TMTPOST— The 2025 World Artificial Intelligence Conference (WAIC) and High-Level Meeting on Global AI Governance is in full swing, and this year’s edition is breaking records across the board, highlighting both the surging global appetite for AI and the rising prominence of embodied intelligence.For the first time, WAIC’s exhibition footprint has surpassed 70,000 square meters, with more than 800 companies showcasing over 3,000 innovations. The lineup includes more than 40 large AI models, 50 AI-powered terminal devices, 60+ intelligent robots, and over 100 products making their global or China debut.Despite high temperatures and intermittent rain in Shanghai, crowds swarmed the venue. Tickets, originally priced at 168 yuan, have been scalped for as much as 3,000 yuan in the resale market, reflecting a near-unprecedented level of public and industry interest.Tech giants including Huawei, Alibaba, Tencent, Baidu, SenseTime, Lenovo, Tesla, and Siemens dominated the exhibition floor with a wide array of next-gen offerings. But it was the embodied AI robots—particularly humanoid and quadruped forms—that captured the public’s imagination, drawing larger crowds than even the flagship AI models and computing platforms.Huawei’s unveiling of the Ascend 384 supernode attracted intense attention, while robot boxing demos at Unitree’s booth caused foot traffic bottlenecks as attendees jostled for space and smartphone shots. Livestreaming and influencer videos filled nearly every corner of the venue.Still, the event wasn’t without notable absences. Major global players such as NVIDIA, Microsoft, and AMD skipped the conference, as did Chinese large model developers Baichuan Intelligence and 01.AI. Cambricon and Hygon—leaders in AI chips and computing infrastructure—were also missing.Yet the enthusiasm on display overshadowed these omissions. At the AI Gala on July 26, Unitree Robotics founder Xingxing Wang and AgiBot co-founder Peng Zhihui (known widely as “Zhihui Jun”) were named “2025 AI Persons of the Year.” Other honorees included Zhipu AI CEO Peng Zhang and BrainCo Technology founder Bicheng Han.On the first floor of the WAIC exhibition center in Hall H2, spanning over 70,000 square meters, a cluster of figures stood out.By June 2025, there were 35,042 AI companies globally, with 5,170 based in China—around 15% of the total. Among 271 global AI unicorns, 71 are Chinese, accounting for roughly 26%. Between 2014 and 2024, Chinese researchers published over 510,000 AI-related academic papers, more than 34% of the global output. Over the same period, China was granted more than 350,000 AI patents—over half the world’s total.These numbers reflect China’s growing weight in global AI, both in research and industrial scale.This year’s most anticipated WAIC showcase was the debut of Huawei’s Ascend 384 supernode, a long-awaited real machine that cost billions of yuan to develop.Huawei debuted its Ascend 384 SuperNode—also known as the Atlas 900 A3 SuperPoD—marking the product’s first offline showcase. The advanced system was recognized as a “Treasure of the Pavilion,” highlighting its significance within this year’s exhibition.The SuperNode is built on Huawei’s proprietary architecture, utilizing high-speed bus interconnection technology to link 384 NPUs with ultra-high bandwidth and ultra-low latency. This setup effectively overcomes communication bottlenecks between compute, storage, and other cluster resources, allowing the system to function with the operational fluidity of a single machine. Optimizations at the system level enable seamless, efficient scheduling and resource management across the entire node.First introduced in May during the Kunpeng Ascend Developer Conference, the Ascend SuperNode delivers industry-leading performance across three dimensions: bandwidth, latency, and model efficiency. Compared with traditional clusters, it offers 15x faster communication between AI processors, 10x lower single-hop latency, and support for global memory addressing. In benchmark scenarios, it’s the only industry platform to break the 15ms decode latency threshold.Real-world tests show that the SuperNode delivers more than 2.5x higher training performance on large-scale dense models such as LLaMA3, and over 3x speedups on communication-heavy workloads like multimodal and Mixture-of-Experts (MoE) models, including Qwen and DeepSeek.In addition to the hardware reveal, Huawei’s booth featured demonstrations of its broader Ascend AI computing stack, including training and inference platforms, open-source software and