Author: Peng Zhao (Founder of Zhicifang, Co-founding Partner of Yunhe Ziben) Original
On the snow-capped plateau in Ganzi, Sichuan, at an altitude of 3,400 meters, temperatures approach -30°C, and strong electromagnetic fields are ubiquitous. Here, it is impossible to expect the massive amount of data collected by sensors to be transmitted back to the cloud thousands of miles away.
Faced with such extreme spatial constraints, engineers from Zhongguo Diankeyuan made a bold decision: they embedded a complex neural network into an ASIC chip dedicated to sensors.
Coincidentally, on a 3C electronics production line in Zhengzhou, over a thousand kilometers away, engineers from Lingyunguang are facing a test in another dimension: strict time constraints.
On the high-speed precision production line, the hand-eye coordination between the robotic arm and the vision system is measured in microseconds. A round-trip latency of tens of milliseconds to the cloud is merely an imperceptible pause when refreshing a webpage in the consumer internet, but on the production line, it represents the life-and-death boundary between misjudgment and good products. The solution from the Lingyunguang team is straightforward: "1-microsecond synchronization must be placed at the edge." They distilled the model into a sufficiently small version, installed it in the industrial control computer next to the production line, and completed real-time inference locally.
Such spatial constraints and time limitations are not uncommon in the IoT field.
As observers of the final evaluation of the 2026 IoT "New Technologies, New Applications, New Models" achievement selection (hereinafter referred to as the "Three News"), Wulianwang Zhiku participated in the entire evaluation process. When we saw these two truly occurring stories, we couldn't help but sigh that the current popular illusion about AI is being ruthlessly pierced.
Because in the industrial environment, AI has never been a central hub sitting in an air-conditioned room directing operations; rather, it is the nerve ending that must struggle and fight in the mud, the cold, and the millisecond-level life-and-death lines.
The closer to the industry, the clearer the rules of survival: Intelligence that cannot form a closed loop at the edge is, to some extent, pseudo-intelligence in the industrial field. With this cognition tempered by the harsh field, we deeply experienced the strong sense of contrast at the final evaluation site of the "Three News".
In the final evaluation meeting room of the "Three News", although the 822 application forms were mostly filled with buzzwords such as AI, large models, and AI agents, the expert review team led by Yin Hao, an academician of the Chinese Academy of Sciences, was not swept away by these dazzling concepts, but instead performed with exceptional calmness and sharpness.
In the previous article, we revealed the outward value change of IoT, where data flow evolved from "for humans to see" to directly "for AI to use". In this article, we will focus inward to explore how IoT and AI will integrate with each other as AI serves as a general-purpose technology, and create replicable capabilities.
Therefore, at this evaluation site representing the highest level of China's IoT, Wulianwang Zhiku recorded the three core questions from the judges:
If AI makes a misjudgment, who will bear the losses?
If the network is disconnected, can the system operate independently without the cloud?
When serving the second customer, do we still need to charge so much money?
Behind these questions lies the life-and-death proposition that has remained unsolved in the IoT industry for decades: fragmented scenarios and the heavy dilemma of the project-based system. This year, in this seemingly dull and even slightly gunpowder-scented meeting room, we finally saw the dawn of untying this industrial dead knot.
The Most Difficult Leap: Taming "Probabilistic AI" with a "Cage of Certainty"
The first question thrown by the judges hit the nail on the head: If AI makes a misjudgment, who will bear the losses?
This sharp questioning reveals the cruel reality of AI's maladaptation in industrial scenarios.
In the greenhouse of the laboratory, the industry often believes that the technological high ground of AI lies in computing power scheduling, model architecture optimization, or pushing the accuracy rate from 98% to 99%. But at the site of the "Three News" selection, as observers, we clearly felt the consensus of the judges: The true deep water zone lies in how to make AI, which is accustomed to probabilistic output, adapt to industrial sites with zero fault tolerance.
This is essentially a fierce conflict of underlying logic.
The operating mechanism of large models is built on statistical probability. Whether it is natural language processing or image generation, it is inherently divergent and uncertain. A 99% accuracy rate is already a huge success for a large model that writes poems and paints, but in the context of industrial manufacturing and critical infrastructure, the fault tolerance is extremely low. That 1% error is a harmless hallucination on the consumer end, but it becomes a tragic accident in a factory, and a devastating disaster in the power grid.
Take the power grid-level test faced by Guojia Dianwang as an example. In high-voltage transmission lines of hundreds of thousands of volts or extremely harsh substations, the equipment must first undergo the physical tempering of life and death. No matter how advanced the AI algorithm is or how massive the model parameters are, if the edge computing node carrying it crashes under strong electromagnetic interference or breaks down under extreme temperature differences, all intelligence instantly drops to zero.
Here, extreme physical reliability is merely a ticket to enter, and intelligence is just a bonus point after surviving.
