Author | ZeR0 Editor | Moying The largest AI event in history was astonishingly hot, both psychologically and physically. Amid a convergence of various factors, AI practitioners attending the World Artificial Intelligence Conference (WAIC) 2026 must have spent an extremely unforgettable few days, surrounded by a scorching atmosphere, explosive information, and Shanghai's weather oscillating between sweltering heat and thunderstorms. Even Yili and Luckin Coffee each took over a massive booth. Yili played on homophones, hanging an advertisement for "True AI (Love) Milk"; Luckin Coffee had already dived in, showcasing its AI coffee pop-up store, smart stores, AI Lucky ordering agent, and the Luckin AI friend circle. During this exhibition, the Zhidx team walked an unprecedented distance—over 71,000 steps in three days, rushing through more than 500 booths across three major halls. Combined with on-site exchanges, forums, and interviews, we have summarized the 15 key trends of WAIC 2026.
AI smartphones enter the second half, transitioning from "traditional phones + AI features" to "native agents." AI agent smartphones and robot phones are highly popular, set to reconstruct the agent operating system.
AI glasses and AI earphones feature more practical and focused functions, emphasizing lightweight design and long battery life. Some are competing in specific functions like translation and sports, while others target AI memory and proactive AI interaction. Home appliance giants were absent, and no one is telling the story of whole-house smart AIoT anymore. However, the muggy weather made all carbon-based organisms realize that some things AI can never replace, such as fans, air conditioners, and ice-cold drinks.
The hype around "OpenClaw Lobster" has passed, but agents are thriving, becoming the only main thread running through terminals, applications, models, computing power, and embodiment. Vertical industries such as power grids, banks, and telecom operators all have agents. In particular, enterprise agents, as "AI colleagues," shoulder the heavy responsibility of driving organizational evolution.
The booths of tech giants like Alibaba, Tencent, Baidu, and JD.com continued the "large and comprehensive" platform route, offering everything from super nodes, chips, large models, cloud computing, embodied intelligence, AI hardware, AI applications to agent matrices and experience zones, striving to be the infrastructure of the agent era. Although ByteDance did not exhibit, Doubao was everywhere.
Regarding flagship large language models, before people could mourn the widespread banning of Claude accounts, domestic Fable-class models collectively "kissed" the spotlight. Over these three days, the 2.8T parameter Kimi K3 went viral upon open-sourcing; some users found themselves in the gray-scale testing of the official DeepSeek-V4; Alibaba threw down the 2.4T parameter Qwen3.8 preview version, claiming it "only loses to Fable 5." A new MiniMax model with an estimated 2.5T~3T parameters is also expected to be open-sourced soon.
Observing the booth setups, the "Six Little Tigers of LLMs" are heading towards divergence: Step Star is cutting into the AI smartphone terminal track; Moonshot AI firmly holds the aesthetic ceiling; MiniMax serves up a full family bucket of large models and agent ecosystems; 01.AI set up a "Boss Personality Test" interaction; and Baichuan AI is going all-in on AI healthcare.
The narrative of domestic substitution is the real tailwind. Staff at multiple computing power booths specifically emphasized to us that their boards and complete machines achieved "100% domestic production." From chips to IDM merchants, everyone is emphasizing "super nodes." Smart computing centers have entered system-level competition, involving key links such as chips, interconnects, liquid cooling, storage, optical fibers, and computing-power-electricity coordination. In the future, computing will focus on optical computing, quantum computing, and space computing, and domestic pioneers have already emerged.
This should be the most comprehensive gathering of domestic top AI chip companies this year, publicly displaying their scaled landing achievements from chips, computing cards, servers, clusters, software to ecosystem applications.
Tokens replace FLOPS, becoming the new metrology and settlement unit across the computing power-model-application three layers. AI Infra vendors are changing their storylines; last year they were selling computing power, this year they are selling Tokens, collectively shifting their positioning to "Token shops" and "Token factories."
Data released by the MIIT today shows that China's quadruped robots account for nearly 70% of global sales, and humanoid robot whole-machine products reach over 400 models, exceeding half of the global total. Robots enter the year of "real work," but scene demonstrations are severely homogenized; wherever there is a performance, there is a crowd. Non-physical AI is simply defeated by human nature for watching the fun. Multiple booths built real-scene smart pharmacies, probably emptying all the drugs from the surrounding pharmacies.
