WAIC 2026 Deep Dive: From "Model Wars" to "Year One of Agents" — China's AI Industry Shifts Gears

Event Overview

From July 17 to 20, 2026, the World Artificial Intelligence Conference (WAIC 2026) was held across four venues at the Shanghai World Expo Exhibition Center, Zhangjiang Science Hall, and Xuhui West Bund. Exhibition space exceeded 100,000 square meters, with over 1,100 participating companies, 3,000+ exhibits, and 300+ global debuts — all core metrics doubling year-over-year. ✅Verified

The most significant signal from this year's conference: the AI industry has officially shifted from the "foundation model competition" phase to the "agent deployment" phase. Among 130 large model and generative AI exhibitors, 63.8% identified primarily as AI Agent/intelligent application providers, while only 18 focused on foundational models. Meanwhile, American AI giants were collectively absent — NVIDIA, Intel, and AMD had no independent booths; Apple, Microsoft, and Tesla sent only minimal staff with no press conferences or partnership signings.

Background

WAIC has been held for nine consecutive years, serving as China's most important annual AI showcase. At the 2025 conference, large models dominated center stage, with companies competing on parameter counts, context lengths, and multimodal capabilities. Just one year later, the industry landscape has fundamentally shifted:

• Models retreat backstage: Model capabilities are converging toward homogenization, and the competitive focus has shifted from "whose model is stronger" to "whose Agent can actually get work done"
• US-China AI decoupling accelerates: The December 2025 US "FY2026 National Defense Authorization Act" expanded export controls on AI chips to China, compounded by 25% tariffs, directly causing mass withdrawal of American companies from Chinese exhibitions
• Domestic compute substitution enters the system-level phase: Short-term catch-up in single-chip process nodes is unfeasible; supernode clusters have become the pragmatic path
• Embodied intelligence moves from labs to production lines: Humanoid robots are beginning mass production and commercial deployment

Core Analysis

I. Agents Become the Undisputed Protagonist

The product logic at this year's conference underwent a fundamental transformation. Previously, companies showcased "what my model can do"; now they demonstrate "what my Agent can complete for you end-to-end."

• Baidu built its entire showcase around DAA (Daily Active Agents). Its DuMate universal agent was upgraded to autonomously decompose tasks and execute cross-application workflows; the Wenxin 5.1 model retreated backstage as the Agent's "engine"
• Tencent's Hunyuan Hy3 topped OpenRouter's global API call volume chart within one week of launch, but more noteworthy was its full-stack embodied intelligence solution — from cloud infrastructure to VLA models to the TairosAgent framework, forming a complete "brain + cerebellum + body" closed loop
• Alibaba upgraded its Miaowu team edition from a personal tool to an organizational productivity platform, paired with Qwen AI glasses, extending the Agent entry point from screens to wearables
• StepFun launched Step AOS, the world's first terminal-native Agent operating system, adapted to approximately 60% of China's top smartphone brands with over 42 million installed units, and introduced China's first L3-certified Agent smartphone

Key Judgment: Agents are not just marketing repackaging of large models — they represent a qualitative shift in product form: from "user issues commands, model produces results" to "user sets goals, Agent executes autonomously." This means AI product interaction paradigms, security models, and business models will all be restructured. ⚠️High Confidence

II. Domestic Compute Enters the "Supernode Era"

Facing US chip export controls, China's compute industry has chosen a "system-level breakthrough" strategy:

• Huawei Atlas 950: 8,192-card supernode, FP8 compute of 1 EFLOPS, 256TB unified global memory addressing, Lingqu 2.0 all-optical interconnect with 3μs RTT latency, launching Q4 2026. Combined with 750+ commercially deployed 384-card supernodes, forming a dual product line of "volume + flagship"
• Alibaba Panjiu AL128: 128 Zhenwu M890 chips per cabinet, proprietary ALink interconnect + ICN Switch 1.0, Pb/s-class bandwidth, inference performance 50% improvement over traditional architectures. Paired with Zhenwu M890 chips (3x previous generation performance, 560,000+ units delivered cumulatively), forming a chip-server-cloud service full stack
• Sugon Dengfeng 8000: China's first 100,000-card fully self-developed AI supercluster, domestically produced across the entire chain from chips to applications
• ZTE OEX: Pioneering orthogonal backplane-free architecture, open and decoupled, compatible with diverse GPU ecosystems

