ZTE's Cableless Supernode to Hit Market in 2026 as AI Competition Shifts to Systems

Competition in large language models is shifting from parameter scale to token efficiency. The proliferation of agent applications means complex tasks require tens or even hundreds of model calls, and the collaborative execution of multiple agents continues to drive up token consumption. Wang Weibin, chief scientist of product planning at ZTE, said that what truly determines user experience is whether the system can continuously output tokens at lower cost and higher speed, which requires stable and efficient responses at the compute level when multiple users access simultaneously. In 2026, domestic supernode solutions will collectively enter the industrialization phase, driven by both demand-side and supply-side factors.

On the demand side, large models are evolving toward trillion-parameter scales, and the mixture-of-experts (MoE) architecture is becoming the industry mainstream. Compared with traditional dense models, MoE models reduce computational costs but create more frequent data-exchange requirements. Each token generation requires parameter synchronization and data exchange among different GPUs, prolonging GPU communication wait times. The traditional approach of horizontally scaling with 8-GPU servers can no longer fully unleash GPU computing power. Even when large numbers of GPUs are deployed in massive AI computing clusters, insufficient cross-node communication efficiency leaves significant room for improving actual compute utilization.

Supernodes organize dozens or even hundreds of GPUs into a unified computing unit via high-speed interconnects, reducing data-waiting time and improving memory utilization and communication efficiency. The industrial chain for this technology has matured over the past few years of accumulation, moving supernodes from proof-of-concept to industrialization. The rapid development of agent applications has further amplified token demand. An agent completing a complex task requires autonomous planning, tool invocation, multi-round reasoning, and repeated execution, multiplying token consumption. With token costs currently high, repeated reasoning and redundant token consumption in complex tasks further push up inference costs, prompting the industry to discuss the concept of token budgeting.

Domestic supernode solutions differ from overseas ones in development paths. Overseas vendors maintain leadership in key technologies such as advanced process nodes, packaging processes, and HBM high-bandwidth memory, giving their individual GPUs more computing power. Domestic vendors have adopted a route of multi-factor synergy and open interconnects, improving system performance by increasing the number of GPUs per rack, enhancing inter-rack expansion capabilities, and using high-performance networks to compensate for interconnect deficiencies. In 2026, the innovation focus of domestic supernode solutions is on interconnect architecture, switching capability, thermal design, and engineering capability, with the industry's discussion shifting from making chips to building systems.

Cableless solutions represent an important direction in architectural innovation. Early solutions used large numbers of copper cables to interconnect compute and switching nodes. A single rack might contain more than 5,000 cables requiring precise insertion, and a single faulty cable could disrupt the entire system. The cableless architecture features a backplane-free, orthogonal design aimed at shortening signal transmission paths and reducing latency. Eliminating cables frees up bandwidth, improves system reliability, and lowers maintenance costs. This design stems from the company's expertise in core routers, achieving efficient integration of data processing and data exchange. Overseas vendors' next-generation supernode solutions have also begun abandoning copper cables in favor of large backplane connections.

Innovation in interconnect methods is another key trend. As GPU scale expands, copper-cable transmission distance and bandwidth capabilities are approaching physical limits. The industry is exploring optical interconnects to replace traditional electrical interconnects to boost bandwidth and reduce latency. Optical fibers offer stronger interference resistance and support longer-distance data transmission. Approaches such as near-package optics, co-packaged optics, and on-chip optical interconnects focus on shortening the distance between optical modules and chips to improve overall efficiency. At present, optical interconnect solutions still need time before large-scale commercialization; the technology has not yet reached the stage for scalable deployment. The cableless architecture remains a more mature and deployable solution.