Many humanoid robots and robot dogs that can walk around and perform have staff holding remote controllers nearby. Heat dissipation and battery life remain challenges, with "hanging," "broken down," and "Ge You slouching" miserable robots seen everywhere. However, as the industry shifts towards landing, the practice of having anime cosplayers lead robot dogs on patrol to attract traffic left many visitors with a poor viewing experience.
What robots perform is not important; the real value lies in the evolution of their "brain," "eyes," and "hands." Embodied intelligence forms a four-chain resonance closed loop of "whole machine—brain—dexterous hand—data." The route for the embodied intelligence brain remains diverse, focusing on embodied intelligence models and world models, with tech giants and large model enterprises starting to enter this track.
"Dexterous hands" involve tactile perception and fine manipulation, making them one of the most popular core components of embodied intelligence. In recent days, related startups such as Xinuo Future, Rimian Kaiwu, Mogan Technology, and Yimu Technology have completed new financing, with Yimu Technology's valuation breaking through 10 billion RMB.
Compared to pure text data, physical AI data is severely scarce. Centralized procurement has become a value highland. Current common collection methods include real-machine practical training, simulation, and wearable and non-embodied collection.
AI is continuously improving medical conditions, increasing medical accuracy, and providing convenience for the health protection of more people. Meanwhile, AI for Science is accelerating scientific research in key fields such as life sciences, critical materials, and earth meteorology. In this rapidly advancing AI era, agents and embodied intelligence are at their peak, technological development is magnificent, capital is expanding at doubling speeds, the birth rate of new unicorns is spectacular, and every industrial layer is seeking innovation and change. This article is very long, divided into 12 chapters (AI Smartphones, AI Hardware, Agents & AI Office, Large Models, Super Nodes, Chips, Token Factories, Robots, Embodied Intelligence Models, Dexterous Hands, Data Centralized Procurement, AI Healthcare & Scientific Research). Every booth was visited by us personally, and the information was recorded purely manually without any AI assistance, guaranteeing no hallucinations. If you want to understand the latest trends in a specific industry, you can directly scroll to the corresponding chapter. We hope this will serve as a very intuitive reference note.
01.The First Year of AI Agent Smartphones, Letting Phones Handle Tasks Themselves The hottest AI hardware during WAIC this year is the AI agent smartphone, aiming to let agents take over the phone and complete long-range complex tasks for you through cross-App operations. Among the three AI phones, the Honor Robot Phone has opened pre-orders across all channels since July 18, and real machines can be experienced at the Alibaba booth in the Expo Exhibition Hall and the Honor booth in Xuhui West Bund. The gimbal that can pop out to observe the world is the soul of this phone; it is very lively and cute when interacting with people, and can also play with more camera movements that fixed cameras cannot achieve during photography. The newly debuted Step large model native smartphone STEPX Neo this week was selected as a "Pavilion Treasure." Equipped with the self-developed agent native operating system Step AOS and the agent Amoo, it has reached the first batch of ecological cooperation with Ctrip, Alipay, Didi, Meituan, etc. Photography of the exhibits is not allowed, but you can watch live demonstrations of the agent doing things for you. For example, when a user gives the command "Help me deliver the file to the company, XXX is rushing me," the system automatically identifies the current location, calls contact information, matches the company address, and completes the errand service order. The entire process did not open any App, nor did it input commands step by step; the agent completed the task planning. The second-generation "Doubao Phone," the Nubia NaviX Ultra, which has not been officially released, is displayed in a glass cabinet at the ZTE booth, but it can only be looked at, not touched. Zhidx personally experienced the second-generation "Doubao Phone" at yesterday's ZTE Nubia media appreciation meeting. The Doubao Phone Assistant Intent Model 2.0 newly supports simultaneous submission of multiple tasks. For example, you can ask Doubao to open Douyin to find a video while simultaneously searching for flight tickets on Qunar and ordering an iced Americano on the Luckin App. Unprocessed tasks will automatically enter a queuing state. The three mobile phone "new species" have reached a consensus—system reconstruction: reorganizing computing, data, application, and service resources at the system level to build an agent-centric operating system. Currently, there are two main ways for system-level agents to take over Apps: first, the GUI Agent route, which obtains simulated touch permissions from the system bottom layer, "understands" the interface, and operates autonomously, bypassing App authorization; second, the protocol-based interconnection route, which can call authorized Apps. Recently, there is good news that has excited the entire mobile phone circle: on July 8, the AI model services of Apple, Huawei, OPPO, vivo, Xiaomi, Samsung, and ZTE Nubia collectively completed filing, equivalent to mobile phone AI being issued a "birth permit" by regulators. The cross-border alliance of "large model