The Ragnarök of Wrapper Startups: Why the AI Project You Bet On Became a Bad Asset Overnight
Prologue
Three months ago, a friend I’ve known for years sent me a voice message late at night. His tone carried a kind of suppressed excitement that I know all too well.
“Old friend, I’m all in. An AI project. We’ve secured angel funding at a thirty million valuation. The product is an AI writing assistant built on GPT-5. I’ve spent three months polishing the interface, with over fifty templates, and the user growth curve is beautiful. This time, I’m going to seize the dividends of this era.”
I listened, then paused for a few seconds. I’ve encountered this brand of excitement too many times in my thirty-year career in credit approval. Every project that later became a non-performing asset began with its founder’s eyes burning with that same light on that first night.
I sent him back a text. A text that he later screenshotted and set as his phone wallpaper, but only after his project had cratered to zero.
“You’re not building a startup. You’re building a villa on someone else’s foundation. The owner of that foundation can demolish your house whenever they want, or even build an identical one for free. What is your core asset? That single line of code calling an API. But in their financial statements, that line of code is on the liability side—they can cut off your ‘supply’ at any moment.”
He didn’t believe me. He said my years in banking had made me too risk-averse, that I saw risk everywhere and was missing the opportunity of a lifetime.
Today, three months later, his project’s valuation is zero. The team is disbanded, the servers have expired, and the website returns a 404 error. His life savings of several million, along with the angel investment, have all gone down the drain.
He called me. His voice was no longer filled with excitement, but with the exhaustion and confusion of someone who has been hollowed out.
“The product was clearly great, user growth was fast, and retention was good. Why did it all fall apart overnight? What did I do wrong?”
I didn’t answer immediately. Because I knew this wasn’t a simple question of “what you did wrong.” This is a thirty-year-old commercial “cold case” that has replayed itself repeatedly across different eras, wearing different masks.
In 1985, a group of people did the same thing. In 1995, another group followed. In 2005, yet another. Each time, they believed they were the trendsetters of their era. Each time, they discovered they were swimming naked only when the tide went out.
Today, from the perspective of a risk officer, I will put on my detective’s glasses and take you back to the beginning of this “cold case.” Using the non-performing loan files I’ve handled personally, the true stories of founders whose hair turned white overnight, and the most fundamental logic of economics, I will help you tear through this final layer of paper.
After reading this article, you will understand: Why is the collapse of “AI wrappers” not an accident, but an inevitability? In the AGI era, what are the truly valuable assets? And why is the greatest risk for the average person not missing a trend, but taking on leverage they shouldn’t?
This story begins in 1985, in Zhongguancun.
Chapter One: Echoes of History: When the PC Assemblers Met Microsoft’s Rent Collector
I. 1985, The “Golden Age” of Zhongguancun
In 1985 Beijing, there was an alley in Zhongguancun that would later be known as “Electronics Street.” It was crammed with storefronts of all sizes, adorned with colorful, handwritten signs: “Custom-built PCs,” “Compatibles,” “286/386.”
Back then, IBM had defined the standard for personal computers, but its machines were prohibitively expensive, costing twenty to thirty thousand yuan—several years’ salary for an average worker. A group of clever individuals spotted a business opportunity: smuggle CPUs, hard drives, and memory from Hong Kong, buy cases and power supplies from Taiwan, source operating systems from the United States, and then assemble these parts into “compatible” machines under their own brand.
These people were later called “PC assemblers.”
Many of them are now entrepreneurs worth billions. But in the winter of 1985, they were just crouching behind counters, screwdriver in hand, assembling machines one by one, debugging drivers over and over, and patiently teaching customers how to use DOS commands.
An old bank president I know got his start assembling PCs. He later moved into investment. Once, over drinks, he spoke of those years with a redness in his eyes.
“You know what the most profitable thing was back then? Not selling whole machines. It was selling ‘optimization software.’ The DOS system was too primitive, memory management was a mess, and the hard drive would crash if it got too fragmented. A few of us wrote some batch scripts, packaged them into a piece of software called ‘Memory Butler,’ and sold it for 50 yuan a floppy disk. The cost was less than 50 cents. That was the real windfall.”
His “Memory Butler” was the “wrapper product” of its time. It was built on top of Microsoft’s DOS, adding a layer of an “optimization interface” and “automation scripts.” Users who weren’t tech-savvy found it incredibly convenient and paid up.
