WAIC 2026 Observation: Behind the $40.9 billion contract signing, how far is AI from being able to “talk” to being able to “do”?

The four-day 2026 World Artificial Intelligence Conference (WAIC) concluded on July 20 at the Shanghai Expo Center. The scorecard announced at the closing ceremony was impressive: 100k square meters of exhibition space, 4,486 exhibits, 351 products making their global debut, more than 400k offline audience visits, and over 3 billion in global online traffic. Participants included 102 countries and international organizations, with 1,117 companies exhibiting; the conference is expected to achieve intended procurement amounts of 20.36 billion yuan, up 25%.

More symbolically, a batch of key Shanghai AI projects were centrally signed at the closing ceremony—covering 32 projects across areas such as AI infrastructure, embodied intelligence, scientific intelligence, and agent application. The investment amount exceeds 40.9 billion yuan. Concrete, real-dollar commitments provide a strong footnote for the start of China’s 2026 AI industry.

However, once the noise fades and capital is secured, a more fundamental question needs to be asked: to what extent have the 40.9 billion yuan in signed deals and the 3 billion in online traffic truly shortened the distance for AI to go from “can chat” to “can do”? This article attempts to think coldly through six dimensions.

From “parameter races” to “implementation races”: has the industry evaluation system truly switched?

The most frequently cited narrative at this year’s WAIC is: “AI bids farewell to the parameter race and enters a new stage where landing value determines the winners.” This judgment was directly echoed on the exhibition hall floor. In embodied intelligence, more than 200 companies exhibited, and nearly 60 humanoid robots were scattered throughout the venue for guidance, explanations, and inquiries. This is the first time in a major domestic exhibition that humanoid robots have been deployed in large scale—up not just for people to take photos with, but actually to work.

Yuanshun’s Spirit G2 Max has been moving and stacking pallets in JD Logistics’ real warehouse 24 hours a day; on an electric motor assembly line co-exhibited by Segway and SAIC Motor/Continental Automotive (HuaYu Automotive), two robots coordinate to complete precise locking in only 10 seconds—this solution has already been running on real production lines for one million sets of electric motors per year. “Shaomai Go,” a robot convenience store incubated by SenseTime, has opened 10 locations in Shanghai, with 20 across the country; real stores average 400 orders per day.

But the narrative of “from showmanship to hands-on work” obscures a fact: these landing cases are still a minority of showcase projects by leading enterprises, still far from becoming the industry norm of large-scale, standardized, and replicable operations. With over 300 dynamically running robots in the hall, most remain in the demo stage. “Deployment mode” was a high-frequency term at this conference, but “deployment” itself indicates that those truly entering production environments are still the early movers, not the majority.

The real weight of 40.9 billion yuan in signed deals: how do capital intensity and conversion efficiency match?

With 32 projects and 40.9 billion yuan in investment, spanning four major areas—AI infrastructure, embodied intelligence, scientific intelligence, and agent applications—this is among the largest batches of centralized signings in WAIC history in terms of scale.

But the efficiency of capital investment depends on two key variables: the technology maturity curve and the length of commercialization paths.

Take embodied intelligence as an example. In this year’s WAIC, the number of embodied intelligence exhibitors increased from more than 80 last year to over 200, with more than 300 real machines. Industry analysts pointed out that, compared with previous years, WAIC 2026 accomplished a critical transition: the industry no longer competes on the extreme stunts of single machines, but instead competes on the underlying capabilities—data foundations and brain-like cognition—along with scene adaptability and commercialization abilities for batch production. This positioning upgrade is correct, but it also implies a reality: the underlying capabilities are not yet mature. The robot brain architecture is unsettled, and sources of training data remain an industry challenge. The debate over whether robots should use VLA or world models was brought directly onto the forum.

This means that a substantial portion of the 40.9 billion yuan will be allocated to technical routes that have not yet formed industry consensus. There is an inherent tension between the high-density inflow of capital and the uncertainty of technology paths. Signed-deal amounts measure confidence, but they do not equal a progress bar for implementation.

AI infrastructure: from “stacking cards” to “building systems”—has the cost turning point arrived?