Look at the real exploration of Yiqi Jituan in the intelligent transformation of production lines. With the advancement of software-defined vehicles, a single vehicle involves over 200 software versions, spanning 9 categories and 157 sets of equipment. In the complex links of software configuration, flashing, testing, and tracing, traditional rule systems and manual verification have touched the physical boundaries of efficiency and consistency. However, when introducing AI agents to intervene in production decision-making, Yiqi Jituan did not blindly hand over all control rights, but built an extremely rigorous safety net mechanism.
Under this mechanism, the decision-making authority of AI is strictly graded: for low-risk operations such as fine-tuning environmental parameters and recording equipment status, AI can execute automatically; but for high-risk operations involving changing core welding parameters and issuing shutdown commands, AI only acts as an advisor to give suggestions, and must be executed only after manual review.
This is by no means a distrust of new technologies, but an instinctive defense of safety boundaries by the industrial system. At the bottom of this safety net mechanism is rigorous authority auditing and second-level rollback response.
Once the AI output is abnormal, the system must be able to perform authority circuit breaking and physical isolation on it within milliseconds, ensuring that the production line instantly recovers to the safety baseline.
Therefore, for current IoT enterprises, the most arduous task is not to endlessly develop a smarter AI, but to personally build a cage of certainty and firmly tame the probabilistic AI within it.
The bars of this cage are forged by five unbreakable industrial iron laws: reliability, safety, cost, adaptation, and delivery.
Only when AI, under the strict constraints of these five checkpoints, sheds its wild probabilistic nature and transforms into a predictable, controllable, and traceable deterministic tool, can it be considered to have truly obtained the pass to enter the industrial site.
And once AI is successfully "tamed", the huge commercial potential it releases may reshape the business model of the IoT industry for decades.
Calculating the Bill for the Second Production Line: From "Manual Customization" to "Asset Reuse"
When AI is successfully tamed within the cage of certainty, the judges immediately threw out that soul-striking question, which also leaves IoT practitioners with mixed feelings: When serving the second customer, can the marginal cost be brought down?
This is one of the most profound inspirations brought by the "Three News" final evaluation site to the industry: The value of AI to IoT does not lie in optimizing the accuracy of a certain sensor or the parameters of a certain device to the extreme, but in the possibility that it will reconstruct the commercial monetization logic of the entire industry.
Let us first uncover the most painful scar of the IoT industry in the past few decades.
As the birthplace and industrial highland of China's IoT, Wuxi has gathered over 3,500 IoT enterprises. In 2025, the revenue above designated size of the city's IoT industry reached as high as CNY 318.4 billion. In the first half of 2026, this industry continued its growth trend, with revenue above designated size reaching CNY 139.92 billion, a year-on-year growth of 8.4%. The prosperity of Wuxi is a microcosm of the vigorous development of IoT in China and even the world.
However, stepping out of Wuxi and looking at the entire industry, an embarrassing common pain point always lingers: Why is it difficult for a track universally recognized as having a trillion-level broad prospect to hatch enough towering trees with a market capitalization of CNY 100 billion?
The answer is hidden in three words: Project-based system.
The fragmentation barriers in industrial sites are extremely strict. Thousands of industries have thousands of faces, and even different workshops in the same factory may have completely different equipment interfaces and process flows. In the past, the business model of IoT enterprises was essentially a customized supplier wearing a tech coat. Winning the benchmark project for the first production line often requires dispatching a team of software engineers to stay up late writing code and debugging interfaces on site, fighting hard for 3 months.
When the enterprise happily accepts the order for the second production line, they despair to find that due to the differences in scenarios, the previous code can hardly be reused, and they can only send people again and endure another 3 months.
The cost of the first production line is high, and the second production line remains high. Although the revenue scale is indeed growing, the meager profits are all swallowed by high engineer travel expenses and on-site development costs. This linear growth model of manpower stacking has firmly locked the path to scaled leapfrogging for IoT enterprises.
But now, the explosion of AI technology finally holds the promise of smashing this ceiling that has imprisoned the industry for many years.
In the representative industrial vision inspection case at the final evaluation site, we saw a shocking qualitative change.
In the past, as long as a new model or new product was introduced on the production line, software engineers who understood algorithms needed to rush to the site immediately, rewrite code and adjust parameters for the machines, usually taking 3 days to get the vision inspection system running again.
After introducing AI agents based on large models, the situation has fundamentally reversed. Today, when introducing a new model, there is no need for software engineers to travel. OT engineers on site who understand the process but do not know the code only need to use simple natural language interaction and graphical drag-and-drop within the system. In just 4 hours, they can make the AI automatically generate and adapt to the new inspection model.
From software engineers staying on site for 3 days to write code to process engineers dragging and dropping for 4 hours, what is saved in between is absolutely not just travel expenses and working hours, but the realization of a leapfrog change in the business model.
After the first demonstration line was successfully verified, the precipitated "golden line" standard was quickly reused on over 30 production lines. This scaled expansion not only increased the line building efficiency by 30%, but also drove the overall line output from about 450 UPH to a substantial increase to 685 UPH, and the yield rate steadily crossed from 95% to over 99.5%.