Key Judgment: The competitive logic for domestic compute has shifted from "single-card performance benchmarking against NVIDIA" to "system-level engineering capability." Supernode architectures compensate for single-chip gaps through high-speed interconnects — the most pragmatic catch-up path currently. However, ecosystem compatibility (CUDA migration costs) remains the core bottleneck for deployment. ⚠️High Confidence

III. Embodied Intelligence Crosses the Mass Production Threshold

Robots are no longer just "performance acts" at exhibition booths — they are entering real production lines and service scenarios:

• Zhiyuan Robotics Expedition A3 Ultra — the world's first mass-producible full-size humanoid robot, with 15,000+ units rolled off the production line, flexible production lines delivering 100,000+ units annually, 7×24h unmanned operations
• Ant Group's Lingbo Smart Pharmacy — LingBot-VLA 2.0 drives three heterogeneous robots in collaboration, completing the full consultation-to-dispensing workflow in 90 seconds, already in live operation at Guoda Pharmacies in Shanghai
• Tencent deployed VLA models to cosmetics production lines with task success rates exceeding 95%
• Unitree Robotics GD-01 manned transforming mecha — 3 meters tall, 500 kg, dual humanoid/quadruped form modes, priced from 3.9 million yuan

Key Judgment: Embodied intelligence is replaying the large model trajectory — from technical showcase to scenario validation to mass production. 2026 is "Year One of robots reporting for duty," but large-scale commercialization still requires solving reliability, cost, and safety certification challenges. ⚠️Vendor Claims

IV. The Signal of American Giants' Collective Absence

NVIDIA's China market share fell from 95% to approximately 8%, H200 export volumes to China are "negligible," and Blackwell remains banned from direct export. Microsoft CEO Nadella only spoke at a parallel forum to criticize Claude Fable's excessive controls, with no product launches. OpenAI, Anthropic, and Meta did not exhibit.

Key Judgment: The US and China are accelerating toward two relatively independent AI technology and governance systems. In the short term, American companies lose Chinese market revenue; in the long term, this may accelerate the maturation of China's domestic substitution ecosystem — the impact on global AI industry dynamics runs far deeper than one conference's absences. ⚠️High Confidence

Vendor Responses

Predictions

• The second half of 2026 will become an Agent product elimination round: Of the current wave of Agent product launches, fewer than 10% will achieve meaningful scale in daily active usage; the rest will quickly fade. Evaluation criteria will shift from "demo quality" to "user retention."
• Domestic supernodes will face their first large-scale commercial validation in Q4 2026: The real-world performance of Huawei Atlas 950 and Alibaba Panjiu AL128 will determine whether domestic compute can genuinely handle large model training workloads.
• Embodied intelligence will enter "real ROI validation" in 2027: Current mass production data is largely capacity demonstration; the true test lies in long-term reliability and cost-effectiveness across factory and service scenarios.
• US-China AI governance systems will accelerate divergence: 29 nations signed the WAICO (World AI Cooperation Organization) with headquarters in Shanghai, forming a parallel structure to the US-EU-led AI governance framework. Companies will need to prepare for "dual-track compliance."
• On-device AI will become standard in smartphones within 12 months: Seven on-device large models have completed regulatory filing, Agent smartphones are launching en masse, and edge-cloud collaborative architecture will become the industry standard by 2027.

*Confidence notes: ✅Verified — sourced from official releases or authoritative media testing; ⚠️High Confidence — cross-validated across multiple reliable media reports; ⚠️Vendor Claims — sourced from vendor self-reporting, pending independent verification.*