From 1985 to 1990, it was the golden age for PC assemblers and wrapper software vendors. They rode two waves of dividends: first, the popularization of the personal computer; second, the imperfections of the operating system. They were like people standing in an elevator, doing nothing while the elevator itself was ascending.
But they forgot that the elevator button was in someone else’s hand.
II. 1995, The Crushing Blow of Windows 95
On August 24, 1995, Microsoft released Windows 95.
The significance of this date only becomes clear in hindsight. It wasn’t just an operating system update; it was a thorough cleansing of the “wrapper ecosystem.”
What did Windows 95 bring?
First, it had a built-in graphical interface. Before, you had to learn DOS commands to use a computer. This gave rise to a plethora of “DOS tutorial software,” “command cheat sheets,” and “batch script assistants.” After Windows 95, you could perform almost any task with a simple mouse click. The value of those “tutorial” products vanished overnight.
Second, it had built-in networking capabilities. Previously, getting online required installing a stack of drivers, configuring numerous parameters, and buying something called “dial-up software.” After Windows 95, networking was a native feature. Those “dial-up assistants” and “internet accelerators” all disappeared.
Third, it came with a built-in office suite. Although Word, Excel, and PowerPoint weren’t bundled directly with the OS, Microsoft’s bundling strategy instantly wiped out the market for third-party office software vendors. The most classic example is Lotus 1-2-3, which once had a market share of over 80%. Within three years of Windows 95’s release, its share plummeted to single digits.
The logic behind Windows 95’s crushing blow was simple: the operating system is the “foundation.” For any “value-added service” built upon it, if the foundation’s owner decides “this is a service I should provide,” they can simply build it into the next version and offer it for free.
This wasn’t market competition; it was the appropriation of an ecological niche.
My old bank president friend, the one who started with PCs, made an immediate decision after Windows 95 was released: he cut all his “optimization software” business and focused solely on hardware sales.
“I didn’t understand it all at the time, but I had a simple intuition,” he later told me. “Microsoft was doing our job for us, doing it better, and not charging for it. What was left for us to sell?”
That intuition, in retrospect, is one of the most fundamental truths of the business world: those who build their business on a platform are forever at the mercy of that platform.
III. A Mirror of History: From “Memory Butler” to “AI Writing Assistant”
Now, let’s pull the timeline back to 2026.
If we compare the Zhongguancun of 1985 with the AI startup boom of 2023-2025, you’ll find the script is almost identical.
In the DOS era of 1985, the underlying system (DOS) was rudimentary and had a poor user experience. Users needed “optimization tools” to use it comfortably. Thus, “Memory Butler,” “Disk Defragmenter,” and “Command Helper” were born.
In the “early stage of large models” from 2023-2025, the underlying models (GPT-3.5/4) had limited capabilities and unstable interfaces. Users needed “wrapper tools” to use them smoothly. Thus, “AI writing assistants,” “AI summarization tools,” and “AI art templates” sprang up like mushrooms.
The core asset of the 1985 “Memory Butler” was a few batch scripts and the author’s understanding and “optimization” of the DOS system.
The core asset of the 2025 “AI writing assistant” is a line of API-calling code, the developer’s prompt engineering, and UI design.
The creator of the 1985 “Memory Butler” had no idea what Microsoft would do with Windows 95.
The creator of the 2025 “AI writing assistant” has no idea what OpenAI will do with its next-generation model.
They are both unaware that they are building on a foundation where the rent can be collected at any time.
In early 2026, OpenAI and Anthropic released their new-generation native multimodal models almost simultaneously. The “capability leap” of these models was not incremental; it was a mutation.
First, native multimodality. Previously, to make an AI “understand” an image, you had to pass the image to an “image recognition API” and then feed the result to the large model. This process required you to write code, handle integrations, and manage errors. Now, the new models can “see” on their own. You just upload the image, and the model understands it directly. What does this mean? It means the core function of all “AI image recognition tools” and “AI image-to-text tools” has been “built-in for free” by the underlying model.
Second, ultra-long context memory. Previously, an AI’s “memory” was short, limited to a few thousand tokens per conversation. If you wanted AI to summarize a 10,000-word article, you had to manually break it into chunks and feed it piece by piece. This created a market for “AI long-text summarizers” and “AI document processors.” Now, the new models can remember millions of tokens, equivalent to an entire book. You can throw a 10,000-word article at it, and it will read and summarize it in one go. Those “chunking and summarizing tools” have lost their reason to exist.