hardware ecosystems, and industry-specific applications spanning telecommunications, finance, energy, education, and research.Beyond Huawei, more than 30 other AI compute vendors—including Moore Threads, Enflame, Biren Technology, Houmo Intelligent, Metax, and Moxin AI—also showcased new products, underscoring the fierce competition and rapid innovation in China's AI infrastructure race.GPU developer Moore Threads positioned itself as a full-stack AI infrastructure contender, showcasing a broad lineup of AI products and solutions anchored by its in-house general-purpose GPU. Built on its self-developed MUSA architecture, the company’s full-featured GPU chips support a wide range of workloads—AI training and inference, graphics rendering, scientific computing, physical simulation, and UHD video processing—all on a single chip design.The company’s booth featured its general-purpose accelerated computing platform, Kuai’e intelligent computing clusters, and vertical industry solutions. Moore Threads emphasized its expanding capabilities in areas including large model development, embodied intelligence, AI for Science (AI4S), and simulation.Founder and CEO Zhang Jianzhong has previously said Moore Threads aims to build an “AI Factory” through system-level engineering innovations to support next-generation model training. He cited the release of DeepSeek V3 as a signal of the rapid advancement in AI capabilities, noting that model intelligence levels have surged by 50% in just six months—outpacing the natural evolution of human intelligence.Enflame Technology marked its largest-ever presence at the 2025 World Artificial Intelligence Conference by unveiling a suite of new AI computing products, including the Enflame S60 inference card and the DeepSeek integrated system series, set for launch in early 2025. On July 27, the company also introduced its fourth-generation all-in-one training and inference platform—the Enflame L600—and debuted a new high-performance computing system: the YunSui OGX series. The flagship OGX400 model connects eight OAM cards within a single machine, featuring 1,152GB of onboard memory, 28.8TB/s storage bandwidth, and a 2.8TB/s total communication bandwidth—enough to fully run DeepSeek’s full-power version on a single unit.Enflame reported that over 70,000 units of its S60 card have been deployed to date, placing it among the top domestic providers in terms of scale. Its solutions now support over 300 application scenarios, ranging from large language models and multimodal systems to search, recommendation engines, and traditional AI use cases. The company has contributed to five national intelligent computing clusters, serving more than one billion end users, and offers full-stack solutions for models of various sizes and compute demands. Looking ahead, Enflame is preparing to scale from tens of thousands to hundreds of thousands of cards in its next-generation deployments.Moore Threads also returned to WAIC, highlighting its Xiyun C600 GPU and offering a full-stack presentation spanning chip-level innovation to cluster-level systems. In contrast, Tianshu Zhixin was notably absent from the exhibition for the first time.Birentech partnered with Xizhi Technology and ZTE to unveil the LightSphere X—billed as the world’s first distributed GPU supernode solution using optical interconnects and switching. MXAI introduced its MX S40 Active Cooling computing card, optimized for rapid deployment across a range of AI workloads with zero configuration required.Edge AI company InnoStar demonstrated its P1 chip across three application domains: PCs, edge platforms, and automotive computing. Built on a 6nm process with a 12-core Arm architecture and integrating CPU, GPU, and NPU, the chip delivers 45 TOPS of edge AI compute and has entered the commercialization phase.CorePower Technology showcased its latest AI processors and solutions, while Axera presented its new edge computing roadmap and products tailored to distributed AI deployment.Yunmai Xinlian brought its proprietary AISC-based YSA-100 interconnect chip to WAIC, alongside three series of smart NIC and DPU products: metaScale, metaConnect, and metaVisor. GMI Cloud introduced an AI cost calculator that demonstrated its cloud solutions can lower overseas IT spending by over 40% and reduce the investment payback period to one-third of the industry average.AI infrastructure newcomer InfiniteCore made its WAIC debut with a broad lineup of performance acceleration tools. These included the InfiniteCore AI Cloud, an intelligent computing