AI infrastructure is one of the major destinations of the 40.9 billion yuan in signed deals. At this year’s WAIC, the core highlight of the computing power exhibition area was no longer the benchmark scores of a single chip, but the real-machine display of domestically made ultra-large clusters. Industry consensus is forming: AI infrastructure is shifting from “stacking cards” to “building systems”—AI chips, accelerator cards, servers, full-rack systems, supernodes, high-speed interconnects, liquid cooling and heat dissipation, storage systems, and scheduling software must all be considered as an integrated whole.

Tianji Zhixin released Tianrui 300 general-purpose GPUs. In the DeepSeek V4 MoE scenario, computational efficiency broke through 70%; in 64k long-context scenarios, attention efficiency is 10% higher than international mainstream solutions. Computing power competition has moved from “squeezing more transistors” to a stage of “squeezing more efficiency.” The “AI factory” displayed by 9章云极 uses two engines—training factories and token factories—to drive AI production, pushing it from a development mode that depends on a handful of expert teams toward a replicable, standardized production pipeline.

These advances point to a positive signal: the cost threshold for large-scale AI deployment is being gradually dismantled. But “gradually” is the key. Improvements in computing efficiency are incremental, while expansions in model parameter scale are exponential. The “scissor gap” between the two remains a fundamental bottleneck limiting AI from “can chat” to “can do.” Whether token costs can continue to decline and whether computing power can truly become a “universal utility resource like water and electricity” depends on sustained investment and technological breakthroughs at the infrastructure level—exactly the long-cycle proposition that needs to be answered by the 40.9 billion yuan in signed deals.

Embodied intelligence’s “pragmatic shift”: from showmanship to doing work—there are still several hurdles in between

The most talked-about change at this year’s WAIC occurred in the embodied intelligence exhibition hall. Humanoid robots are no longer the only form factor—manned exosuits, semi-humans (centaur-like robots), and wheeled chassis have also crowded in. Behind the diversification of forms lies a pragmatic logic driven by scenarios: industrial scenarios need stability and load-bearing capability, not “human-like” charm from biped walking.

But the “pragmatic shift” also reveals another side of the industry’s early stage. Xiao Bu Mi, launched by Songyan Power, is priced at 9,998 yuan, becoming the first mass-produced humanoid robot on the market under 10,000 yuan. Price cuts signal the prelude to large-scale mass production, but they also imply that the industry has not yet found a sufficiently high value anchor to support higher product premiums.

More challenging is the technical barrier of “doing work” itself. The packaged fruit and organizing pen cases demonstrated by Daimeng robots look simple, but in reality are among the hardest things for robots to tackle. If the clamping is too heavy, the item breaks; if it is too light, it slips. Mango surfaces are smooth; one slip and it drops. Transparent plastic lids have almost no effective visual cues, so only tactile sensors can be relied on for perception. Organizing pen cases is even harder: pen cases are soft—every step of grabbing, placing, and pulling the zipper involves deformation. These “small problems” in real scenarios are precisely the “big hurdles” that robots must cross to move from “can perform” to “can do the job.”

Agents from “conversation” to “execution”: can DAU replace the parameter race?

Agents are another high-frequency term at this year’s WAIC. Unlike previous years, agents here no longer stop at dialogue demos on the exhibition floor; instead, they are pushed into real work flows, transaction pathways, and production workshops.

A metric shift worth noting is the introduction of DAA (Daily Active Agents, daily active agent count). Compared with traditional metrics such as model parameters and token consumption, DAA focuses on how many agents actually enter business processes every day, complete tasks, and create value. A research report on DAA released by IDC shows that the number of global daily active agents in 2025 was 28.6 million, is expected to reach 79.4 million in 2026, and will grow to 100k by 2030.

The shift in metrics reflects an upgrade in industry understanding, but it also needs cold thinking: what are the statistical definitions and criteria for DAA? How is a “daily active agent” defined—does it mean completing a single task dispatch, or achieving an end-to-end business closed loop? Differences in definitions among vendors may lead to lack of data comparability. Just as “parameters” once became marketing phrasing in the era of large models, “daily active agents” also faces the risk of being misused. From “conversation” to “hands-on execution,” the real test of agents is not the inflation of daily-active numbers, but whether they can perform tasks stably, reliably, and with auditability in complex scenarios.

