Tencent Publishes 102.4T NPO Switch Design Specification, a Step Closer to Large-Scale Deployment?

C114 News, October 20 (Shui Yi): During WAIC 2026 in July, Tencent Cloud executives said the company would deploy domestically produced computing power on a large scale to minimize inference costs. The company expects to deploy NPO super nodes in Q4 2026 and called for unified NPO industry standards at home and abroad.

Recently, Tencent detailed on its Echang Wangshi WeChat account a 102.4T NPO switch designed and developed based on 3.2T NPO module products, and simultaneously released the 102.4T NPO Switch Design Reference Specification. Covering system architecture, chassis structure, fiber cable management and software operations, it provides a complete NPO switch design solution to support large-scale NPO deployment.

Why Publish a Design Specification?

Large model parameters continue to climb and agentic applications are emerging rapidly, driving up computing demand and pushing compute clusters toward large-scale super-node architectures. As a result, the focus of AI computing competition is shifting from the peak performance of individual chips to system-level cluster efficiency, and control over system-level interconnect design has, for the first time, become a new battleground among cloud providers. Leading cloud service providers are all developing their own complete intelligent computing architectures.

In fact, as cloud computing scaled up, leading cloud service providers began developing their own servers, switches and other ICT equipment, defining products tailored to their own business scenarios. In the AI era, systematic solutions have largely taken shape. For example, Tencent developed Xingmai Network, a software-hardware co-designed high-performance computing network system, while Alibaba developed HPN (High Performance Network), a high-performance network architecture.

Customized white-box business has accordingly become an important growth driver for ICT system vendors in recent years. Cloud service providers' product design planning plays a major role in this. System vendors also participate in the design, jointly advancing the rollout of products and solutions.

As early as January this year, Tencent released technical specifications for its 3.2T NPO engine, covering product definition, performance metrics and more. However, Tencent said modules alone are far from sufficient; optical engines must rely on matching system equipment to form a network. The NPO switch is precisely one of the key system platforms that will carry NPO technology toward large-scale deployment.

All of the above are important reasons for releasing the 102.4T NPO Switch Design Reference Specification, which to some extent signals that Tencent Cloud is ready for large-scale NPO deployment.

According to the introduction, the 102.4T NPO switch uses Tencent's self-developed universal switch hardware platform, stands 4RU high, and features a front-mounted management card that supports hot-swap field maintenance. The system is equipped with one 102.4T switch chip, 32 3.2T NPO modules and 32 ELSFP external light source modules, and provides 128 800G MPO16 optical interfaces externally. The 32 NPO modules are arranged symmetrically around the switch chip, with each module allocated 32 100G (112G PAM4) channels and providing four MPO16 optical interfaces externally.

The core design concepts of Tencent's 102.4T NPO switch comprise three points. First, near-chip integration: the 32 3.2T NPO modules are laid out around the switch chip, compressing electrical links to the extreme and solving high-speed signal integrity issues at the source. Second, system co-design: NPO-specific engineering challenges in layout, cable management, pressing and thermal dissipation are addressed one by one through structural innovations such as self-developed power supply modules, cable management racks and spools, and combined cover plates. Third, openness and decoupling: the NPO and ELSFP modules are designed to standard specifications, the software management is compatible with the traditional optical module ecosystem, and the overall system design is opened to the industry chain in the form of a reference specification.

Why is large-scale NPO deployment needed?

In recent months, debate over NPO and CPO has continued across the industry. To be sure, CPO's performance advantages along the evolution of optical interconnect technology are hardly disputed; the question is whether it can be trusted.

CPO integrates optical and electrical components into a single package, but optical devices and electrical chips differ greatly in material systems and sensitivity to heat and mechanical stress. Co-packaging therefore raises a series of issues in reliability, manufacturability and serviceability. CIR, a research firm, previously wrote in a report that CPO could mean replacing an entire board, line card or even a whole switch assembly when an optical component fails, a prospect widely interpreted as "vendor lock-in disguised as innovation."

In fact, Tencent announced last year that it had begun deploying its self-developed CPO switches on the live network, gaining first-hand experience in deployment and operations and evaluating the feasibility and performance of CPO technology in complex network scenarios, but little further information has been disclosed since. This year, the company is going all out to advance large-scale NPO deployment.

NPO places the optical engine near the main chip, greatly shortening the electrical link. It offers technical advantages including low power consumption, low latency and module-level serviceability, making it an open, disaggregated solution that balances performance, cost and availability while largely preserving customers' usage habits and the existing industry ecosystem.

