WAIC 2026 Preview: AI Agents Move from Chat to Action — The Era of Real Utility Begins

From July 17 to 20, the ninth World Artificial Intelligence Conference (WAIC) returns to Shanghai across three venues under the theme "Intelligent Partners, Co-creating the Future." This year's exhibition spans over 100,000 square meters for the first time, hosting 1,100-plus enterprises, more than 3,000 exhibits, and over 300 world-premiere products — all historic highs.

If the past few editions of WAIC revolved around "model emergence" and "capability leaps," 2026 delivers a sharper message: the competitive logic of the AI industry has decisively shifted from parameter races to production-grade agent deployment.

From Chat to Action: Agents Take the Baton

2026 is widely recognized as the "Year of AI Agent Production Deployment" — and the numbers back it up.

IDC projects a 139% CAGR for the global agent market from 2025 to 2030. Gartner estimates that 40% of enterprise applications will integrate agent capabilities this year. IBM and Salesforce jointly forecast over one billion AI agents running globally by the end of 2026. According to Boao AI, 79% of global organizations have already launched AI agent rollouts in 2026, with the market leaping from $7.63 billion in 2025 to $10.91 billion.

The industry's focus has shifted from "what models can do" to "what agents can accomplish."

The Compute Foundation: Huawei Atlas 950 Debuts

One of the most anticipated hardware reveals at WAIC 2026 is the first complete physical showcase of Huawei's Atlas 950 SuperPoD. Built on the domestic Ascend 950DT chip, this massive compute cluster scales up to 8,192 NPU cards in full interconnect, delivering 8 EFLOPS of FP8 compute with 1,152 TB of shared HBM and 16.3 PB/s of interconnect bandwidth.

Compared to NVIDIA's NVL144, the Atlas 950 offers 56.8x the card count, 6.7x the total compute, 15x the memory capacity, and 62x the interconnect bandwidth. Its self-developed Lingqu (UnifiedBus) all-optical interconnect protocol enables fully symmetric communication among CPU, NPU, and memory units — 160 physical racks functioning as a single logical compute entity. Huawei's rotating chairman Eric Xu noted that the SuperCluster built on Atlas 950 has already surpassed 500,000 cards in scale.

Agent OS and the AI Phone: Agents Go System-Level

Sharing the exhibition floor with these compute behemoths is something far more pocketable: the world's first AI Agent phone. Its core innovation isn't a "smarter assistant" — it's system-level agent capability, using a GUI Agent architecture to visually interpret and operate phone interfaces, autonomously executing multi-step tasks across apps. Nubia, in partnership with ByteDance's Doubao, has already prepared 80,000 to 100,000 units, with an initial batch totaling 500,000.

Stepfun (Jieyue Xingchen) will debut its Agent operating system at WAIC. Built on the open-source Step 3.7 Flash — a sparse MoE model with 196B parameters activating only 11B per inference — it already covers roughly 60% of top Chinese phone brands with over 42 million installs. Stepfun's three-year trajectory — from finance-specific LLMs to automotive Agent OS previews to a universal Agent operating system — mirrors the industry's broader consolidation from isolated breakthroughs to integrated systems.

The Token Economy: Turning Spend into Value

Agent deployment at scale is driving exponential growth in token consumption. By March 2026, China's daily token calls exceeded 140 trillion, a more than 1,400x increase from 100 billion in early 2024. An enterprise agent completing a single complex task can consume hundreds of thousands to millions of tokens — orders of magnitude beyond a simple chatbot exchange.

According to the "2026 Agent Top Ten Keywords" released by CAICT's AI Research Institute, the stable duration for agents executing long-horizon tasks doubles every eight months, with the current record standing at 16 hours. In Q2 2026, a cutting-edge model was released every 2.8 days on average.

Tokens are evolving from a technical metric into the core unit of production in the AI era. The battleground is shifting from parameter counts to token ecosystems, and the center of gravity in compute demand is moving from training to inference.

An Industry Logic Rewired

When 300 AI products make their global debut simultaneously in Shanghai — when agents move from concept into operating systems and phones, when humanoid robots walk out of labs and onto factory floors — the signal is unmistakable: the AI industry is transitioning from "demonstrating capability" to "delivering value."

From Huawei's Atlas 950 compute foundation to Stepfun's system-level Agent OS to the first AI Agent phone, China's AI supply chain is evolving from isolated technical breakthroughs toward systematic industrial realization. WAIC 2026 is not just a technology showcase — it's a collective confirmation of where the industry is headed. Agents and embodied intelligence are defining the next decade of AI.