WAIC 2026 showcases domestic computing breakthroughs and edge-side AI agent advances after Shanghai summit closes

The 2026 World AI Conference & High-Level Meeting on Global AI Governance, themed “AI Partnership for a Brighter Future”, concluded on 20 July. The event continues to shape industry discussions across global artificial intelligence circles. Exhibition spaces hosted 4,486 exhibits, with 351 products making their global debuts. Visitor numbers exceeded 400,000, and preliminary intended procurement deals are valued at approximately 20.36 billion yuan, marking a year-on-year rise of 25 per cent.

Beneath the bustling display of innovations spanning super-node computing infrastructure and home-grown semiconductors, token economics and edge-deployed AI agents stand out as key trends. These developments reflect sustained efforts to consolidate independent intelligent computing foundations and advance the development of next-generation smart terminals.

Computing chips are evolving beyond isolated deployment models. Huawei unveiled the physical prototype of its Ascend 950 SuperPoD for the first time at the conference. The configuration arranges 20 cabinets, comprising eight computing cabinets and four Lingqu interconnection cabinets. A total of 1,024 neural processing units operate in high-speed interconnection, delivering 1 EFLOPS of FP8 computing power and 2 EFLOPS under FP4 precision. Super-nodes have emerged as a fresh paradigm for constructing AI infrastructure. Physically interconnecting vast arrays of chips while operating them logically as a unified computing system creates the capacity to satisfy the heavy computing demands of large language models.

A range of competing domestic super-node solutions were presented side by side. These include Muxi Semiconductor’s Xijing S600 new-generation AI super-node product, an all-domestic super-node built with 128 processing units per cabinet combining T-head’s Zhenwu M890 chips and Alibaba Cloud Panjiu AL128 servers, Sugon’s Dawn 8000 large-scale AI supercluster capable of supporting tens of thousands of accelerators, and ZTE’s OEX super-node platform.

Industry practices advance alongside accelerated formulation of unified standards. The White Paper on Super-node Definition and Practice, jointly drafted by the Pengcheng Laboratory and the Global Computing Alliance, was formally launched during WAIC 2026. The document establishes consistent industry benchmarks for super-node architecture after years of fragmented technical definitions across the sector. It identifies unified memory addressing across discrete physical nodes as the core architectural feature, alongside three defining technical attributes: memory semantics and unified memory addressing, ultra-low latency and ultra-high bandwidth.

Innovation activity continues to expand across the domestic semiconductor supply chain. Over one hundred firms focused on intelligent computing chips displayed more than 200 cutting-edge products at the conference, of which 67 received their first public presentations in China. Muxi Semiconductor rolled out its new Xisuo X series chips. Eastern Computing Core exhibited the DF1000, the world’s first AI chip built on mature domestic manufacturing processes integrating software-defined frameworks and 3D stacked near-memory computing. Enflame Technology partnered with First Packaging Technology to release CoPoS panel-level advanced packaging samples, exploring glass substrate applications for semiconductor packaging. Microcore Intelligent demonstrated its Reconfigurable 3.0 technology and flagship TX82 chip, pushing three-dimensional computing-in-memory chips from conceptual research toward practical deployment. Bluechip Computing delivered a full demonstration covering RISC-V server hardware, optimised software stacks and model benchmark testing, targeting bottlenecks within domestic computing infrastructure originating at the chip layer.

Independent computing capability cannot be reduced to simple one-to-one replacement of imported chips with domestic alternatives. Chinese developers are building competence to define chip specifications, construct computing systems and cultivate complete industrial ecosystems, forming robust foundations for self-reliant intelligent computing infrastructure.

While computing infrastructure strengthens at the backend, AI agents are accelerating deployment onto edge terminals. Progress in embedding agents within consumer hardware gathered clear momentum throughout WAIC 2026. Stepverse unveiled the STEPX Neo smartphone, presented as the world’s first handset built natively for large-model AI agents alongside its dedicated agent operating system. Alibaba Group worked with Honor to create handset-optimised solutions built around the Qwen large model, already deployed on the Honor Robot Phone. Nubian smartphones equipped with Doubao mobile assistant move into mass production.

Research divides the evolution of edge operating systems into three distinct phases. The initial stage adds AI functions onto conventional operating systems without fundamental architectural restructuring for models and agents. The second phase delivers AI platform operating systems optimised for large models and silicon hardware, yet these platforms are not designed with native AI support from inception. The third tier covers AI-native operating systems, engineered from the ground up to accommodate heterogeneous accelerators including NPUs and GPUs, enable lightweight models to run efficiently on end devices, and sustain continuous operation of intelligent agents.

Hardware manufacturers roll out successive product iterations, yet widespread consumer adoption remains unconfirmed. Operators developing mobile AI agents must prioritise identifying high-frequency, valuable real-world scenarios to close loops connecting model development, engineering implementation and user feedback, enabling mass user engagement with intelligent systems.

Embodied robotics represents another vital frontier for commercial AI rollout. Embodied intelligence sits within the broader field of edge intelligence, requiring models to integrate perception, reasoning, decision-making and physical action capabilities. Large language models have reached the highest maturity level, while multimodal perception systems stand at roughly 70 per cent technical readiness. Actuation modules lag, with maturity levels below 40 per cent. Embodied intelligence demonstrates commercial potential within automotive internal and external perception systems, yet viable applications for complex physical manipulation remain limited.

A structural shift reshapes computing demand dynamics. Token consumption has historically been driven by human users, yet the bulk of token generation and exchange will take place outside direct human observation, triggered by agent planning, tool invocation and automated code writing. Industry focus shifts from sheer computing scale to token value. Tokens evolve into core value carriers and pricing benchmarks within the AI sector. Clients weigh token generation efficiency and delivery quality rather than simply purchasing computing hardware.

China holds competitive advantages for lowering token production costs, supported by globally leading open-weight large models, scalable memory production capacity and stable, affordable power supplies.

Market participants across the industrial chain continue tracking iterative upgrades of domestic silicon products, edge agent deployment progress and shifts in token-driven computing demand. Ongoing technical innovation will keep reshaping competitive dynamics across the artificial intelligence ecosystem.