platform, and terminal AI solutions geared toward cross-regional clusters, heterogeneous environments, and streamlined AI deployment.InfiniteCore CEO Xia Lixue described today’s computing power landscape as analogous to building giant power plants—while application-layer AI startups need “plug-and-play” infrastructure. Xia emphasized the importance of creating agile, modular AI infrastructure to meet the needs of small and midsize developers seeking flexible, scalable compute solutions.Qingcheng AI debuted Chitu, a domestically developed, open-source large model inference engine built in partnership with Tsinghua University. Positioned as production-ready, Chitu supports deployment across major LLMs and local computing platforms. The company also showcased Bagualu, its in-house high-performance large model training system, which has been validated on several large-scale domestic clusters and deployed in its AI chatbot and intelligent painting platforms.PPIO, which recently filed for a Hong Kong IPO, launched what it called China’s first Agentic AI infrastructure service platform. Co-founder and CEO Yao Xin said PPIO is the country’s first infrastructure provider focused on Agentic AI, offering a full-stack platform—from the nation’s first Agent sandbox to model services and end-to-end agent tools—for developers and enterprises.Beidian Digital Intelligence unveiled its Qianjin·AI Computing Platform, which enables unified scheduling and efficient collaboration across heterogeneous compute resources. It also launched the Xintian·Agent Platform, which supports the flexible orchestration of mainstream models and industry-specific algorithms via a low-code/no-code development interface. Its retrieval-augmented generation (RAG) system reportedly achieves over 95% accuracy, boosting development productivity by more than tenfold.The Shanghai Institute for Science and Intelligence, together with Fudan University and startup Infinite Lightyear, launched the Xinghe Qizhi Scientific Intelligence Open Platform. Targeted at the research community, it integrates high-value scientific data, open-source models, intelligent computing, and dry-wet lab experiment loops. Qi Yuan, Distinguished Professor at Fudan and founder of Infinite Lightyear, said AI’s evolution is shifting from a race for scale to one driven by value creation, trust, and industry fluency—especially in science, finance, and healthcare.While some continue to chase computing supremacy, many of China’s top tech firms are focusing on practical AI deployments. Companies including Alibaba, Ant Group, JD.com, Tencent, Baidu, SenseTime, NetEase, Kuaishou Keling AI, China Telecom, and Tesla, are prioritizing use-case-driven implementations in complex real-world environments.Alibaba has officially joined the so-called “Hundred Mirror Battle,” with subsidiary Quark unveiling its first self-developed AI glasses. The device has completed R&D and is expected to launch this year. On the enterprise front, Alibaba demonstrated three open-source large models that claimed top rankings in international benchmarks. Ye Jieping, VP of Alibaba Cloud, said Tongyi Qianwen now supports full-scale, multimodal open source, with over 400 million downloads and 140,000 derivative models. The company plans to invest RMB 380 billion over the next three years in cloud and AI infrastructure.Ant Group showcased a suite of AI-powered applications across healthcare, finance, and daily life, headlined by its AQ Health Assistant, which won the WAIC SAIL Award TOP30 and “Treasure of the Exhibition.” Since its soft launch on Alipay last September and standalone app release in June, AQ has served over 100 million users.Built on Ant’s proprietary medical large model, AQ enables voice-based health queries, diagnostics, report interpretation, and chronic disease management. It connects to over 5,000 hospitals, nearly one million doctors, and 269 specialist AI agents. It also integrates with major wearable and health management devices, offering users personalized, AI-powered healthcare services at scale.Ant Group CEO Han Xinyi said on July 27 that the company’s first two decades were defined by democratizing financial access and embedding QR code payments into daily life. Now, Ant is pivoting to an “AI-first” strategy anchored in three pillars: the Alipay dual flywheel, AI-centric transformation, and accelerated globalization. Since announcing the initiative in 2023, Ant has focused on developing AI-powered health, finance, and lifestyle assistants designed for mass adoption. The company is also