Third, a dramatic improvement in reasoning ability. Previously, an AI’s “thinking” was shallow. You had to use complex prompts to “guide” it toward a high-quality answer. This gave rise to the “voodoo” of “prompt engineering” and a slew of “prompt template libraries” and “prompt optimization tools.” Now, the new models can “think” for themselves. You don’t need to meticulously design a prompt; you can just ask directly. Those “template tools” have become utterly obsolete.
This combination of blows rendered the “core assets” of wrapper AI companies—their UI, template libraries, prompt engineering, and multimodal integrations—completely “overwritten for free” by the new version of the underlying model.
Just as Windows 95 integrated a graphical interface, networking, and office software, the new generation of large models has integrated image understanding, long-text summarization, and reasoning optimization.
This is not competition. This is a crushing blow from a higher dimension.
IV. The Villa on the Foundation
Looking back at the history from 1985 to 1995, we can extract three fundamental laws. These laws apply equally to the AI track today.
Law 1: The value of the platform is always greater than the applications built upon it.
The operating system is the platform; applications are the products on it. The platform owner controls the “rules of the ecosystem.” They can decide who gets to build a house and can also build the house themselves in the next version, kicking out the original owner. This is the cruelest truth in business: profits always flow to the controller of the ecological niche, not its occupants.
Law 2: The evolution of a platform trends toward consuming downstream value-added services.
Why do platforms evolve? Because their owners are always looking for new growth drivers. Once the basic functions are mature, a platform will extend upward, consuming the market of companies that provide “value-added services” on top of it. Microsoft, moving from DOS to Windows 95, consumed the market for “optimization software” and “office software.” OpenAI, moving from GPT-4 to GPT-5/6, is consuming the market for “wrapper AI.”
Law 3: When you do business on a platform, you are always leveraging “someone else’s capital.”
In economics, we call the dividends from an overall industry upswing a “Beta” return. The “Memory Butler” vendors of 1985-1990 reaped the Beta returns of DOS popularization. The “wrapper AI” companies of 2023-2025 reaped the Beta returns of large model popularization. But Beta returns are borrowed leverage. When the platform owner enters the game and the tide recedes, the companies with only Beta returns and no “Alpha” (value they create themselves) will be the first to fall.
My old bank president friend, the “Memory Butler” creator, made a decision in 1995: cut the software business and pivot to hardware. That decision saved him. Because while hardware had thin margins, it was a “physical asset.” Microsoft couldn’t “build in for free” a physical computer.
And those “wrapper software” vendors who didn’t pivot in time? They were almost completely wiped out in the three years following 1995.
Today, the same script is playing out on the AI stage.
The question is: which side are you on this time?
Chapter Two: The Current Conundrum: Deconstructing the Truth of Wrapper AI’s Bad Assets
I. The Risk Officer’s Scalpel: Why Your Core Asset is Actually Their Liability
Let’s step away from history for a moment and return to the present day of 2026.
I am now going to do what I’ve done for thirty years: conduct a full “credit assessment” for a “wrapper AI” company. This isn’t an academic exercise. This is a risk officer taking a scalpel to a company’s balance sheet, peeling it back layer by layer, to see what its “core assets” really are.
Let’s call my friend’s company “SmartWrite Technologies.” What did its balance sheet look like three months ago?
Assets:
- Cash: 3 million (Angel investment)
- Intangible Assets: An AI writing platform codebase, valued at 20 million
- User Data: 50,000 registered users, 3,000 paying users, valued at 5 million
- Other: Servers, office equipment, etc., 1 million
- Total Assets: 29 million
Liabilities:
- Accounts Payable: 500,000 in outstanding API fees
- Other Payables: 300,000
- Net Assets: 28.2 million
On the surface, this looks like a healthy company with “core assets.” The code is worth 20 million, the users 5 million. Together, these 25 million in intangible assets account for 86% of total assets.
But to a risk officer like me, these “intangible assets” are not assets. They are “toxic assets.”
Why? Let me tell you about a real case I handled.
It was 2008, and I had just become the general manager of the head office’s credit approval department. A company that made a “mobile phone input method” came to us for a loan. Their product was incredibly popular, with tens of millions of users and high daily active usage. Their core asset was the input method’s code.
In my approval opinion, I wrote four words: “Denied. Asset not independent.”