company + hardware manufacturing company" is becoming a mainstream organizational form. For example, the "Doubao Phone" is a collaboration between ByteDance and ZTE; Honor, in addition to its self-developed model, also announced a joint creation with Alibaba Qwen large model to develop a terminal large model solution deeply optimized for mobile phone scenarios; additionally, Qwen confirmed it will be integrated into the Apple Intelligence for the China version; and ModelBest's on-device model will be loaded on Samsung phones. The national App Alipay has also cooperated with OPPO and Step Star to support mobile phone AI assistants calling the AI version of Alipay's agent services. In mobile photography, Intsig's CamScanner demonstrated a very useful "anti-reflection" technology, which can remove the pain points of reflection blocking and shadow interference that easily occur when shooting museum exhibits and ID materials with one click. Besides phones for adults, Lingyuzhou created the AI terminal "Little Square Machine" for children, which can be held or worn on the chest, transforming the real environment into a conversable and learnable intelligent space, accompanying children to learn whenever they ask. At the opening ceremony of WAIC 2026, Yin Qi, Chairman of Step Star, proposed that the agent wave will promote three structural changes: new systems, new carriers, and new networks. At the new system level, agent capabilities are jointly determined by model capabilities and agent operating system capabilities. Agentic OS decides how "far" the agent can go by connecting models with data, tools, interfaces, and devices; at the new carrier level, terminal design concepts will further move from "human-centric" to "human-machine symbiosis." Computers, mobile phones, cars, and robots will become different "bodies" of the same agent in different scenarios, supporting intelligent cross-terminal continuous existence, free migration, and collaborative task execution; at the new network level, the A2A network will connect humans and agents. Agents will have their own identity, capabilities, and credit, and will autonomously find partners, organize collaboration, and complete transactions, becoming important infrastructure for the agent economy.
02.AI Glasses and AI Earphones Form Differentiated Routes, Home AIoT Grows an Agent Brain AI wearable devices pay more attention to lightweight design, natural interaction, and seamless collaboration, allowing AI to integrate into people's daily lives more imperceptibly and handle tasks for people more proactively.
AI/XR Glasses The "Hundred Glasses War" is forming differentiated routes. During this WAIC, over 25 AI glasses were exhibited, generally weighing less than 50 grams, with the lightest at only 14.9 grams. Functionally, they do not pursue being large and comprehensive, but each has its own trade-offs. Besides general AI glasses, some are mainly aimed at segmented scenarios like translation and sports, such as BleeqUp's 4-in-1 sports shooting glasses and Liangliang Vision's professional AR translation glasses. Among them, iFLYTEK AI glasses were selected as a "Pavilion Treasure," focusing on simultaneous interpretation translation, including call translation, online simultaneous interpretation, face-to-face translation, and photo translation, supporting translation in 122 languages. Xingyi Intelligence exhibited the world's first holographic optical waveguide smart glasses, weighing only 28 grams, with newly added smart ring control. Active memory and all-day wear have become a major focus of competition. AI glasses manufacturers have reached a consensus: as the long-term memory of agents becomes infrastructure, glasses worn on the face all day are the most natural entry point for collecting and invoking personal memory. Qwen AI agent glasses added full-duplex voice, eye tracking, and real-time user vital sign monitoring functions, enabling the device to generate personalized health reports and form agent memory. Ant Willit AI glasses cover functions such as scan-to-pay, city companion tour, and real-time information reminders, focusing on a lightweight AI entry point for high-frequency daily use. Xiaodu AI glasses obtained the L3 level test certificate of the national standard series "Artificial Intelligence Terminal Intelligence Grading." Li Weike Technology released the X-AI memory glasses. Based on the AI memory capability system, it can extract, understand, and summarize key information heard throughout the day to form searchable structured memory, making the agent understand you better. Xinmou Tech's Moonix AI glasses just went on sale, with a whole machine weight of only 14.9 grams, focusing on "imperceptible recording + AI generation," cutting out functions like translation, teleprompter, and navigation. VITURE XR smart glasses and the Pro neck ring were created by the original Google Glass team members, showcasing 6DoF gesture spatial computing game functions and 2D to 3D experiences on-site. JD Aidol Creation Camp exhibited AI/AR glasses such as INMO GO3, XREAL 1S, and RayNeo X3 Pro; Rokid released the glasses native AI operating system YodaOS; Weiguang Xuanjing focuses on modular "hardware evolvable"; MoJie, based on self-developed resin diffraction optical waveguide technology, reduced the glasses' weight to 25 grams. INMO's AIR and GO series AR glasses and Yudong Ruiyi's HOYi consumer-grade AR glasses also appeared at the venue. Overall, the core components of AI glasses are highly domesticated, but this has also intensified hardware homogenization, with differentiation shifting towards operating systems, software, and ecosystem services.
