Author: Peng Zhao (Founder of Zhicifang, Co-Founding Partner of Yunhe Capital) Original
Recently, I had an in-depth conversation with Tao Jianhui, founder of Taos Data. He once shared a highly visual detail: when he visited a cigarette factory and chatted with an on-site engineer, the engineer looked him over and said, "You don't even smell like smoke."
This remark was about him, but it actually applies to today's AI large models. No matter how unprecedented their language capabilities are, once they enter a factory, they are just "new employees." They have no idea what "the outlet temperature of Furnace No. 3 shows 620 degrees" actually means, which production line this furnace belongs to, what the normal fluctuation range is, or how its upstream and downstream linkage mechanisms work.
The dilemma of how large models can enter factories has been discussed in the industry for a long time. In this conversation, I compiled some data from the prospectuses and annual reports of five leading Industrial Internet companies, totaling over 1,700 pages.
If we put these five financial reports and archives alongside overseas industrial digitalization samples from the past decade or so, we can answer a question that predates and is even more critical than "how large models enter factories": Why did the previous generation of Industrial Internet platforms, which "built the roads" for large models, fail one after another?
The Shattered Platform Myth: A 14-Year "Experimental Report"
Let us first turn our attention overseas. In 2012, GE (General Electric) first introduced the concept of the Industrial Internet globally, and in 2013, it officially launched the Predix platform product. GE once invested over $4 billion in its digital transformation, with its digital department peaking at nearly 28,000 employees. However, after GE restructured its digital business in 2018, the grand narrative of its "horizontal platform" came to an end. Siemens MindSphere made a high-profile debut in 2016 but was quietly renamed Insights Hub in 2023, downgraded to a product component. In the same year, Google IoT Core and the IBM Watson IoT platform were successively and quietly shut down. SAP's Leonardo brand had also already faded away.
Bringing our focus back to China. The Ministry of Industry and Information Technology's (MIIT) "Dual-Cross Platform" list expanded rapidly from 10 companies in 2019 to 51 in 2023, before subsequently shrinking as the evaluation system began to implement tiering. In the capital markets, Rootcloud withdrew its IPO application from the STAR Market in 2023, marking one of the turning points.
According to Frost & Sullivan's statistics, in the domestic market for platform-based industrial data intelligence solutions, the market share of the top-ranked enterprise is only 1.7%. After more than a decade of grand industry narratives and seven years of competition for the Dual-Cross list, with over 50 top-tier enterprises selected successively, the aggregated "industry leader" holds a market share of less than one-fiftieth.
This is not just a single company's report card; it is a ruthless industry experimental report.
The conclusion is obvious: In the industrial sector, there is no such thing as "platform economics."
The word "platform" was borrowed from the consumer Internet. The core premise for its success is called "network effects": the more users there are, the more valuable the network becomes for other users. However, industrial data inherently lacks this attribute. Connecting Factory A's production line to a platform does not make Factory B's connection more valuable; the cross-customer externality is virtually zero. Without network effects, there is no winner-takes-all.
The second dilemma lies in the fact that the marginal cost of service is "people."
The server cost for a consumer platform serving 100 million people is barely different from serving 1 million. However, for an industrial platform, every additional customer requires adding another project team. GE's digital department once swelled to nearly 28,000 people; Rootcloud's employees surged from over 500 to 1,500 within two years. In 2021, the per capita revenue was 340,000 RMB, but the per capita allocated loss that same year exceeded this figure. The revenue generated by each employee could not even cover the financial deficit allocated to them.
The third dilemma is that the "anchor" for pricing was set incorrectly.
The previous generation of platforms wanted to charge entry fees: connection fees, platform usage fees, and subscription fees. However, looking at the past fifty years of industrial history, customers are only willing to pay for two things: one is the verified certainty of results, which is why the project-based model prevails. But the cost of the project-based model is obvious: once downstream capital expenditure hits the brakes, project-based revenue drops to zero in the current period, with no cushion whatsoever.
The other is the "annuity" with extremely high switching costs, and the standard answer is OSIsoft (PI System). This company, with an annual revenue of approximately $400 million, was first acquired by AVEVA for $5 billion, and Schneider Electric subsequently took AVEVA entirely private at a valuation of over $10 billion. What are the giants buying? They are buying decades of data accumulation across more than 20,000 sites globally, along with the massive applications built on top of it. Customers simply cannot switch, so they have to pay maintenance fees year after year.
