The four-day 2026 World Artificial Intelligence Conference (WAIC) concluded at the Shanghai Expo Center on July 20. The closing ceremony revealed impressive results: 100,000 square meters of exhibition space, 4,486 exhibits, 351 products making their global debut, over 400,000 in-person attendees, and more than 3 billion global online views. Representatives from 102 countries and international organizations participated, with 1,117 companies exhibiting. The total intended procurement amount reached 20.36 billion RMB, marking a 25% year-over-year increase.
Even more significant was the concentrated signing of key Shanghai AI projects at the closing ceremony—32 projects spanning AI infrastructure, embodied intelligence, scientific intelligence, and agent applications, with investments exceeding 40.9 billion RMB. This substantial capital injection sets a strong tone for the launch of China’s AI industry in 2026.
However, once the excitement fades and capital settles, a fundamental question arises: To what extent do the 40.9 billion RMB in signed contracts and 3 billion online views actually bridge the gap between AI being "able to chat" and "able to act"? This article aims to examine this issue from six perspectives.
From "Parameter Competition" to "Deployment Competition": Has the Industry Evaluation System Truly Shifted?
The most frequently cited narrative at this year’s WAIC was that "AI has moved past parameter competition and entered a new phase where real-world value is the benchmark." This was evident at the exhibition. In the embodied intelligence sector, more than 200 companies participated, with nearly 60 humanoid robots deployed throughout the venue for guiding, explaining, and answering questions. This marks the first time a major domestic exhibition has put humanoid robots to work—not just as photo ops, but actually performing tasks.
For example, Zhiyuan’s Elf G2 Max operates 24/7 in JD Logistics’ real warehouses, handling stacking and transportation. On the motor assembly line jointly exhibited by JAKA Robotics and Huayu Automotive, two robots work together to complete precision fastening in just 10 seconds; this solution is already running on a real production line with an annual output of one million motors. SenseTime’s robot convenience store "ShaomaiGo" has opened 10 locations in Shanghai and 20 nationwide, with each store averaging 400 orders per day.
Yet, the narrative of "moving from showmanship to practical work" masks a key reality: these real-world cases remain demonstration projects by leading companies, far from becoming scalable, standardized, and replicable industry norms. Of the more than 300 robots operating dynamically at the exhibition, most are still in the demo stage. "Deployment" was a high-frequency term at this year’s conference, but the act of deployment itself underscores that only pioneers—not the majority—are entering true production environments.
The Real Weight of 40.9 Billion RMB in Contracts: How Do Capital Density and Conversion Efficiency Align?
The 32 projects, totaling 40.9 billion RMB in investment, span four main areas: AI infrastructure, embodied intelligence, scientific intelligence, and agent applications. In terms of capital density, this was one of the largest concentrated signings in WAIC history.
However, the efficiency of capital investment hinges on two key variables: the technology maturity curve and the length of the commercialization path.
Take embodied intelligence as an example. This year’s WAIC saw the number of embodied intelligence exhibitors jump from over 80 last year to more than 200, with over 300 real machines on display. Industry analysts note that compared to previous years, WAIC 2026 marked a pivotal shift: the industry no longer focuses on the extreme capabilities of individual machines, but rather on data foundations, brain-like cognitive architecture, scenario adaptation, and the ability to commercialize through mass production. This upgrade in positioning is correct, but it also signals a reality: core capabilities are not yet mature. The architecture for robot "brains" remains unsettled, and sourcing training data is still a major challenge. Debates such as "should robots use VLA or world models" were front and center at the forums.
This means a significant portion of the 40.9 billion RMB will be invested in technical routes that have yet to achieve industry consensus. There is inherent tension between high-density capital inflows and uncertainty in technical direction. Contract value measures confidence, but it does not equate to progress toward real-world deployment.
Computing Infrastructure: From "Stacking Cards" to "Building Systems"—Has the Cost Inflection Point Arrived?