AI Earphones AI earphone exhibits generally adopt an ear-clip design to reduce burden, with functions converging on translation, minutes, and health. Alibaba Qwen's first AI agent earphone was exhibited for the first time, adopting an ear-clip design. By integrating Qwen AI assistant capabilities, it supports simultaneous interpretation, meeting minutes, health recording, and other functions. NetEase Youdao's first exhibition of Youdao OpenPods strings together recording, transcription, summarization, and translation into a closed loop for meetings and cross-language communication, supporting simultaneous interpretation, meeting minutes, precise recognition and correction of professional terms, and for the first time turning the earphone case into an independent recording and translation terminal. Lightfan Tech's AI earphone with a camera, Lightwear, appeared at the MiniMax booth, capable of real-time recognition of real scenes and proactively reminding users of memos; Nothing's multiple earphones appeared at the Baidu Intelligent Cloud booth. Mobvoi also exhibited the TicNote series AI recording earphones and AI smart recording watches, and hung a wall of AI recording cards.
Smart Home AIoT Facing the home scenario, Xiaodu released the "Home Agent Hub" strategy, supporting the agent upgrade of Super Xiaodu, capable of autonomously decomposing goals, coordinating and scheduling tools, and delivering tasks in a closed loop. JD.com built a real-scene AI Home space, arranging AI TVs, AI tea bars, AI projection desk lamps, smart mattresses, Kaimisi AI smart robot dogs, and AI West Highland Terriers in the living room, study, and bedroom.
Other AI Terminals Yanshan Tech demonstrated a brain-controlled game interaction system based on Yanshi brain-like self-research, where players control the game "Black Myth: Wukong" with their minds. Quectel created a smart sleep acoustic solution SleepHub for electric bed manufacturers, supporting continuous dialogue, active and passive voice broadcasting, and featuring snoring detection, dual-zone space recognition, and bed posture linkage. L'Oréal's PPD professional hair styling magic mirror allows consumers to virtually experience a variety of overall styling effects, including hairstyles and makeup. There are also many diverse AI terminals and application displays, such as Quwan Network's generative AI guitar, Xianghang Tech's naked-eye 3D real image technology, uAI gesture interactive naked-eye 3D display system, and ViewX Liber 15.6 naked-eye 3D spatial intelligent AI display.
03.Agents Become "AI Colleagues" and "AI Employees," AI Office Transforms Organizational Productivity Industry trends change so fast. OpenClaw, which went viral at the beginning of the year, once set off an agent storm, but now it is a thing of the past. Only a few "shrimp boxes" and "lobster all-in-one machines" in the exhibition area maintain its presence. Replacing it is the comprehensive explosion of agent software, hardware, platforms, and infrastructure. Tech giants like Alibaba, Ant, Baidu, Tencent, and JD.com all exhibited comprehensive layouts for agents, ranging from agent applications to infrastructure such as agent development platforms, model services, and computing power platforms. Taking Tencent as an example, the booth packaged and displayed the "Efficiency Agent Toolset," including QClaw, ima, QQ Browser, and Marvis for individual users, as well as WorkBuddy, CodeBuddy, and Agent Suite for workplace office and enterprise users. The WeCom Agent "Da Yuan" was also exhibited for the first time. In terms of hardware, companies like Lenovo, iFLYTEK, 6Lian Intelligent, Moore Threads, and Cixin Technology exhibited terminal products such as AI PCs, Agent Computers, AI computing modules, "lobster" boxes, desktop AI supercomputers, and home AI hubs; SOPHGO displayed the OpenClaw large lobster all-in-one machine; China Unicom made a Yuanjing UniClaw portable AI terminal that stuffs an AI workstation into a USB drive. Houmo Intelligent also jointly launched a holographic interactive personal smart computing center ClawHouse X1 Pro with Liaoqu Intelligent, capable of running real-time screen rendering, large model local inference, and low-latency voice interaction. In terms of software, Baidu Intelligent Cloud's general agent Baidu Dazi was selected as a "Pavilion Treasure." It can autonomously decompose steps, call tools, and complete the full-process delivery of file processing, data analysis, and