In contrast, the previous generation of platforms occupied neither end: they lacked the hardcore delivery capability to charge by results, nor did they have legacy assets that made it impossible for customers to switch.
The Turning Point for Breaking the Deadlock: How AI Unties "2.5" Knots
The turning point has finally arrived. The emergence of AI has directly shaken the foundations of "two and a half" of the three major dilemmas mentioned above.
Let's look at the second dilemma first: marginal cost. Tao Jianhui told me that over the past year, he has been using AI to write code and documents almost every day, and the company's internal R&D efficiency has improved by at least 5 times. The production cost of software has achieved a cliff-like drop first. More importantly is the service cost: deployment, Q&A, troubleshooting, reporting, dashboards... In the past, every customer required a dedicated project team. Now, these tasks can be entirely handed over to AI agents in the system. The marginal cost of service has been compressed exponentially for the first time in the history of industrial software.
Now let's look at the first dilemma: network effects, which has been half-solved: semantics. In the past, doing data standardization and building information models was essentially doing "public welfare" for the next project. The work was done for the current contract, but the benefits went to the next one, and no customer was willing to pay for it separately. As a result, standardization could never be pushed forward; fifty Dual-Cross platforms might as well have had a hundred sets of models.
For example, when the number "620" exists in isolation, it means nothing. It could be temperature, rotational speed, or even just an arbitrary code. The so-called "semantics" is the entire set of instructions attached to this number. It represents temperature, the unit is Celsius, the measurement point is at the outlet of Furnace No. 3, belonging to Production Line 1 in Workshop 2. Its normal fluctuation range is between 580 and 650 degrees, and it must trigger an alarm if it exceeds 680 degrees. Its upstream process is batching, and its downstream process is cooling. Only by attaching these tags can an AI large model truly understand what this number is and what it is related to when it receives it. This is like compiling a "dictionary" for all the factory's data, allowing the AI to look up the true meaning of each word one by one.
AI large models may change the revenue structure of this endeavor. Whether the ontology is well-built directly determines whether the AI can do the job. Only when every object has an identity, every relationship is explicitly expressed, and every data point carries business semantics, can the large model truly understand the phrase "620 degrees."
For the first time, semantics has a direct buyer, and the "dictionary" may become a core product. Whose semantic model becomes the de facto standard will likely hold the entry point for reading industrial AI.
The third dilemma, the pricing anchor, has shifted the most profoundly, because the anchor for pricing has completely changed.
AI large models are inherently probabilistic and uncertain, while the industrial site demands absolute certainty. Therefore, basic "functions" can be free, as AI has significantly reduced the production cost of code and functions; but "responsibility" cannot be free. What enterprises are truly willing to pay for is the "insurance" to ensure this system does not make mistakes, including service guarantees, compliance audits, permission grading, and critical oversight with humans in the loop (Human-in-the-loop).
This constitutes a solid paywall between the commercial and free versions. To summarize in one sentence: Functions are free, responsibility is paid for.
Here, we need to introduce a new concept. The annuity of PI System is anchored on "data cannot be moved away," but AI is continuously breaking through migration costs. Code can be generated, data can be moved, and even the ontology model might be exported with one click. The asset of "cannot be switched" is rapidly depreciating. The annuity model will not disappear, but the anchor may shift: from the passive binding of "data cannot be moved away" in the past, to the active service of "someone must bear the responsibility" today.
With the 2.5 dilemmas explained, the remaining half is: the network effect at the customer data layer, which AI has still not created. Factories still do not want to see other companies' data, and a single-factory closed loop remains the iron rule of the industry...
But miraculously, network effects seem to have sprouted.
New Growth Flywheel: Competing for "Model Memory" and "AI Sovereignty"
The first real growth flywheel in the history of industrial software may grow within the "memory of AI large models."
In the past, software companies needed to expend massive sales resources to convince enterprise CIOs to buy their products. But in the AI era, more and more code writing, system integration, and technology selection are being automatically completed by AI agents. The decision-makers for software procurement are shifting from "humans" to "models."
If your industrial software has a high degree of open-source and appears more frequently in AI training corpora and tool registries, the AI will default to prioritizing your tools when generating solutions. As deployment increases, it will generate more corpus, forming a positive cycle.
