WAIC 2026: AI Finally Stopped Chatting and Started Working

Shanghai World Expo Exhibition Center. 40 degrees outside. Even hotter inside—packed with human bodies moving at a crawl.

WAIC 2026 broke records: over 100,000 square meters of exhibition space, 1,100+ exhibitors, 3,000+ exhibits. Day one was jammed; day two was worse.

But the biggest change wasn't more models. It was that AI finally figured something out: stop chatting, start working.

Agents: From Chat to Chores

The dominant feeling walking through Hall H1: every major AI company is building agents now.

StepFun turned its booth into an AI phone showroom. Their STEPX Neo, one of the exhibition's "Top Ten Treasures," features an agent called Amoo that can plan business trips, book flights on Ctrip, and complete payments via Alipay—all without you switching between apps.

Tencent's booth was even bolder: a "Digital Employee Experience Center." Staff placed a handwritten task card on a device, and WorkBuddy autonomously parsed the request, broke it into steps, called the right tools, and delivered results. According to Tencent, over 95% of their engineers now use CodeBuddy, cutting coding time by about 40%. Under the hood? The newly released Hunyuan Hy3 model—295B total parameters, 21B active, 256K context window.

Ant Group showed off AI-powered payments. Abao (the AI version of Alipay) handles train bookings, ride-hailing, and package delivery with a single voice command—all within Ant's own ecosystem. What surprised me more was "Afu" in healthcare: after pairing with mainstream fitness bands and blood pressure monitors, it delivers personalized AI health advice, online consultations, and automated hospital registration.

Alibaba wasn't idle either. They launched Meoo Team, an enterprise AI application platform, plus their first AI earbuds—clip-on design with Qwen AI built in, supporting real-time translation, meeting transcription, and health tracking.

Models kept evolving too. On the eve of the show, Moonshot AI dropped Kimi K3—the first open-source model reaching 2.8 trillion parameters, ranking just behind Claude Fable 5 and GPT-5.6 Sol in overall capability. MiniMax showcased M3 and teased the multimodal H3. Baichuan unveiled the M4 medical model, already deployed as AI oncologists and pediatricians in partner hospitals.

Robots: Just Get the Job Done

Last year's robots flipped, danced, and boxed. This year? They clocked into factories.

Unitree still drew crowds, but beyond the boxing and dance routines, the real head-turner was GD01—the world's first rideable transforming mech, standing 2.8 meters tall, switching between bipedal and quadrupedal modes. Price tag: 3.9 million yuan.

But the robots that mattered were the ones sweating on assembly lines. Leju built a real factory production line at its booth. Humanoid robots autonomously unboxed materials, loaded parts, and straightened crooked boxes. The entire hardware-software stack is domestically developed, already deployed across automotive, logistics, and a dozen other industries.

Zhiyuan launched five new products in one go, including the Expedition A3 Ultra and the world's first cycling robot, demonstrating industrial automation for loading and unloading. Lingchu Intelligence partnered with fiber optics giant YOFC on precision optical module manufacturing—inspection, sorting, and packaging, all automated to minimize component damage.

Commercial and home scenarios are accelerating too. Ant Lingbo recreated a real pharmacy at its booth: robots identify drugs, sort prescriptions, and deliver them, powered by LingBot-VLA 2.0, a unified embodied intelligence brain already compatible with 17 robot manufacturers and 20+ configurations. Star Dynamics brought the industry's first commercially validated logistics robot solution, already running routine operations at SF Express and China Post locations.

Fourier set up a living room in its booth, debuting an "Embodied Home" solution for companionship. Galaxy Universal had robots toasting bread, pouring drinks, and plating food in real time. No more backflips—they're going into domestic service.

One new direction worth watching: brain-computer interfaces merging with robotics. BrainCo demoed the Revo 3 bionic dexterous hand, capable of executing brain-wave commands. OYMotion's EMG wristband offers embodied AI a high-quality way to capture physical interaction data.

Computing Power: The Super-Node Era

This year, domestic chipmakers got their own entire exhibition hall. The focus wasn't single-card benchmarks anymore, but one shared concept: super-nodes.

Huawei's Ascend 950 super-node stole the spotlight: 1,024 cards in a single unit, delivering 1 EFLOPS (FP8) with 3-microsecond latency, supporting trillion-parameter model training and inference. The Atlas 950 SuperCluster built on it can scale to 500,000 cards. Expected to ship in Q4 2026.

Sugon unveiled China's first fully domestic 100,000-card AI super-cluster, the Sugon 8000, powered by Hygon chips with immersion phase-change liquid cooling. Baidu Kunlun's Tianchi super-node improved throughput by 25% with 256 cards; its 512-card variant already handles trillion-parameter training. Alibaba's Zhenwu M890x Panjiu AL128 super-node, another "Top Ten Treasure," uses T-Head's in-house training-inference chip and has surpassed 400 customers.

Startups joined the fray. MetaX, Enflame, and Biren all debuted their first super-node offerings. Tsinghua professor Wei Shaojun's EastStar Computing showed the DF1000, a near-memory computing 3D chip that stacks compute and memory to shrink data transfer distance, aiming to break through the memory wall, bandwidth wall, and power wall.

Across the hall, companies rallied under the banner "Built for the Token Factory." SenseTime launched the "Galaxy Initiative" for 10,000-card intelligent computing clusters. Wuwen Core unveiled a "front-shop-back-factory" three-layer computing architecture. PPIO claimed its AI agent cloud service could slash computing costs by over half.

One chip industry insider told me orders have exploded this year. Supply-chain constraints—production capacity, memory—mean the shortage won't ease meaningfully for another one to two years. Training demand is converging but per-user requirements are growing larger, while inference demand is growing exponentially.

Final Thoughts

Three years ago, the AI conversation was "will it replace me?" After WAIC 2026, the conversation has shifted to "how much work can it do for me?"

Agents are taking over daily services. Robots are entering factories and homes. Domestic computing power is sprinting toward 100,000-card clusters. AI is evolving from "good at talking" to "good at working," and WAIC 2026 might just be the milestone that marks this turn.

The road ahead is long—embodied intelligence still wrestles with data challenges, computing supply remains tight, and agent ecosystems remain fragmented. But the direction is right: get AI out of the chat box and into creating real value.