PPT creation. In NetEase Youdao's full-scenario Agent application matrix, the office assistant Agent Youdao LobsterAI can schedule files, terminals, browsers, and local projects to complete complex office tasks. The innovation of enterprise-level agents lies in upgrading from single-point tool empowerment to business process reconstruction, ensuring AI is controllable, efficient, and replicable within the organization. Kingsoft Office's organizational-level AI office product WPS Comate can achieve three connections (knowledge, data, capabilities), two managements (cost, security), and one platform (unified platform), integrating tasks, experts, teams, skills, Wiki, applications, automation, and other capabilities, distilling personal experience into organizational capabilities. Mobvoi's enterprise-level AI-native collaboration platform CodeBanana positions itself as a "collaborative version of Codex," taking "projects" as the core unit. Each project simultaneously exists as a trio of group chat, Agent, and independent workspace, helping organizations upgrade AI from a personal tool to a collaborative, executable, and precipitable organizational operating system. Lingyang brought four types of AI employees—marketing, customer service, e-commerce operations, and sales—to the scene, demonstrating agents automatically running market insights, breaking down creative plans, and providing landing strategies on the spot. Xianyuan Tech's "Leadeep AI" consists of digital avatars, AI expert think tanks, and organizational engines, forming an individual-team-organization intelligent closed loop, helping enterprises precipitate business processes and promote autonomous organizational intelligence evolution. Yunzhisheng, Midu, Fourth Paradigm, LanZhou Tech, Tezign, Dipu, and Boshiyun also displayed enterprise-level agent operating systems, platforms, or products. Transwarp displayed a GPU-native cognitive database designed for agents, integrating SQL analysis, knowledge retrieval, memory, and context management capabilities on a unified base; it also simultaneously launched cognitive database cloud services, including "Data Factory" and "Token Factory" services, to optimize enterprise data and knowledge services and enterprise-level AI inference efficiency. Agent deployment is inseparable from security protection. Cisco displayed a zero-trust security architecture covering the full life cycle of AI agents. Sangfor Technology displayed how agents equipped with security guardrails can accurately intercept attacks. Facing content creation scenarios, Kujiale brought a new AI video creation Agent integrated with its self-developed world model, which can convert floor plans into interactive 3D virtual studios; Wondershare upgraded its film-level storyboard director Agent for AI short drama production; ZhiXiang Future also released the world's first unlimited-duration content creation agent during WAIC. Facing education and training scenarios, Hailiang Tech formed a complete education Agent system including smart lesson preparation, adaptive learning, multimodal classroom analysis, and digital teacher "Xingledu." Dance with AI turned its booth into an AI learning laboratory; ZhiShi YiYou focuses on enterprise training, bringing AI coaches into sales and service scenarios. Facing the power grid scenario, China Southern Power Grid's AI-native distribution network planning agent "Da Wate · YunRui" was selected as a "Pavilion Treasure." It can autonomously understand tasks, call power professional tools, and complete problem diagnosis, scheme generation, simulation verification, investment evaluation, and report compilation. Facing industrial scenarios, Siemens' Eigen engineering agent can improve engineering efficiency by up to 50%. Haier's COSMOPlat demonstrated how its industrial AI assistant directs agents to work collaboratively. Facing swarm intelligence, China Telecom Research Institute's Smart Transmission Network exchanges computing for bandwidth, greatly reducing bandwidth consumption and achieving efficient multi-agent collaboration; Nativex strings together the full marketing chain of market analysis, creative generation, and ad placement through multi-agent collaboration. Qiming Venture Partners Managing Partner Zhou Zhifeng predicts that in the next 12-24 months, AI-Native organizations will move from concept to proof, and a group of enterprises will achieve per capita output several times that of traditional organizations.