Scientific intelligence and global governance: institutional breakthroughs on a long-termism track

Among the four directions in the 40.9 billion yuan in signed deals, discussions on scientific intelligence (AI for Science, AI4S) and global governance were relatively “quiet,” but their long-term significance may be even deeper.

Scientific intelligence signals AI stepping out of digital application scenarios such as internet content and office work, and fully penetrating real-world research fields such as biomedicine and new materials. It deeply participates in scientific hypotheses, simulation and reasoning, and experimental verification, reshaping traditional research models. During WAIC, the “Excellence-Scale and Intelligence Fusion Platform” was officially released, delivering a breakthrough in China’s scientific intelligence compute infrastructure track. Crystal Technology Holdings showcased AI4S innovation achievements, accelerating R&D iterations and transformation of results in new drug and new materials domains.

In terms of global governance, during the conference, 29 countries signed an agreement to establish the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai. This is the first time in AI that a global governance framework has been built through equal multilateral negotiations. The conference brought together 11 winners of the Turing Award, the Nobel Prize, and the Fields Medal. These institutional breakthroughs may not directly produce commercial returns, but they provide governance infrastructure for the long-term healthy development of the AI industry.

If embodied intelligence and agents address the question of “what AI can do,” then scientific intelligence answers “what AI can discover,” and global governance answers “how AI should be governed.” These three questions operate on different time scales: the answers to the first may become gradually clearer within 3-5 years, while the answers to the latter two may take 10 years or even longer to become visible. Even if the proportion of signed deals directed toward scientific intelligence and governance frameworks is limited, its strategic value should not be underestimated.

Summary

WAIC 2026 delivered a standout scorecard with 32 projects, 40.9 billion yuan in signed deals, and 3 billion in global online traffic. AI infrastructure moving from “stacking cards” to “building systems,” embodied intelligence shifting from “showmanship” to “doing work,” and agents evolving from “conversation” into “hands-on execution”—these trends are real and deserve recognition.

But the “cold thinking” perspective reminds us: 40.9 billion yuan is a vote of confidence, not a guarantee of implementation. The distance from “can chat” to “can do” must be measured through sustained iteration of technological breakthroughs, engineering capabilities, data accumulation, and business model improvements. Whether a computing cost turning point has truly arrived, when robot brain technical routes will converge, whether agent task completion rates can continue improving, and whether scientific intelligence can move from concepts to substantive scientific discoveries—the answers to these questions are not on WAIC exhibition booths, but in the industry’s practical efforts over the next 3-5 years.

The value of WAIC 2026 may not lie in how many “things already done” it showcases, but in how clearly it helps people see what “must be done next.”

FAQ

Q: What specific areas do WAIC 2026’s 32 AI projects cover?

They cover four major areas: AI infrastructure, embodied intelligence, scientific intelligence, and agent applications, with a total investment amount exceeding 40.9 billion yuan.

Q: What are the global online traffic and exhibition scale for WAIC 2026?

Global online traffic exceeded 3 billion; offline audiences exceeded 400k visits; 1,117 companies exhibited; 4,486 exhibits were presented; and 351 products made their global debuts.

Q: What important changes did embodied intelligence see at this WAIC?

Embodied intelligence formed its own independent exhibition hall for the first time. The number of exhibiting companies rose from more than 80 to over 200; there were over 300 real machines, with almost all operating dynamically. The industry shifted from “showmanship” to “doing work,” and robots began entering real production lines and commercial scenarios.

Q: What trend did agents (Agent) show at WAIC 2026?

Agents moved from “conversation” to “hands-on execution.” They no longer stayed at booth demos, but entered real work flows and production workshops. The number of global daily active agents is expected to reach 79.4 million in 2026.

Q: What are the main obstacles for AI to move from “can chat” to “can do”?

Main obstacles include: the scissor gap between computing costs and model scale, the robot brain technical routes not yet converging, difficulties in obtaining training data, and the need to improve task completion rates and stability in complex scenarios.

Q: What outcomes did WAIC 2026 achieve in AI global governance?

29 countries signed an agreement to establish the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai—marking the first global governance framework in AI built through equal multilateral negotiations.