In the first half of the AI infrastructure era, there was a prevailing belief that more powerful chips were winning. However, this assumption officially became outdated at the World Artificial Intelligence Conference in Shanghai in 2026. The main trend is no longer about teraflops on a single board, but rather supernodes—rack-scale systems that combine hundreds or thousands of accelerators into a unified memory.
The competition is now focused on the overall system efficiency rather than the peak performance of individual chips. Biren Tech has positioned itself in the optical interconnect camp. Its NPO optical supernode for 1024 cards utilizes the internal BLink 2.0 protocol and near-packet optics to implement the 'copper exit, optical entry' principle in the scaling domain. Up to 1024 GPUs can share a single addressable memory space. The company's product line ranges from an electrical base configuration of 16 cards to a high-density rack of 128 cards and an optical supernode of 1024 cards.
Yixin Intelligent took a different approach. At its announcement at WAIC, it presented the industry's first RISC-V AI supernode based on its proprietary Epoch cloud AI chip, which is the first mass-produced high-performance AI accelerator based on RISC-V in China. The Epoch chip already supports FP8 precision with block quantization and plans to transition to EXFP4 and MXFP4 in the next generation. The supernode uses an orthogonal backplane-less connection, reducing hardware connection costs by 80 percent and narrowing latency to the hundreds of nanoseconds range.
The rest of the exhibition hall showcased competing concepts. Infinigence CoreX demonstrated dense general-purpose silicon for training. Kunlun Chip, a division spun off from Baidu, presented its third-generation architecture. SingularMOL showed a chiplet-based interconnection. Infinigence, the cloud side orchestrator from Infinigence Cloud, presented the software aspect: a dispatching layer capable of integrating disparate domestic hardware into a single virtual supercluster.
The economic underpinning of this situation is causing discomfort. GPUs in a poorly designed cluster spend most of their time waiting for data. Bandwidth, not computational power, has become the bottleneck for training and inference. Communication barriers mean that placing a thousand cards in a rack without a system interconnection strategy yields a 'token conversion rate' that may only be a tenth of the theoretical value. Chinese supernode suppliers are competing for token utilization per dollar of capital expenditure.
Equipment selection is no longer a decision made by a single vendor. A savvy Chinese AI lab in 2026 will acquire at least two different accelerators and wrap them in a software layer that distributes tasks across various vendors. Labs creating this orchestration layer will own the next bottleneck. WAIC 2026 will be remembered as the year when Chinese AI infrastructure suppliers stopped selling chips and started selling systems.
Computational power is becoming a critical resource as tech giants actively strive to create and deploy artificial intelligence supernodes. This frenzy is driven by competitions to increase model parameters, which creates unprecedented demand for infrastructure.
Companies, including Huawei, ZTE, H3C, and Sugon, are competing to build and deploy AI supernodes scalable from 1024 to 100 thousand GPUs. The increasing demands, driven by models like Kimi K3 and GLM-5, push computational capacity to its limits. Supernodes have replaced models as the central focus of the WAIC 2026 exhibition, as computational capability has become the main limiting factor in AI development. The shift from clusters with 8 cards to clusters of 100 thousand cards makes supernode capacity a direct competitive advantage; consequently, H3C demonstrated a 58% growth, and Inspur showed 41% growth over three weeks.
The most evident demand pressure is observed at the forefront of model development. GLM-5 models from Zhipu AI and Kimi K3 from Moonshot AI, boasting 2.8 trillion parameters, have been forced to raise prices and restrict subscriptions due to the enormous inference demand. Deploying the K3 model requires a supernode configuration with more than 64 GPUs, setting a minimum computational threshold for advanced models. In response, Zhipu acquired Zhongke Jiahe and built a domestic AI data center with a capacity of 1 GW, leading to a 37% rise in the company's stock in one day. The math of the situation is telling: industry participants report computation costs of approximately $0.80 per dollar of revenue from the model, yet every dollar of revenue corresponds to a multiple increase in valuation multipliers, making the scale of computation the primary lever for growth.
Existing achievements include the use of Huawei Ascend 950 with 1024 NPUs and deployed over 750 A384 sets, as well as ZTE OEX with 128 GPUs per rack and a cluster of 10 thousand. Sugon introduced Dawn 8000—the first domestic system with 100 thousand cards. H3C is building an AI factory capable of delivering 240–300 kW to a test point. Technological directions diverge: Huawei utilizes a proprietary full-stack model—from chip to interconnect and cooling—allowing for deep optimization across layers but requiring colossal R&D investment. ZTE and its partners prefer an open approach using multiple chips to avoid dependence on a single vendor and ensure flexibility for clients. Other players, such as Moore Threads, Pingtouge, and Baidu Kunlun, focus on specific cloud or vertical scenarios, paying close attention to inference efficiency and economic performance. The overarching challenge lies in maintaining system integrity across the architecture of chips, interconnect, software stack, and operational tools, as advantages gained in one area cannot compensate for systemic gaps.
Model parameters have not yet reached the 2.8 trillion mark; forecasts indicate reaching 10 trillion within two years. The consensus was clear: supernodes define the future. Companies that cannot secure computational power at the supernode scale will be structurally excluded from the AI frontier competition.