pushing the frontier in AGI and embodied intelligence while investing in foundational infrastructure—from AI-native payment systems to security frameworks—to drive AI’s industrial-scale application.JD.com is making a full-scale entry into embodied intelligence. On July 27, the e-commerce giant rebranded its Yanxi large model lineup as JoyAI, now spanning model sizes from 3B to 750B to meet varied enterprise demands. The models support language, speech, image, video, and digital human modalities. Leveraging dynamic layered distillation and cross-domain data governance, JoyAI improves inference efficiency by 30% and slashes training costs by 70%, enabling lightweight, fast-deployable AI solutions. The newly released 750B model reportedly matches the performance of leading global LLMs.Tencent showcased its latest AI breakthroughs, including the Hunyuan 3D World Model and its embodied intelligence platform TairoS. “AI is not for showing off, but for solving real-world problems,” said Cai Guangzhong, Vice President of Tencent.Other tech leaders also rolled out their AI advancements:Baidu exhibited its AI digital human technology and new solutions from Baidu Intelligent Cloud.SenseTime launched WuNeng, its embodied intelligence platform built on the SenseTime Embodied World Modeland supported by its large-scale edge and cloud infrastructure. The platform equips robots and smart devices with visual navigation, multimodal interaction, and world perception capabilities, while also enabling integration with edge chips for flexible deployment.Kuaishou’s Koling AI introduced Ling Drawing Canvas, an AI creation tool featuring infinite canvas space, intelligent co-creation, and real-time multi-user collaboration.AI video generation startup Aishi Technology made its WAIC debut with Paiwo AI and the Paiwo AI Open Platform, marking a new push into domestic AI content production.Midu unveiled ModelTong 2.0, an AI document assistant that streamlines drafting, editing, proofreading, and formatting for official documents.iFLYTEK brought over 20 AI exhibits to the 2025 World Artificial Intelligence Conference (WAIC), underscoring its ambition to make AI ubiquitous and practical. As WAIC’s exclusive translation partner, iFLYTEK provided simultaneous interpretation across multiple academic forums and technical exchanges—including the highly anticipated first China appearance of Geoffrey Hinton, the 77-year-old “Godfather of AI” and winner of both the Turing Award and 2024 Nobel Prize.The company’s upgraded Speech Simultaneous Interpretation Model has achieved a breakthrough in latency, reducing the first-word response time for Chinese-English translation to just two seconds—on par with top-tier human interpreters—while maintaining high accuracy.iFLYTEK’s simultaneous interpretation services now span more than 50 countries and regions, serving over 400 million users and supporting 400,000+ conferences. Its latest Spark X1 large model supports over 130 languages and handles tasks including Q&A, math problem-solving, translation, and text generation, with marked improvements in stability and industry applicability.“The fourth wave of AI has entered deep waters,” said Zhao Xiang, Vice President of iFLYTEK. “The real test now is how large models integrate into real-world products and generate tangible user value.” He added that 2025 is shaping up to be a breakout year for “AI for All” and “AI for Every Industry.” iFLYTEK’s smart devices—including translators, voice recorders, and office tablets—are already facilitating cross-border communication, trade, and conferences for both Chinese and global users.Siemens made its WAIC debut this year, standing out as one of the few multinational exhibitors. The company introduced its Industrial Copilot intelligent agent system to the China market, showcasing its push to embed generative AI into industrial workflows. Siemens highlighted efforts around “industrial foundational models” and “industrial AI agents” designed to power next-gen manufacturing intelligence.Robots Take the SpotlightThis year’s WAIC saw robots outshine even the much-hyped “Six Little Tigers” of Chinese LLMs, as embodied AI moved to center stage. AI firms such as Zhipu AI and Stepverse highlighted robotic technologies at their booths, signaling the industry’s shift toward physical intelligence.Stepverse officially launched Step 3, its next-generation foundational large model, to be open-sourced globally on July 31. Described as the company’s first full-scale native multimodal reasoning model, Step 3 balances performance