providers. First, looking at the upstream chip and material end, compound semiconductor materials, CMOS process nodes, silicon photonics foundry services, and high-precision laser chips will see explosive demand. In particular, chips with built-in light sources place higher requirements on performance, reliability, and lifespan, driving the accelerated iteration of upstream material science and wafer foundry.
For the packaging and manufacturing end, traditional Outsourced Semiconductor Assembly and Test (OSAT) facilities and substrate suppliers are accelerating their transformation towards co-packaging capabilities. The NPO packaging of Hi-ONE involves process challenges such as micron-level precision alignment and Fiber Array Unit (FAU) coupling, and yield control on the packaging mass production side will be the decisive core competitiveness.
Meanwhile, Huawei has also joined forces with over 20 upstream and downstream enterprises in the computing unit industry chain, including China Mobile Research Institute and JD Cloud, to jointly launch the "OPEN NPO" project and release the country's first NPO optical interconnect Multi-Source Agreement (MSA). Regarding international interconnect protocols for NPO, the Open Data Center Committee (ODCC) and the Optical Internetworking Forum (OIF) will successively release detailed rules such as NPO hardware interfaces, physical channel models, and slot mechanical standards for rate ranges from 3.2T to 7.2T, with Huawei once again taking the lead.
Conclusion
It can be clearly stated that the emergence of Huawei's first NPO product, Hi-ONE, and its Ascend 960 Super Node not only paves a broad avenue in technology but also demonstrates the powerful vitality of system-level innovation to the global market in terms of commercialization. We believe that as the "OPEN NPO" ecosystem continues to expand and the construction of domestic intelligent computing centers fully accelerates, NPO will no longer be a "transitional form," and running parallel with CPO will become the norm.
Reference: "Huawei Releases the World's First Super Node Adopting NPO - Ascend 960 Super Node"