WAIC 2026: When AI Stopped Showing Off and Started Counting Beans

From July 17 to 20, 2026, Shanghai's World Expo Exhibition Center saw 300,000 visitors braving 39°C heat. The 2026 World Artificial Intelligence Conference (WAIC) left behind topics even hotter than the weather.

This was no ordinary tech expo. Over 1,100 companies exhibited, more than 300 products made their global debut, and the exhibition area exceeded 100,000 square meters for the first time. Behind these numbers lies an industry-wide pivot.

If you've attended WAIC for three consecutive years, the contrast is stark. In 2024, booths touted "parameter counts," "context lengths," and "benchmark rankings." In 2025, the theme shifted to "model deployment." This year, parameter comparison boards vanished—replaced by finished products and mass-production figures that actually deliver in real scenarios.

The three-year LLM parameter arms race is officially over.

Agents: From Chat Windows to Production Lines

One word dominated WAIC 2026: Agent.

Over 10% of forum topics were agent-related, and four of the ten "Treasures of the Hall" were agent products. More telling was the shift in the conversation—not "can we build it," but "can we use it."

Stepfun turned nearly its entire booth into an AI phone experience zone. Its STEPX Neo—a native AI terminal powered by large models—was named one of the ten Treasures. Staff demonstrated the Amoo agent autonomously planning a business trip: booking flights and hotels on Ctrip, completing payment via Alipay, and connecting to Didi and Meituan for follow-up services—all with almost no app switching required.

Tencent transformed its booth into a "Digital Employee Experience Center," with agents spanning design, creation, office, and lifestyle forming a complete Agent matrix. Siemens unveiled an industrial engineering agent capable of independently handling PLC programming, equipment debugging, and production line design—boosting engineering efficiency by 2-5x.

But beneath the polished demos, the industry wrestled with a deeper question: how do agents move from demo environments into production environments?

Laiye CTO Hu Yichuan put it bluntly: many companies merely plug AI into existing workflows—"humans work, AI assists." True AI transformation means restructuring human-centric processes into agent-centric ones. Decathlon VP Xiao Lu was even more direct: AI won't automatically accelerate your company; it will first expose the flaws in your digital infrastructure—messy processes, inconsistent data standards, siloed departments. These debts must eventually be paid.

Token Factories: Compute Becomes an Industrial Commodity

If Agents were WAIC's public face, the Token economy was its backbone.

"Token used to be just a unit of measurement inside large models. This year, it became the hottest business term in AI." NVIDIA CEO Jensen Huang popularized the concept in June, likening modern AI data centers to "factories producing tokens"—every token convertible into code, answers, designs, and ultimately, profit.

At WAIC, "Token Factories" were everywhere. iSoftStone revealed its Beijing No.1 Token Factory's daily capacity: 1.4 trillion tokens. Infinigence AI's Agentic MaaS platform saw daily token calls surge roughly 40x since the start of the year, with inference costs dropping 10x. xFusion proposed a dual-center model: "Token as the cost center, Agent as the profit center."

Even more notable was the business model evolution. PPIO CEO Yao Xin observed "pulse-like" Agent calling patterns—sandboxes idling 90% of the time, spiking every half hour—forcing billing granularity down to the second. His prediction: "In three years, our customers might not even be human anymore." Over 90% of calls would come from agents.

Metering and provisioning compute like water and electricity—Token Factories represent the upgrade of AI capability from "artisan workshop" to "industrial assembly line." The industry has lost patience with distant AI idealism. AI is pragmatically entering its ROI phase: prove you can make money.

World Models: AI Learns Physics Common Sense

More futuristic than Token Factories: World Models.

In simple terms, a world model lets AI grasp the fundamental laws of the physical world: a cup shatters when dropped, an object moves when pushed, picking something up requires a certain amount of force. With this "common sense," robots can genuinely function in the real world.

GigaVision's booth was mobbed. Its complete "World Generation Model" and "World Action Model" system made its full debut: GigaWorld generates interactive virtual training worlds for embodied intelligence, GigaBrain serves as a universal embodied brain deployable across robot bodies, and the Shiguang S1 robot autonomously completed tasks like meal preparation, table clearing, and tidying in a home setting.

Zhiyuan Robotics launched its embodied foundation model GO-2 and world model GE-2, enabling robots to learn and evolve autonomously within a "model world" for the first time. ACE Robotics released Kairos World Model 3.1, extending the "understanding—generation—prediction" unified architecture. Across WAIC, 208 robot models no longer just danced—they transported goods, assembled components, folded clothes, and guided visitors.

World models serve more than robotics. GigaVision founder Huang Guan's thesis set the tone: "AGI has manifested linguistic instinct. Next will be physical instinct." As large models move from the digital to the physical world, AI must do more than talk—it must understand space, objects, motion, contact, failure, and feedback.

Computing Foundations: China's SuperPods Rise

Underpinning all of this: computing power.

WAIC 2026 debuted a 10,000-square-meter dedicated Chip & Computing Fusion Hall. Huawei's Atlas 950 SuperPoD made its first public appearance—1,024 NPU cards interconnected at high speed, delivering 1 EFLOPS FP8 compute with 256TB unified memory. Sugon unveiled the "Sugon 8000," China's first fully domestic 100,000-card AI supercluster. T-Head's Zhenwu M890 chip paired with the Panjiu AL128 SuperPod server was named a Treasure of the Hall.

Huatai Securities dubbed 2026 the "Year of Domestic SuperPods," projecting a market reaching 341.4 billion yuan by 2028, at a 194% CAGR. The logic of AI industry competition is shifting from individual chip performance to the organizational efficiency of entire computing systems.

Closing Thoughts

From the Hundred Model War of 2023, to the parameter arms race of 2024, to the deployment exploration of 2025, to the ROI era of 2026—WAIC's four-year arc is a microcosm of China's AI industry maturing from euphoria to pragmatism.

This year's WAIC taught the industry one thing: AI's value lies not in how smart the model is, but in whether it can truly enter production environments, execute reliably, and deliver measurable business results.

When parameters disappear from exhibition boards, when robots stop dancing, when Token becomes a line item on the balance sheet—that's when AI truly comes of age.