The bank president asked me what I meant. I said, “This company’s input method runs on Nokia’s Symbian OS. If Nokia decides tomorrow to build its own input method into the system, this company is finished.”
What happened next? In 2010, Nokia did start building in its own input method. In 2011, that company went bankrupt. The founder called me and said, “Mr. Zhang, you were right.”
What is the core logic of this case? It’s the issue of “asset independence.”
In a bank’s risk assessment framework, an asset’s value isn’t determined by how much money it’s making now, but by how “independent” it is. If an asset’s existence and appreciation depend entirely on the “cooperation” or “non-competition” of another entity (the platform owner), then it’s not a true asset. It’s “borrowed leverage.”
What does the value of “SmartWrite’s” code depend on?
First, it depends on OpenAI not “building its features in for free.” If OpenAI’s next-generation model can do AI writing itself, and do it better, faster, and cheaper, then “SmartWrite’s” code is worthless.
Second, it depends on OpenAI not changing its API pricing strategy. If OpenAI suddenly increases API fees tenfold, “SmartWrite’s” costs will explode. It will either operate at a loss or raise prices, and if it raises prices, its users will flee.
Third, it depends on OpenAI not banning its API key. If OpenAI decides that “SmartWrite” is “freeloading” or encroaching on its market share, it can terminate its API key at any time, instantly paralyzing the company.
Do you see it now? “SmartWrite’s” core asset is a “liability” on OpenAI’s balance sheet.
OpenAI has to provide stable API services for these wrapper companies, offer technical support for their users, and “endorse” their business models. But the moment OpenAI decides these wrappers are no longer valuable, or are beginning to threaten its market position, it can “write off” this liability—just as a bank writes off a bad loan.
This is the point I keep stressing: when you build a villa on someone else’s foundation, the owner can tear it down at any moment, or even build an identical one.
II. Switching Costs: Why Your Users Were Never Your Users
Now, let’s go one level deeper and look at “SmartWrite’s” other “core asset”: 50,000 registered users and 3,000 paying customers.
My friend was most proud of this user data. “Look,” he said, “users are willing to pay, which proves my product has value. Isn’t that a moat?”
No. In the business world, having “users” and having a “moat” are two different things.
Let me illustrate with a classic example. In the 1990s, a company called Netscape made a web browser, Netscape Navigator, which at one point had a market share of over 90%. Hundreds of millions of people used it, far more than any “wrapper AI” today.
Then what did Microsoft do? It bundled Internet Explorer for free with Windows 95.
The result? Netscape’s market share plunged from 90% to single digits in just three years. In the end, Netscape was acquired by AOL for $4.2 billion—which sounds like a lot, but was a fraction of its peak valuation.
Why couldn’t hundreds of millions of users save Netscape?
Because the “switching cost” was zero.
What are switching costs? They are the price a user pays to switch from one product to another. This price can be monetary (e.g., losing a subscription), temporal (e.g., needing to learn new software), or data-related (e.g., the hassle of migrating large amounts of content).
What was Netscape’s switching cost? The answer is zero.
Because all browsers do the same thing. A user could open IE, type in a URL, and browse the web just the same. No new learning, no data migration, no extra cost. So, when Microsoft gave away IE for free, users switched in a heartbeat.
Now, let’s calculate the switching cost for “SmartWrite.”
What is the core need of a user of “SmartWrite’s” AI writing assistant? To “write good articles.” Why did they choose “SmartWrite”? Because it was “a little more convenient” than the official interface, or “a little cheaper.”
Now, suppose OpenAI’s official interface becomes just as convenient, or even more so, and the price is the same, or even free. What price would a user pay to switch to the official interface?
Step one: Open a browser and type chat.openai.com. Step two: Register an account (or log in with Google). Step three: Start using it.
The entire process takes less than two minutes. Moreover, the user doesn’t need to learn anything new, as the official interface’s logic is nearly identical to “SmartWrite’s.” They don’t need to migrate any data, because they weren’t storing any “proprietary data” on “SmartWrite” to begin with—their articles were saved on their own computers.
So, the switching cost for “SmartWrite” is also zero.
What does this mean? It means “SmartWrite’s” 3,000 paying users are not its “assets”; they are “temporary rentals.” The moment a better, cheaper, or more convenient alternative appears, these users will churn within 24 hours.
True user assets are those users who “can’t leave even if they want to.”
Take WeChat, for example. Want to switch from WeChat to another messaging app? The cost is enormous. All your friends, your social feed, your subscriptions, your payment history, your linked bank cards, your mini-programs are on WeChat. Switching is equivalent to replacing your entire digital life.