The previous generation of platforms competed for CIO budgets; this generation of infrastructure competes for "model memory."
At the same time, what is even more interesting is that Palantir, one of the most successful data companies overseas, emphasized another selling point in its financial report: "AI sovereignty."
Palantir's revenue reached $4.475 billion in fiscal year 2025, a growth of 56%; in the second quarter of 2026, revenue was $1.94 billion, a growth of 93%, with a single-quarter net profit of $1.06 billion and a net profit margin as high as 55%. In his signed commentary in the 2025 annual report, CEO Alex Karp referred to this trend as "the commoditization of cognition." Half a year later, he pushed the selling point further to "AI sovereignty." He repeatedly assured customers that their factories' private data and process advantages would absolutely not be used to train any public models.
On one side, open-source tools gain AI's favor and distribution by "entering the corpus"; on the other side, Palantir sells trust to customers by "isolating the corpus."
This seems contradictory, but it actually pieces together a complete commercial landscape for industrial AI: as "tools" for infrastructure, the more public they are, the better, as this determines whether they can be widely called by AI; while for customers' specific "process contexts," the more private they are, the better, as this determines the factory's core competitiveness.
What is "context"? Context is the data point and its surrounding background, the overall situation at that time and place. Merely recognizing "the outlet temperature of Furnace No. 3 is 620 degrees" is far from enough. What product batch is this furnace producing right now? Is it in the heating-up phase just after ignition, or has it been in stable production for 8 hours? Was a new batch of raw materials swapped in last night? Does the veteran worker at this factory habitually keep the temperature 10 degrees lower than the industry average? The same 620 degrees is completely normal during the heating-up phase, but during stable production, it could be a precursor to a fault. All the background information needed to judge whether this data point is good or bad is the context.
In short: Semantics allows the AI to "read the words," while context allows the AI to "understand the situation." Semantics is the label for a single data point, while context is the ecology of the entire site.
The cigarette factory engineer's remark at the beginning of the article, "You don't even smell like smoke," precisely exposes the AI's dilemma of lacking context. The large model recognizes the two characters for "temperature," but it hasn't struggled and worked in this factory; it doesn't understand the rules, habits, and current emergencies here.
Understanding this layer makes it clear why, commercially, semantics and context must be treated separately.
Dictionaries (semantics) can and must be standardized. Temperature should be uniformly called temperature; it cannot be Temp in one factory and T01 in another. Whose dictionary format becomes the universal standard wins the ecosystem, so it is naturally suited to be built as public infrastructure.
But within the context lie each factory's "unique skills"—formulas, parameters, and the decades of hands-on experience from veteran workers. This is the core asset for enterprises to beat their peers, and it must never be reduced to public knowledge, let alone become training data for someone else's model.
As a tool, the semantic standard is better the more public and widely used it is; while as a core asset, the on-site context is more valuable the more private it is. The entire business of this generation of data infrastructure is about doing these two opposite things well simultaneously.
Two routes diverge from here:
One is Palantir's "high-priced bespoke" route: helping clients build private models + close-proximity engineer services + taking responsibility for the results. This route is extremely expensive and picky about clients. The 55% net profit margin in a single quarter proves its extremely high ceiling, but it is essentially a high-end project-based model, bypassing platform economics.
The other is the "scale path": open-source underlying engines, free basic platforms, and betting on scale and ecosystem. This is exactly the path that the previous generation of Industrial Internet platforms failed to build. The bet is that AI can truly bring down the marginal delivery cost.
Conclusion
Looking back over the past 14 years, the pioneers of Industrial Internet platforms cannot be said to have not worked hard, but they tried to force-fit the "platform economics" of the consumer Internet onto the complex reality of industry, ultimately failing to escape the gravitational pull of business models.
Today, the emergence of large models does not merely provide a smarter chat box; it restructures the cost structure and value anchors of industrial software.
When the marginal cost of code generation is significantly diluted, when semantic models make data truly readable, and when the business model shifts from "selling tickets" to "charging for responsibility," the gears of industrial digitalization truly begin to mesh.
In the second half of large models entering factories, the competition is no longer about grand narratives, but about who can fastest help AI acquire that "smell of smoke" in the muddy industrial sites.