04.Large Model Pursuit: Towards 3 Trillion Parameters, Waiting for "Domestic Fable 5" In early 2025, DeepSeek-R1 aligned with OpenAI-o1 at extremely low cost, turning "open source + high efficiency" into the collective route for Chinese large models. Entering 2026, the Hong Kong stock market gave birth to the "first large model stock," the hundred-model war calmed down, developers became disenchanted with "topping the leaderboard," and turned their attention to practical effects and cost-effectiveness. Of course, as it stands, whoever first becomes the recognized Chinese version of Fable 5 will become the next "domestic light" standing in the C-position. This WAIC exhibited a total of 261 large models, but the headline was occupied by something outside WAIC: Moonshot AI released the 2.8 trillion parameter Kimi K3, surpassing the 1.6 trillion DeepSeek V4-Pro to become the world's largest open-source weight model, and its scores in some benchmark tests surpassed Fable 5. But at the booth, Moonshot AI consistently adopted a high-end design style, presenting exhibits extremely restrainedly, with only one Kimi K3 poster and two screens, mainly promoting through audiences completing tasks on social platforms to get rewards. Among the "Six Little Tigers of LLMs," except for Zhipu AI, the other five all participated. MiniMax mainly displayed the ecological progress of its large model and agent family bucket. For example, there were hardware demonstrations like Vita Dynamics robot dogs and AI guitars on-site, and you could also play the claw machine to take away MiniMax's red little octopus. Besides exhibiting the Step series model matrix, Step Star differentiated itself with AI terminals, setting up agent smartphone and agent car exhibition areas.01.AI displayed the Wance platform, Boss AI, Investment Officer AI, and Sales Champion AI, helping enterprises make key decision judgments and executions, and set up a personality test interactive game "What kind of boss are you" at the booth. Baichuan AI focused on AI healthcare, displaying the Baichuan-M4 medical large model, the BaiXiaoYi application, and AI medical enhancement capabilities co-built with top-tier hospitals. From a technical route perspective, "sparse architecture + million Token context" has become standard for flagship models. The MoE architecture can expand model scale without sacrificing inference speed. Model capabilities are moving towards balancing intelligence and efficiency, continuously strengthening long-range programming and agent reasoning capabilities. Qiming Venture Partners Managing Partner Zhou Zhifeng predicts that in the next 12-24 months, top models will internalize most "plug-in" capabilities, including task planning, tool calling, multi-Agent collaboration, and some Harness engineering capabilities. In terms of multimodal large models, SenseTime's SenseNova U1 Pro produced stunning visual effects in image generation, capable of generating both high-definition and delicate Chinese paintings of the 24 solar terms and workflow posters with dense formulas and rich text details. Shengshu Technology's MotuBrain world model can real-time analyze actions captured by cameras, analyzing and generating 3D motion data; Yingmou Technology and VAST focused on exhibiting 3D generation models. Lianhui Tech displayed the on-device AI landing ecosystem achievements of multimodal large models, such as AI visual assistance and embodied equipment; Mosi Intelligence, originating from Fudan University's MOSS team, makes multimodal large models based on situational intelligence, setting up interactive gameplay like voice personality tests and AI podcast production at the booth. Vertical large model exhibits cover all walks of life, such as Hikvision's Guanlan large model for the smart IoT industry, GeGe AI's music large model, Zhongke Wenge's decision-making large model, PetroChina's Kunlun large model, PipeChina's pipeline industrial world model, COSMOPlat's industrial large model, Guotai Haitong's trusted and controllable large model, Yaosu Tech's 3D human-derived organoid model, etc. China Tower also exhibited multiple segmented industry large models.