and inference efficiency via a Mixture of Experts (MoE) architecture with 321 billion total parameters and 38 billion active ones. The release marks a strategic leap in scaling through architectural and algorithmic innovation.In a lighter moment at the booth, Stepverse teamed up with entertainment giant Canxing Culture, the producer of “Sing! China”, to debut “AI Sing! China.” Visitors could scan a QR code to generate AI-powered rap songs, dance alongside robots, and even receive real-time performance scores—blending entertainment with cutting-edge AI interaction.Animated image sourced from the internetBeyond the singing and dancing robots, this year’s World Artificial Intelligence Conference (WAIC) saw a surge in embodied intelligence demos, with robots performing increasingly complex real-world tasks—including boxing, folding laundry, cooking, and object handling.A wide array of companies showcased their robotics technologies, including Unitree Robotics, Zhiyuan Robotics, Fourier Intelligence, Galaxy Universal Robotics, YunDeep, COLMO, Stardust Era, Beijing Humanoid Robot Innovation Center, Kepler, Keenon Robotics, Lingbao, Qianxun Intelligence, Leju, Zhipingfang, Uric Unix AI, Qiongche Intelligence, and Variable Robotics.Among the most talked-about exhibits was Unitree Robotics’ G1 combat robot, which stole the spotlight by transforming its booth into a full-fledged boxing arena at the Shanghai World Expo Exhibition & Convention Center. The humanoid robots captivated crowds with agile dodges, swift punches, and impressive coordination in live matches. Parents and children jostled for views, with many struggling to even squeeze a phone camera into the packed ring.During WAIC, Genie Robotics released the industry's first truly real-world-oriented dual-arm robot world model open-source platform, Genie Envisioner (hereinafter referred to as GE). This platform integrates three core capabilities—prediction, control, and evaluation—providing an end-to-end, unified solution for robots to transition from "seeing" to "acting." It enables robots to rehearse in the digital space before being deployed in the real world.At the same time, Genie also launched the industry's first general-purpose embodied foundation model—QiYuan Large Model, featuring "one brain, multiple forms" capability, allowing it to adapt to various heterogeneous robotic bodies. Additionally, on the application side, Genie Robotics has redefined the intelligent productivity paradigm through its "robot + embodied model" technology, achieving groundbreaking deployments in four major scenarios: industrial manufacturing, warehouse logistics, power inspection, and interactive guidance. Earlier, Genie established its own professional data collection facility, creating and open-sourcing the world's largest dataset, AgiBot World.Yao Maoqing, Partner & President of the Embodied Intelligence Division at Genie Robotics, stated that Genie has integrated the robot body, motion intelligence, interaction intelligence, and operational intelligence, achieving a full-stack, closed-loop development of "one entity, three intelligences." The company has also built a "data-model-body-scenario" four-wheel-driven embodied intelligence flywheel system. Every rotation of this "data-model-body-scenario" flywheel further empowers a wide range of industries with embodied intelligence.In addition, Genie will release a new generation of dexterous hands on August 18, offering up to 19 degrees of freedom at a weight of just 750g, while the 16-degree-of-freedom version weighs only 500g.GalBot showcased its Galbot robot at WAIC, playing the role of a “sales assistant” in supermarket operations. Behind this is GalBot’s independently developed end-to-end embodied large model. Currently, GalBot has secured orders from 100 stores and plans to roll out nationwide by the end of the year. In Beijing alone, nearly ten stores have already completed deployment and achieved regular operations. Each store, covering about 40 square meters, is stocked with over 5,000 types of medicines and 6,000 product slots, with tens of thousands of boxes of medicine managed solely by the Galbot “robot,” which is on duty around the clock.Fourier Intelligence also presented its upcoming third-generation GRx series product, the GR-3, at WAIC, marking its first public appearance ahead of its official launch on August 6. The GR-3 is Fourier’s first medical rehabilitation and wellness robot focused on interactive companionship. Building on previous iterations, it further expands the robot’s capabilities