The 2026 World Artificial Intelligence Conference (WAIC) commenced at the Shanghai World Expo Exhibition Center on July 17. This year marked the debut of a dedicated zone for the computing power sector, featuring over 200 domestic vendors. A notable shift was evident on the exhibition floor: the spotlight was no longer on spec sheets for individual chips but on server racks, clusters, and comprehensive system solutions.
A vendor's comment captured the industry's changing priorities: "Two years ago, people asked, 'How many cards do you have?' Now they ask, 'What's your cluster's MFU?'" This highlights the transition. Once a certain scale of computing power is achieved, maximizing the efficiency of every single processing unit becomes the true differentiator.
Sugon's 8000 (Dengfeng) ultra-large AI supercomputing system was selected as one of the conference's top ten "Treasures of the Hall." The Sugon 8000 adopts a hyper-intelligent fusion technology path, enabling native integration of scientific computing and intelligent computing. It supports full-precision computation from FP64 to INT8, covering diverse scenarios such as scientific computing, large model training and inference, and industrial simulation. The system integrates key components including chips, computing, storage, networking, cooling, applications, and services, providing full-stack coordination to meet composite computing demands for tasks at a scale of hundreds of thousands of processing units.
Achieving efficient operation at such massive scale tests the innovative capabilities of systems engineering behind ultra-large clusters. The Sugon 8000 employs a globally pioneering high-density rack structure, increasing the computing density per unit by up to 20 times compared to other super-nodes, enabling higher compute integration within limited space. Simultaneously, it utilizes a self-developed scaleFabric, an IB-like native RDMA high-speed interconnect technology, achieving ultra-high-speed and stable interconnection for hundreds of thousands of units, providing the network backbone for ultra-large-scale parallel computing and resource coordination.
Currently, the Sugon 8000 is connected to the national integrated computing power network via the National Supercomputing Internet. It has completed optimization for over 300 key applications, spanning more than 20 research and industrial fields including large models, robotics, innovative drug development, new materials, quantum computing, and astronomy/weather. Over 70 applications have achieved scaling to tens of thousands of units. It supports continuous operation of scientific AI applications like protein folding simulation, trillion-atom-scale water molecular dynamics simulation, and ultra-large-scale turbulence simulation, as well as AI applications such as large model training and scientific/industry-specific agents.
If Sugon and Hygon address the question of "where does the computing power come from," then Transwarp Technology answers "how to efficiently feed data to the computing power." A nine-year WAIC exhibitor, Transwarp this year showcased its GPU-native cognitive database, which for the first time deeply integrates GPU-native computing capabilities with cognitive abilities. It unifies data analysis, knowledge retrieval, trusted context provision, and long-term memory management within a single data foundation, achieving a leap from "data management" to "cognitive services."
This technological approach has been validated in industries like finance, government, and energy. Exhibition staff demonstrated an intelligent investment research scenario: an AI Agent, based on the GPU-native database, completed multimodal data retrieval and reasoning within milliseconds, with response speeds several times faster than traditional architectures.
This year, JiuZhangYunJi presented with the concept of an "AI Factory," driven by dual engines: a Training Factory and a Token Factory. Its core innovation lies in the DCU unified computing power measurement system—where 1 DCU = 312 TFLOPS × 1 hour. This system unifies the computing power measurement of heterogeneous chips from Nvidia, Ascend, Hygon, and others onto a single scale.
It's reported that the newly upgraded JiuZhang Intelligent Computing Cloud 3.0 achieves a full-process closed loop from "computing power metering" to "computing power governance." It integrates intelligent queue scheduling, resource quota management, and a development tool suite, transforming computing power from "something you can buy" to "something you can use well."
Client data validates this model: in a project with a leading model developer, comprehensive costs were reduced by up to 82.1%, and elastic scheduling of 1 to 999 GPUs effectively matched business fluctuation demands.
Lingxiong Technology approaches the computing power industry from another angle—computing equipment leasing. Its digital closed-loop DaaS (Device-as-a-Service) solution provides enterprises with full-stack services for AI servers, storage servers, and other computing equipment, including subscription, recovery, and maintenance, helping enterprises reduce initial investment by up to 97.4%.
"Many SMEs don't lack a need for computing power; they simply can't afford to buy or operate it. We enable computing equipment to be used on-demand, like water and electricity," a company representative stated. For robotics leasing, Lingxiong collaborates with leading firms like Unitree to offer integrated "robot hardware + cloud brain + local delivery" services, with pay-per-use and flexible lease terms. This model attracted intense interest during the exhibition, highlighting strong demand from SMEs for lightweight AI infrastructure.
In an interview, Zhongcheng Hualong pointed out that China is experiencing explosive growth in AI inference calls. Leveraging the massive domestic demand for inference iteration, domestic large model inference chips are carving out a differentiated path. "We have abandoned the traditional training-centric underlying design logic of training-inference integrated GPUs. By deeply coupling algorithms with silicon and reconstructing the computing paradigm, we are building a dedicated, inference-native heterogeneous computing architecture," said Wang Jiacheng, Chairman of Zhongcheng Hualong. Its HL200, HL200Pro, and HL400 AI chips will natively support FP8/FP4, with performance targeting international mainstream levels. They are designed to fully meet the inference demands of next-generation generative AI and AI Agent applications while achieving a fully domestic supply chain.
Walking through the entire computing power exhibition zone, three key trends are noteworthy. Competition has shifted from "comparing single cards" to "comparing clusters." The operational efficiency of clusters with tens of thousands or hundreds of thousands of units has become the core metric, with single-card performance numbers no longer the primary evaluation standard. Domestic substitution is moving from "usable" to "easy to use." Examples like the training loss curve for MoE-236B matching international levels, EvoPhys-World topping global rankings, and the Sugon 8000 supporting over 300 application optimizations represent industry-level validation, not just lab data. Computing infrastructure is permeating downstream sectors. A significant number of technical decision-makers from finance, manufacturing, energy, healthcare, and other industries were present, focused on one question: "Can this system run effectively in my business scenario?"
When computing power truly becomes a fundamental utility like water and electricity, this industry is just beginning its journey.
Comments