05.The First Year of Domestic Super Nodes, 100K-Card AI Super Cluster Debuts Models define computing power demand; computing power determines model cost. When releasing the DeepSeek-V4 preview version in April this year, DeepSeek wrote a preview: "It is expected that after the Ascend 950 super node is mass-produced in the second half of the year, the price of DeepSeek-V4-Pro will be significantly reduced." Moonshot AI also suggested "deploying Kimi K3 on a super node configuration of 64 or more accelerators." Super nodes are the number one keyword in this year's WAIC computing power exhibition area. According to the "Super Node Development Report" released by Huawei and CAICT, super nodes weave hundreds or thousands of AI processors into a logically unified high-density computing body through nearly non-blocking high-bandwidth interconnection. Huawei's Ascend 950 super node, Alibaba Cloud's Panjiu AL128 super node, Sugon's fully domestic100K-card AI cluster, and ZTE's OEX super node were all selected as WAIC 2026's top ten "Pavilion Treasures." Huawei's Ascend 950 super node was exhibited in a real-machine form as the "industry's largest-scale super node." It high-speed interconnects 1024 NPUs, achieving a logical level like a single machine, providing 1 EFLOPS FP8 computing power, 2 TB/s ultra-large bandwidth per card, global 256 TB memory unified addressing, and 3us RTT ultra-low latency. According to previous calculations by SemiAnalysis, the performance of a single Ascend chip is about 1/3 of NVIDIA's Blackwell, but the BF16 cluster performance of the previous generation Huawei CloudMatrix 384 super node has reached 1.7 times that of the GB200 NVL72, and its total memory capacity is 3.6 times that of it. The route of making up for single-point gaps through system innovation has proven highly effective. Sugon's Sugon 8000 "Dengfeng" is a fully domestic 100K-card AI super cluster, adopting a "super-intelligent fusion" technical route, abandoning traditional partitioning methods to achieve native integrated fusion of all types of computing, facing the composite needs of high-precision scientific computing and low-precision intelligent computing, supporting full precision from FP64 to INT8. Alibaba Cloud's Panjiu AL128 super node, equipped with the Zhenwu M890 chip, adopts a single-cabinet 128-card design, achieving Pb/s-level bandwidth and hundred-nanosecond-level latency through an orthogonal cable-free architecture and the Scale Up interconnection ALink System. ZTE, together with optical chip and AI chip partners such as Lightelligence, Biren Technology, MetaX, Enflame, and Iluvatar CoreX, created the OEX+dOCS domestic high-performance cluster super node, pioneering orthogonal backplane-free "zero-cable" interconnection, optoelectronic fusion networking, hundred-nanosecond-level low latency, reducing interconnection costs by 80%, and dynamically reconstructing topology according to computing power needs. H3C's UniPoD S80000 super node covers a full series of products from 32 cards to 1024 cards, expandable up to 16,384 cards, with training performance improved by 70% and inference performance improved by 3 times. Booths of Inspur, Lenovo, xFusion, CEC Cloud, Ruijie, Qianshi Tech, etc., all had super node exhibits. Luxshare Intelligent Computing exhibited the 10K-card level AI cluster TITAN-iX. Multiple AI chip companies also exhibited important super node achievements, including: Kunlunxin: 32/64-card super node products and 256-card super node clusters. Moore Threads: MTT C256 super node, cabinet-merged up to 256 cards. Tsingmicro: Reconfigurable super node system, integrable with 4096 TX81 chips. Lightelligence: Optical Leap LightSphere X distributed optical circuit switching all-optical super node achieves 2000-card scaled landing. Linxi Tech: Brain-like super node cluster LynAInfra, cascading thousands of brain-like chips at the single-cabinet level. Zhonghao Xinying: Super node, directly connecting up to 2048 Xuyu chips. Enflame: Orthogonal architecture solution YunSui ESL64-O super node, Cable-tray solution YunSui ESL64-C super node. XiWang: HuanWang super node, up to 256 cards per cabinet. MetaX: Xijing S600 super node, starting at 64 cards per cabinet, expandable up to 320 cards. Taichu Yuanqi: 128-card super-intelligent fusion computing system HyperIntelliX. Domestic computing power vendors are exploring how to efficiently organize chips, exploring more effective paths towards high computing power with system-level capabilities. Optical communication, liquid cooling, and computing-power-electricity coordination have entered the core agenda. ZTE OEX's orthogonal backplane-free "zero-cable" design, Matrix super node's computing-network integrated liquid cooling, H3C's computing-network fusion, Luxshare Precision's high-speed interconnect components, CICT's 6.4T NPO and 10K-core level ultra-large core-count optical cables, YOFC's hollow-core optical fibers, Vertiv's thermal management and power supply and distribution, Schneider Electric's operating systems, and other solutions