CPO’s ideal, NPO’s reality: Why are domestic supernodes betting first on “near-package optics”?

```

As domestic computing power competition shifts from single card performance to system efficiency, NPO is becoming a key piece in bridging the high-speed interconnect capability of super nodes. Tencent has clarified a 26Q4 deployment timetable and is promoting unified industry standards, which not only signals demand for optical chips, optical engines, and packaging equipment, but also means that domestic AI chips, switching systems, and servers will usher in new system-level growth. CPO represents the long-term ideal, but will NPO become the realistic answer to large-scale domestic super node deployment first?

1. What happened?—Tencent to deploy NPO super nodes in 26Q4

1. Tencent gives a timetable, NPO moves from technical alternative to engineering deployment:

On July 21, a Tencent Cloud executive stated at a sub-forum of the World Artificial Intelligence Conference that in order to minimize inference costs, the company will massively deploy domestic computing resources and plans to deploy NPO near-packaged optical super nodes in Q4 2026, also calling for unified global NPO industry standards. Unlike previous technology validation and small-batch testing by the supply chain around NPO, Tencent has now given implementation targets, deployment timelines, and standards, meaning NPO is moving from conceptual discussion to engineering implementation by top cloud vendors.

It is noteworthy that Tencent places "domestic computing power" and "NPO super nodes" within the same cost optimization logic. Domestic AI chips solve the issue of supply and supply chain security, while NPO addresses how clusters of many domestic chips can communicate with each other at lower power, higher bandwidth, and lower latency. Only by addressing both the existence of chips and their efficient synergy can domestic computing resources move from single card substitution to system-level applicability.

This strengthens an emerging industry judgment: CPO represents the long-term direction of deep optoelectronic integration, while NPO better meets current AI super node needs for deployment speed, maintainability, and supply chain maturity. In traditional pluggable solutions, optical modules are on the switch panel side, with high-speed electrical signals traveling a longer PCB trace before optical-electrical conversion; NPO moves the optical engine inside the motherboard, closer to the switching ASIC or AI chip, shortening typical electrical interconnect distance from about 300mm to 150mm, decreasing signal loss and power consumption while retaining the capability for independent optical engine removal and replacement. CPO can further reduce electrical interconnect distance to within 50mm, but joint optoelectronic packaging brings more complex yield, thermal, device life, and maintenance issues.

Therefore, NPO is not the most extreme performance solution, but is a more scalable solution at present. It sits between pluggable optical modules and CPO: closer to the chip and lower power than the former; less packaging difficulty and more controllable maintenance cost than the latter. For cloud vendors eager to improve their domestic computing cluster efficiency, such "good-enough performance and engineering feasibility" balance is often more important than pursuing theoretical optimum.

2. Why has market focus shifted from CPO to NPO?

For years, CPO has been the most visionary direction in optical interconnects. As switch chip speeds and AI cluster bandwidth rise, the long-distance high-speed electrical connections required by traditional pluggable modules are reaching physical limits. CPO, by deploying the optical engine near the switching ASIC, minimizes SerDes loss, DSP power consumption, and system latency.

The issue is that CPO's advantage depends on deep integration, making mass production more difficult. The switch chip, optical engine, laser, and advanced packaging must all meet datacenter reliability for yield, cooling, and longevity. If a single optical component fails, repair scope may expand from one module to an entire high-value package. For cloud vendors, what matters is not the lab's lowest power, but the full lifecycle cost determined by procurement, deployment efficiency, fault rate, maintenance time, and system utilization.

NPO sidesteps the hardest step by shortening the electrical connection from chip to optical engine without removing modularity, thus addressing the transition from 800G to 1.6T and from scale-out to scale-up optical demands ahead of CPO mass production. CPO is still the long-term direction, but before its large-scale production, NPO has gained a unique industry window.

2. Why is it important?—NPO is a key piece of domestic super nodes

1. NPO is evolving from Nvidia's "backup route" to a key building block for domestic super nodes:

A so-called super node is not just dozens or thousands of AI accelerator cards in a cabinet—it's a system where high-bandwidth, low-latency interconnect and unified scheduling make multiple physical nodes logically behave more like a single large computer. The super node derives its three core features from unified memory address space, ultra-low latency, and ultra-high bandwidth. At WAI 2026, companies such as Tencent, Huawei, Enflame, Moore Thread and Biren exhibited super nodes, indicating a shift in competition from single card parameters to system-level capability for domestic computing power.

This shift is especially important for domestic computing. Leading overseas GPU vendors possess not just single chip advantages, but comprehensive systems integrating GPUs, HBM, switches, high-speed interconnects and software. Although domestic AI chips are advancing rapidly, their architectures, protocols and server forms are still diverse, so as clusters scale from 8 to 64 to 256 or 1024 cards, communication, fault management and software coordination become vastly more important.

Tencent's NPO deployment is essentially using optical interconnect to expand high-bandwidth domains inside domestic super nodes. As traditional copper links struggle to balance distance, speed and power, placing optical engines closer to AI chips shortens high-speed electrical transmission on PCBs, letting more accelerator cards maintain high-frequency data exchange at lower power. For MoE models, long-context inference, and high-concurrency services, chips must continually exchange parameters, caches and intermediates—interconnect efficiency directly affects utilization and cost-per-token.

Tencent's goal is not simply to use more domestic cards, but to enable "less waiting and more computing" in clusters. If single card performance is not advantageous, systems can narrow practical workload gaps with more chips, larger interconnect bandwidth and better scheduling. This means domestic computing is moving from "chip replacing chip" to "system compensating for individual deficits."