in companionship and emotional interaction. The GR-3 features an innovative soft skin-like covering and a fully immersive interaction system, pushing the boundaries of user experience in companionship and emotional engagement.Mech-Mind Robotics also demonstrated a variety of applications for its products at WAIC, including object picking, sorting, placement, and folding clothes. By enhancing the robot’s vision, cognition, and manipulation capabilities, the company aims to help people with everyday tasks such as serving coffee, passing plates, and folding clothes.Keenon Robotics unveiled the world’s first bipedal embodied service robot, XMAN-F1, in a global debut at WAIC. XMAN-F1 can not only make popcorn smoothly but also prepare a variety of drinks according to customer preferences, offering options such as “with or without ice.” Additionally, it can intelligently collaborate with other specialized service robots from Keenon, enabling multiple robots to complete assigned tasks as a team.DEEP Robotics brought real-world embodied intelligence operation scenarios to the exhibition hall, showcasing for the first time the industry-grade quadruped robot’s full-process inspection capabilities on site. They also presented the latest iteration of the Jueying series robots along with multiple industry application solutions.ROBOTERA made a brilliant appearance with its core embodied intelligence product lineup. Through live demonstrations of the Xingdong L7, Xingdong XHAND1, and Xingdong Q5, the company comprehensively showcased its technical strength in “full-stack, self-developed hardware-software integration” for embodied intelligence, as well as its commercial achievements “from technological breakthroughs to real-world applications.”In terms of dexterous hands, Beijing Lingxin Qiaoshou demonstrated piano playing during WAIC, with its fingertips striking the keys precisely, movements coordinated and rhythmic, achieving a level of fluency and sensitivity close to that of a human hand. Reportedly, after upgrading to about 12 active degrees of freedom, the product’s lateral finger flexibility and enveloping grip capability have been significantly enhanced, allowing it to adapt to more complex object shapes, with a price under one thousand yuan. The current flagship model has further advanced to 16 active degrees of freedom, with joint control precision and movement smoothness now even closer to human hand logic, further expanding its operational scenarios.Yinshi Robotics also unveiled its new RH56F1 dexterous hand at the event; Daimeng Robotics presented its next-generation DM-Hand1 dexterous hand; AI smart appliance brand COLMO showcased embodied intelligent robots’ walking and waving postures at its booth, and has taken the lead in applying AI Agent technology in the industry. Bilibili also featured robots playing soccer, performing the “Ip Man squat,” and running cross-country in the robotics area, allowing visitors to experience the real-time remote control of robots on site.Xi Wei, head of Midea’s Humanoid Robot Innovation Center, stated that humanoid robot development is still in its early stages, but costs will rapidly decrease as the industrial chain matures. Starting from first principles, more integrated designs that reduce the number and weight of components can simultaneously lower costs and improve performance.In fact, as Yao mentioned, the embodied intelligence industry is still in its early stages and developing at a rapid pace. When large models endow robots with intelligence and soul, the gears of fate for the emergence of physical intelligence begin to turn. Facing the data desert of embodied intelligence, we chose to plant the first tree, hoping it will one day become a forest.With the accelerated development of AI large models, these diverse robots are gradually achieving efficient and precise operations by sharing advanced VLA embodied intelligence models. This signals that in the next 5 to 10 years, embodied intelligent robots will soon enter everyone’s work, life, and homes.Yao revealed that the next-generation robot in AgiBot Genie series—the Genie G2—is about to be released. Compared to the previous generation, the G2 has been comprehensively upgraded in terms of motion accuracy and scenario adaptability. It features native data collection capabilities and can handle a variety of scenarios, including industrial, commercial, and household environments, further expanding the application boundaries of embodied intelligence.更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App