jointly form the engineering base for future ultra-large-scale AI clusters. In terms of computing-power-electricity coordination, SenseTime's SenseCore released a computing-power-electricity coordination Agent, achieving an 80% increase in Token output per unit of electricity cost, with an average electricity price 10% lower than that of IDCs in the same region; Damao Tech launched the Computing-Power-Electricity Coordination 2.0 Platform · AIDC Green Power Direct Connection Energy Operation Operating System; Shendu Zhikong and China Southern Power Grid both displayed solutions for computing-power-electricity coordination and liquid cooling intelligent control. On the storage side, Seagate's Magic Box 4+ can provide the world's highest density mechanical hard drive storage solution, with a single disk capacity up to 44TB; Toshiba set up an AI photo machine; Western Digital exhibited the world's first 11-disk commercial CMR mechanical hard drive platform, with a single disk capacity up to 30TB; DapuStor, Memblaze, and Xinzhansu displayed enterprise-level SSDs; Zentech displayed DRAM particles and AI storage-computing solutions; Derui Lingxin displayed the PCIe 5.0 storage product matrix. Facing optical computing and future data centers, Lightelligence displayed an optical matrix of "optical computing + optical interconnection + co-packaged optics." The Optical Leap LightSphere X optical interconnection and optical switching super node cluster, jointly researched by Lightelligence, ZTE, and Biren Technology, achieved the landing of thousand-card results, marking the official completion of the first domestic, commercialized optical switching thousand-card computing power cluster. Facing quantum computing, Zhongqi Wuliang released the world-leading computing center-level neutral atom quantum computer "Qinghe No.1"; SuanFeng released the domestic first set of super-quantum fusion mixed training computing base Shanghai Cube Pro, connecting domestic quantum computers and domestic high-density super node GPU computing power clusters, expected to be delivered to customers in the second half of this year; DataCanvas and Bose Quantum announced full-stack adaptation and collaboration around smart computing clouds and commercial optical quantum hardware; TuringQ released the quantum-classical hybrid agent platform QAgent. Facing space computing, ASPACE displayed the world's first space computing constellation composed of 12 computing satellites, with a total computing power of 5 POPS; SenseCore and ASPACE cooperate, expecting to build a SenseCore space computing power constellation composed of thousand-level computing satellites and over 10,000 P total computing power by 2030; UCloud and USST released a "space computing power" prototype; Enflame and Xingji Tiansuan jointly released "Space-based Computing Power Application Scenarios."
06.AI Chips Move Towards Scaled Landing, Exploration of Fully Domestic Large Computing Power Chips Shows Initial Results The development of domestic large models has begun to be linked to the production capacity of domestic AI chips. From WAIC exhibits, domestic AI chip architectures are diversified, with routes including CPU, GPU, TPU, DSA, reconfigurable, data flow, near-memory computing, and storage-computing integration. Large computing power chips are iterating towards large VRAM, high bandwidth, multi-precision support, Advanced Packaging, and multi-Chiplet. (1) GPU-type cloud AI computing chips: Alibaba, Kunlunxin, Hygon, Moore Threads, Xindongli Tech, Biren Technology, Iluvatar CoreX, MetaX, XiWang, Denglin Tech, and Lisuan Tech all exhibited their latest GPU chip products. (2) Non-GPU-type cloud AI computing chips: Including Huawei Ascend chips, Tsingmicro reconfigurable computing chips, Yixing Intelligent RISC-V cloud AI large computing power chips, CloudWalk large computing power inference chips, Dongfang Suanxin near-memory computing 3D chips, Lightelligence optoelectronic hybrid computing chips, Guangbenwei optical in-memory computing chips, Linxi Tech brain-like computing chips, Zhonghao Xinying TPU, Weiyan Tech3D heterogeneous memory architecture large model inference chips, and Taichu Yuanqi heterogeneous fusion computing chips. (3) Edge and device AI computing chips: Including Listenai large model AI inference chips, AXera edge AI chips, Houmo Intelligent storage-computing integration chips, SOPHGO TPU, Hanxu Tech MRAM+SRAM weight resident data flow chips, Gravity full-function spatial computing chips, Rockchip AIoT chips, Huixi Intelligent embodied intelligence chips, Lingchuan Tech smart video processing chips, Tashan Tech dynamic tactile perception chips, and Yintai Tech AI capability chips (Nexus). (4) Automotive AI chips: BYD, Li Auto, Black Sesame, NIO Shenji Technology, and SiEngine all exhibited self-developed smart driving chips. (5) AI PC SoC: Cixin Tech's Agentic SoC full-scenario product system made its debut. (6) CPU: Including T-Head Arm server CPUs, SpacemiT RISC-V AI CPUs, Lanxin Computing RISC-V+AI smart fusion server CPUs, and SOPH