AI infrastructure is a major destination for the 40.9 billion RMB in contracts. At this year’s WAIC, the focus in the computing power exhibition area shifted away from chip benchmark scores to the unveiling of domestically produced, ultra-large-scale clusters. Industry consensus is emerging: AI infrastructure is moving from "stacking cards" to "building systems"—AI chips, accelerators, servers, rack systems, supernodes, high-speed interconnects, liquid cooling, storage systems, and scheduling software must be considered as an integrated whole.
Tianshu Zhixin released the Tiangai 300 general-purpose GPU, which achieved over 70% computational efficiency in DeepSeek V4 MoE scenarios and outperformed international mainstream solutions by 10% in 64k long-context attention tasks. The competition in computing power has shifted from "stacking transistors" to "squeezing efficiency." Jiuzhang Yunji’s "AI Factory" uses dual engines—"training factory" and "token factory"—to push AI production from expert-driven R&D to standardized, replicable assembly lines.
These developments signal a positive trend: the cost threshold for large-scale AI deployment is gradually being lowered. But "gradually" is the key word. Improvements in computing efficiency are incremental, while model parameter scale is expanding exponentially. The gap between these two trends remains a fundamental bottleneck preventing AI from moving from "able to chat" to "able to act." Whether token costs continue to fall and computing power truly becomes "as accessible as water and electricity" depends on sustained investment and breakthroughs in infrastructure—precisely the long-term challenge that the 40.9 billion RMB in contracts must address.
Embodied Intelligence’s "Pragmatic Turn": From Showmanship to Real Work—What Obstacles Remain?
The most talked-about change at this year’s WAIC occurred in the embodied intelligence pavilion. Humanoid robots are no longer the only form—piloted mechs, centaur designs, and wheeled platforms have emerged. This diversity reflects a pragmatic, scenario-driven logic: industrial settings require stability and load capacity, not just bipedal "human-likeness."
Yet, this pragmatic turn also reveals another facet of the industry’s early stage. Songyan Power’s Xiaobumi, priced at 9,998 RMB, became the first mass-produced humanoid robot on the market under 10,000 RMB. Lower prices signal the onset of mass production, but also indicate the industry has not yet found a high enough value anchor to support greater product premiums.
The bigger challenge lies in the technical hurdles of "real work." Daimeng Robotics demonstrated fruit packing and pencil case organization at their booth—seemingly simple tasks, but actually among the hardest for robots. Squeeze too hard and the fruit breaks; too lightly and it slips. The mango’s smooth surface means it can easily fall; transparent plastic lids are nearly invisible to vision sensors, so tactile sensors are needed. Organizing a pencil case is even tougher: the case is soft, and each step—grabbing, placing pens, zipping—changes its shape. These "small problems" in real-world scenarios are exactly the "big hurdles" robots must overcome to transition from "performing" to "working."
Agents: From "Dialogue" to "Action"—Can DAU Metrics Replace Parameter Competition?
Agent is another high-frequency term at this year’s WAIC. Unlike previous years, agents are no longer limited to dialogue demonstrations on booths—they’re now integrated into real workflows, transaction chains, and production lines.
A notable shift is the introduction of the DAA (Daily Active Agents) metric. Compared to traditional metrics like model parameters and token consumption, DAA focuses on how many agents actually enter business processes, complete tasks, and create value each day. According to IDC’s "DAA Research Report," global daily active agents numbered 28.6 million in 2025, are expected to reach 79.4 million in 2026, and grow to 2.216 billion by 2030.
Switching metrics reflects an upgrade in industry understanding, but warrants careful consideration: What is the statistical definition of DAA? What qualifies as a "daily active agent"—a single task dispatch or a full end-to-end business cycle? Differences in definitions across vendors may lead to data that lacks comparability. Just as "parameters" became marketing buzzwords in the era of large models, "daily active agents" face the risk of misuse. The real test for agents is not the inflation of DAU figures, but whether they can deliver stable, reliable, and auditable task execution in complex scenarios.