This is precisely the profound significance endowed by AI to IoT: It has truly realized the assetization of experience.
In the past, solving the fragmentation problem highly relied on human experience, whether it was the intuition of an old master or the code of a senior engineer. Because humans cannot be infinitely copied, the customization cost has always been at a high level.
Now, AI digests these implicit, non-standardized industry experiences, and finally refines them into replicable recipes, generalizable models, and readily available AI agents.
When experience is transformed into digital assets, IoT truly leverages the economic lever against the curse of the project-based system: diminishing marginal costs.
Training an AI large model that understands industry and creating the first benchmark project indeed faces high initial R&D costs, requiring enterprises to invest massive computing power, data, and top talent. However, when this AI model containing profound industry know-how moves to the second, tenth, or even the hundredth production line, the marginal cost of its scaled deployment will significantly decrease.
The bill for this second production line, today in 2026, finally has the possibility of being amortized. Crossing this threshold, IoT enterprises are expected to bid farewell to the fate of the project-based system of doing business order by order, and truly transform into value engines with core digital assets.
The "Three Rulers" for Measuring AIoT: From "Technology Parade" to "Scaled Landing"
When Wulianwang Zhiku listened to the judges' in-depth analysis of the core value of the projects at the final evaluation site of the "Three News" selection, we realized that this selection has long surpassed a simple technology parade and transformed into a weathervane for the evolution of China's IoT industry.
The three sharp questions thrown by the judges are essentially the "three rulers" for measuring the commercial and technical quality of AIoT:
The first ruler measures depth: At which layer does the intelligence actually fall?
It forces enterprises to answer whether AI is merely pseudo-intelligence staying on PPT presentations and visual dashboards, or whether it has truly penetrated the OT control layer. Can it build a real closed loop that can perceive, decide, and execute amidst the roaring machines and rapidly running production lines?
The second ruler measures the bottom line: How to provide a safety net when the system makes a misjudgment?
It questions the technology: whether to let the large model have fatal hallucinations in the industrial site, or whether a rigorous authority audit and rollback mechanism has been established. Has the enterprise successfully tamed the probabilistic and uncertain AI behemoth with a cage of certainty?
The third ruler measures the future: What is the marginal cost when serving the second customer?
It strikes at the soul of the business model: whether the enterprise continues to sink deep into the mire of fragmentation, acting as a customized outsourcing provider relying on manpower stacking, or whether it has already realized the assetization of experience with the help of AI. Has it truly crossed the chasm of diminishing marginal costs and endowed products with replicable scaled attributes?
The real cases at the final evaluation site are precisely the most powerful response to these three rulers.
Taking the practice of Yiqi Jituan as an example, when the system was replicated across factories from the three major factories of Hongqi to the Yancheng base, its platform core, data models, and rule assets achieved comprehensive reuse, and the only things that truly needed to be redone were the on-site semantics and equipment access.
The team frankly stated that among the six layers of capabilities built, the most invested and hardest bone to chew is the evidence base, which is to unify the heterogeneous protocols, vehicle model information, software versions, and quality results of 9 major categories and 157 sets of equipment into traceable standard semantics. However, once this thick base is constructed, every subsequent replication substantively amortizes its early-stage costs, allowing the marginal cost to significantly decrease.
Liantong Shuke, on the other hand, confirmed this logic from another side. They unified more than 140 robot communication protocols into a single platform, enabling standardized and unified centralized management of robots of different brands and forms.
The mire of protocol fragmentation that has plagued the IoT industry for decades is no longer an unsolvable industry curse here, but has been reduced in dimension to an engineering problem that can be directly conquered through a standardized platform.
These three rulers not only accurately measured the hardcore value of this "Three News" selection, but also clearly outlined the direction for breaking the deadlock in the next phase of the IoT industry in Wuxi and even China.
Conclusion
On October 30, the curtain of the 2026 Shijie Wulianwang Bolanhui will officially open on the shores of Taihu Lake. At that time, the "Three News" achievements that have withstood the strict measurement of the "three rulers" will officially step onto the stage under the spotlight.
Through these weighty achievements, we have already glimpsed a clear picture of industrial reconstruction: the heavy applications from Fortune Global 500 companies and C9 Gaoxiao announce the comprehensive entry of leading forces; the track renaming from Industrial Internet to Industrial AI Agents reflects the underlying logic of technological evolution; giants such as power grids and oil and gas have incubated self-use scenarios into standard products, completing the identity transformation from demand side to supply side; and the large-scale sinking of algorithms to chips and industrial control computers, as well as the reshaping of the business model from selling equipment to selling capabilities and services, have truly consolidated intelligence to the forefront of the industry.
The participating projects may not be perfect yet, and may still be struggling to integrate with the dust and noise of the industrial site, but they represent the most real and bravest exploration of China's IoT. When AI truly sheds its magical halo, turning into lines of deterministic instructions in the workshop and becoming tangible profits on the reports, the intelligent era we call for will truly arrive.
And for the answers to all this, we will see you in Wuxi in autumn.