According to Tencent, compared with conventional pluggable switches, the 102.4T NPO switch delivers multiple benefits: it eliminates the DSP chip inside the electro-optical conversion module, cutting total system power consumption by about 1000W and reducing end-to-end transmission latency by nearly 100ns. The high-speed signal path is short, providing strong support for linear direct drive, with measured BER better than 1E-10 across all ports. The current electromechanical platform can be smoothly upgraded to 204.8T and has the potential to evolve to 409.6T.

Tencent is not alone. Industry practice shows that Huawei officially launched the industry's first supernode using an NPO optical engine, the Ascend 960 Supernode, during Huawei Connect 2026. Alibaba Cloud also showcased a 102.4T UPO switch (note: UPO stands for Ultra Performance NPO) at Apsara Conference 2026, while Panjiu AL144 can scale to tens of thousands of GPUs through SNPO (ScaleUp NPO) optical interconnects.

As Wang Yachen, Vice President of Tencent Cloud and General Manager of Tencent Network, observed during WAIC 2026, optical interconnects have become the key to breaking through bandwidth and power bottlenecks as supernodes grow in scale, and NPO is currently the more pragmatic technical route for building supernodes with domestic GPUs.

In Tencent's view, Scale-up networks in supernode architectures are more sensitive to interconnect power consumption and latency, and NPO switches hold clear advantages in both. As signaling rates rise to 224G, the power and performance benefits of NPO switches will widen further, making NPO an important product path for supporting future Scale-up bandwidth and scale growth.

Openness and disaggregation are the critical issue

It is not hard to see that the key reason cloud service providers such as Tencent are championing NPO is openness. At the module level, leading optical module makers including Zhongji Innolight, Eoptolink, Accelink Technologies and HG Genuine have all introduced ELSFP external light source solutions. During CIOE 2026, the IPEC International Photonics and Electronics Committee, together with multiple organizations, completed a cross-vendor interoperability demonstration of a 51.2T NPO switch system.

On the standards front, in July this year Huawei, together with China Mobile Research Institute, JD Cloud, Baidu and more than 20 other industry-chain organizations, launched the OPEN NPO project and released China's first NPO optical interconnect MSA. In September, the Open CPX Specification 1.0 was released ahead of ECOC 2026; Tencent, together with Alibaba Cloud and others, initiated a 3.2T NPO standardization project under ODCC. Earlier, in May, OIF approved the project "12.8Tb/s Near-Package Optical Module," jointly submitted by CAICT, Huawei, Tencent, Alibaba Cloud, Meituan and China Mobile.

The more specifications there are, the slower convergence becomes. Zhao Wenyu, deputy director of the Institute of Technology and Standards at the China Academy of Information and Communications Technology, noted that NPO has yet to converge on electrical/optical connectors, baseplate design, light source placement, interoperability verification, and operations and maintenance management protocols. Luo Chuanneng, CTO of HG Genuine, said the hardest part of NPO is not the technology but the lack of convergence on solutions; for example, every company's Socket proposal is different.

At a panel discussion during the 2026 OCP Global Summit, Daniel Kuchta, chief hardware system architect at NVIDIA, said that multi-source supply is not viable at present. Today, each vendor's NPO samples are different from one another, and there are four related multi-source agreements, all of them new. In his view, for NPO to reach the volumes forecast by LightCounting, the early stage will have to rely on a single supplier for a large batch of products, and expecting NPO to deliver multi-source supply is "fooling yourself."

Zhongji Innolight previously stressed at an investor relations event that, because NPO products are highly customized, customers will not open up many opportunities. Suppliers that are first to enter customized development therefore stand a good chance of gaining a competitive advantage.

Daniel Kuchta also said that many customers are evaluating NPO, that NVIDIA will work with system contract manufacturers on it, and that its own interconnect ecosystem will support it. He expects CPO, pluggable and NPO to coexist in scale-out. Coherent also said around the 2026 OCP Global Summit that its customer collaboration projects for both the CPO and NPO paths each exceed 10, and that both have secured key customers and long-term agreements.

Looking back, alongside publishing the 102.4T NPO Switch Design Reference Specification, Tencent also summarized a series of lessons learned and developed further thinking on the technical evolution of future NPO products, aiming to move NPO from a proprietary solution toward an open ecosystem. Tencent has already begun R&D on a next-generation 6.4T NPO module and a 204.8T NPO switch, providing more efficient and open all-optical interconnect solutions for AI computing infrastructure.