Scientific Intelligence and Global Governance: Institutional Breakthroughs on the Long-Term Track
Among the four directions for the 40.9 billion RMB in contracts, scientific intelligence (AI for Science, AI4S) and global governance received relatively "quiet" discussion, but their long-term significance may be greater.
Scientific intelligence marks AI’s shift beyond internet content and office applications, deeply penetrating fields like biomedicine and new materials. AI is now actively involved in scientific hypothesis, simulation, and experimental validation, reshaping traditional research models. During WAIC, the "Quantum-Super Intelligence Fusion Platform" was officially launched, representing a breakthrough in domestic scientific intelligence computing infrastructure. Crystal Tai Holdings showcased AI4S innovations accelerating R&D and commercialization in new drugs and materials.
On the global governance front, 29 countries signed an agreement to establish the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai. This is the first globally negotiated governance framework for AI. The conference brought together 11 recipients of the Turing Award, Nobel Prize, and Fields Medal. These institutional breakthroughs may not yield immediate commercial returns, but they lay the governance foundation for AI’s long-term healthy development.
If embodied intelligence and agents answer the question "What can AI do?", scientific intelligence answers "What can AI discover?", and global governance answers "How should AI be managed?" The timelines for these questions differ—the first may become clear within 3–5 years, while the latter two could take 10 years or more. Although the proportion of investment in scientific intelligence and governance frameworks is limited in the 40.9 billion RMB contracts, their strategic value is significant.
Conclusion
WAIC 2026 delivered an impressive report card: 32 projects, 40.9 billion RMB in contracts, and 3 billion global online views. AI infrastructure is evolving from "stacking cards" to "building systems," embodied intelligence is shifting from "showmanship" to "real work," and agents are moving from "dialogue" to "action"—these are real and commendable trends.
Yet, a "cold analysis" reminds us: 40.9 billion RMB is a vote of confidence, not a guarantee of implementation. The journey from "able to chat" to "able to act" must be measured by ongoing technological breakthroughs, engineering capabilities, data accumulation, and business model iteration. Whether the cost inflection point for computing power has truly arrived, when robot brain architectures will converge, whether agent task completion rates can keep improving, and if scientific intelligence will move from concept to substantive discovery—these answers won’t be found at WAIC’s exhibition booths, but in the industry’s practice over the next 3–5 years.
The true value of WAIC 2026 may not lie in how much has already been accomplished, but in how clearly it reveals what must be achieved next.
FAQ
Q: What fields do the 32 AI projects at WAIC 2026 cover?
They span four major areas: AI infrastructure, embodied intelligence, scientific intelligence, and agent applications, with a total investment exceeding 40.9 billion RMB.
Q: What were WAIC 2026’s global online traffic and exhibition scale?
Global online traffic surpassed 3 billion views, with over 400,000 in-person attendees, 1,117 companies exhibiting, 4,486 exhibits showcased, and 351 products making their global debut.
Q: What were the key changes for embodied intelligence at this WAIC?
Embodied intelligence had its own dedicated pavilion for the first time. Exhibiting companies grew from over 80 to more than 200, with over 300 real machines, nearly all operating dynamically. The industry shifted from "showmanship" to "real work," with robots entering real production lines and commercial settings.
Q: What trends did agent technology show at WAIC 2026?
Agents moved from "dialogue" to "action," no longer limited to booth demonstrations but entering real workflows and production environments. Global daily active agents are expected to reach 79.4 million in 2026.
Q: What are the main obstacles for AI transitioning from "able to chat" to "able to act"?
Key obstacles include: the gap between computing power costs and model scale, unresolved technical routes for robot brains, difficulties in acquiring training data, and the need to improve task completion rates and stability in complex scenarios.
Q: What were WAIC 2026’s achievements in AI global governance?
Twenty-nine countries signed an agreement to establish the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai—the first globally negotiated governance framework for AI.