2. What industrial signals does Tencent's deployment plan send?

The first signal is that NPO demand is shifting from overseas chip and switching vendors to domestic cloud vendors. Previously, market research focused on Nvidia, Broadcom, Marvell, and leading overseas optical device players, with Chinese companies mainly seen as North American optical communication ecosystem extensions. By binding NPO to domestic computing power, Tencent signals that a CSP-led ecosystem of domestic AI chips, domestic switch platforms, and local optic interconnects is emerging in China.

The second signal is optical interconnect moving from cabinet connections to inside super nodes. Traditional optical modules primarily handled server-switch, switch-switch (scale-out) connections, but NPO moves optoelectronic conversion next to chips, pushing optical connections deeply into boards and node interiors. Even if the total number of compute cards doesn't change, the per node count of optical engines, lasers, PICs, and fibers will noticeably increase.

Thirdly, domestic computing may require more interconnect per unit. If domestic chips require parallel card arrays for the same scale tasks, then per-unit-compute values for switch ports, interconnect bandwidth, and optical device value may exceed single-card-strong overseas systems. Domestic substitution will not necessarily lower high-end optics demand; it may increase NPO's demand elasticity due to larger clusters.

3. Why is Tencent calling for unified industry standards?

One of the biggest industrial obstacles for NPO is not device performance, but the lack of a unified product form. Different chip and cloud vendors may use different optical engine sizes, electrical interfaces, optical interfaces, connection types, laser configs, and protocols. Once each firm develops a closed custom architecture, suppliers must repeatedly develop and validate, making large-scale production difficult and locking customers into single ecosystems.

Unified standards help first to lower costs. When optical engines, external lasers, connectors, and test specs are reasonably common, more suppliers can join, manufacturing scales up, and NPO can shift from expensive custom solutions to replicable standard products.

Second, unified standards help with interoperability for heterogeneous domestic compute. China has many different GPU, NPU and switch chip vendors; if each needs a separate NPO system, cloud vendors bear high adaptation cost. Unified interfaces lower switching costs and enable multi-supplier ecosystems.

More fundamentally, standard setting rights matter. China has recently promoted computing power tech standards to resolve issues in compatibility, synergy and resource utilization. As NPO is still at a stage where tech routes haven't converged, if domestic cloud vendors can get chip, optical, server and equipment companies to set standards together, they can embed the real needs of domestic super nodes into industry standards—instead of passively adapting post hoc to mature overseas ones.

3. What's next?—Follow the NPO industry chain and domestic super nodes

1. NPO industry chain: Focus on optical chips, optical engines and packaging equipment

With Tencent announcing Q4 2026 NPO super node deployment, industry chain research should shift from "does it have the NPO concept" to whether companies have entered sampling or validation for domestic AI chips, switches or cloud customers. In this regard, client validation and mass production matter more than mere technical reserves. Only when a company moves from sampling to system validation, to small-batch to formal orders, will large NPO super node deployments translate into real demand.

Three main categories for focus:

First—optoelectronic chips such as lasers, modulators, detectors and silicon photonic PICs. As NPO moves optoelectronic conversion near the ASIC and interfaces move from 800G to 1.6T, the rate, power and reliability requirements for optical chips rise in sync.

Second—optical engines and high-density optical components such as FAU, MPO, etc. As optical interconnect penetrates from rack to super node interior, the number of optical engines and fiber connections per node will increase.

Third—precision packaging, coupling and test equipment. NPO requires higher precision in optical chip placement, active coupling and optoelectronic testing—areas where Chinese companies may achieve breakthroughs first.

2. Domestic super nodes: Key beneficiaries are AI chips, interconnect, and whole systems

Compared to individual NPO devices, a bigger industry signal is that domestic computing is shifting from "single card replacement" to "super node deployment." Three areas are direct beneficiaries.

First: domestic AI chips. Super nodes amplify the number of chips, expand interconnect bandwidth and boost parallel efficiency, offsetting domestic chips' disadvantages in single-card performance and software ecosystem. Massive deployment by giants like Tencent will expand use cases and demand for domestic GPU, NPU, but true beneficiaries must ensure stable supply and good cluster scalability.

Second: switch chips and high-speed interconnects. The core of super nodes is not "more cards" but efficient AI chip coordination. NPO only handles optoelectronic transmission—switch ASICs, NICs, SerDes and interconnect protocols all support it. As super node scale grows, domestic switch chips and high-speed networks rise in importance, potentially rivaling the value of the optics themselves.

Finally: AI servers and system integration. Domestic AI chips, NPO optical engines and switch systems are brought together at the server level, with mainboard, chassis, power and cooling all re-designed. Server makers with abilities in chip adaptation, whole-machine design, cluster deployment and operation will be the biggest system-level winners from super nodes.

Conclusion:

CPO remains the long-term direction for deep optoelectronic convergence, but in the stage when domestic computing needs to rapidly form cluster competitiveness, NPO is more likely to take root first. Its value chain beneficiaries are not just traditional optics modules, but a domestic super node industry system encompassing domestic AI chips, high-speed interconnects, NPO optical engines, and servers as a whole.

Risk warning and disclaimerThe market has risks and investment should be cautious. This article does not constitute personal investment advice and has not accounted for individual users’ special investment objectives, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their specific situation. Investment based on this is at your own risk. ```