This is the “high switching cost” created by network effects. The more users WeChat has, the higher the switching cost, and the wider the moat.
But the users of a “wrapper AI” are like tenants in a rental apartment. They live there because the rent is cheap and the location is convenient. The moment a better apartment opens up next door, they pack their bags and leave.
This is the harshest truth of business: users with no switching costs are not your assets; they are your liabilities.
III. The Sunk Cost Fallacy: Why You Sink Deeper and Deeper
Now, let’s move to a deeper level: the psychological one.
My friend actually had a chance to “cut his losses” before his project collapsed. Two weeks before OpenAI announced its new model, there were already whispers in the industry. Some of his investors advised him to pivot quickly, or at least downsize the team and conserve cash.
But he didn’t listen. He said, “I’ve already invested millions, the product is taking shape, and users are growing. If I give up now, wasn’t it all for nothing?”
In economics, that statement has a specific name: the sunk cost fallacy.
What are sunk costs? They are costs that have already been incurred and cannot be recovered. They include the time, money, and energy you’ve invested, as well as your emotional attachment and expectations.
The greatest danger of sunk costs is not the costs themselves, but how they “hijack” your decision-making.
Let me share a personal experience.
In 2003, I was a vice president at a branch bank. A steel trading company applied to renew its loan. The owner, Mr. Liu, was one of the hardest-working people I’d ever met. His business was large, with several warehouses across the country.
But the steel market had already begun its downturn. I looked at his financial statements and saw that his inventory turnover days had ballooned from 60 to 180, and his accounts receivable were also rising sharply. According to the bank’s risk control standards, his company had already crossed a red line.
I advised Mr. Liu to cut production, clear inventory, and scale down. He wouldn’t listen. “I’ve been in this industry for twenty years,” he said. “My warehouses, my team, my sales channels—they are my life’s work. If I give up now, what am I?”
When he said that, his eyes had a look I would later see in many founders of non-performing assets—it wasn’t confidence; it was a refusal to accept defeat.
What happened next? The steel market continued to fall, Mr. Liu’s inventory depreciated, his receivables turned into bad debt, the bank called in its loan, and his capital chain broke. In the end, his company was liquidated, and he was left with a mountain of personal debt.
I went to see him. He was sitting in his empty warehouse and said one thing to me: “Actually, I knew this would happen all along. But I just couldn’t accept it.”
“Couldn’t accept it”—those three words are the most vivid illustration of the sunk cost fallacy.
In the case of “SmartWrite,” what had my friend invested? Several million in capital, three months of development time, a team of a dozen people, and all his dreams and expectations for an “AI startup.”
These were all sunk costs. They were spent and could never be recovered.
But the sunk cost fallacy led him to do one thing: use the money he had already spent to “extend the life” of a future that was doomed to fail.
He used his last bit of cash to pay API fees instead of severance for his employees. He spent his last bit of energy optimizing the UI instead of looking for a new direction. He pinned his last hope on “just holding on a little longer, maybe things will turn around.”
The result? No turnaround came. The money ran out.
This is my one piece of advice for all entrepreneurs and investors: in the world of business, the highest form of wisdom is not “perseverance,” but “knowing when to quit.”
There is a saying in economics: sunk costs are not costs. Because costs are the “price you have to pay in the future,” while sunk costs are the “price you have already paid.” They should not influence any of your future decisions.
But human weakness makes us hostages to sunk costs. Because we’ve already invested, we can’t bear to let go. Because we can’t let go, we sink deeper. Because we sink deeper, we end up completely buried.
IV. The Final Game of Beta and Alpha
Let’s return to “SmartWrite’s” balance sheet.
On the day OpenAI released its new model, what happened to this document?
Assets:
- Cash: All spent
- Intangible Assets (the code): From 20 million to 0—because the new model had all the features built-in.
- User Data: From 5 million to 0—because all users churned within a week.
- Total Assets: From 29 million to 0.
Liabilities:
- Accounts Payable: 500,000 in API fees, still had to be paid.
- Other Payables: 300,000.
- Net Assets: From 28.2 million to -800,000.
A star project valued at thirty million three months ago became an insolvent, bankrupt enterprise three months later.
This is the core concept I’ve been repeating: the Beta return of the wrapper startup is the platform’s Alpha return.
For the past two years, all “wrapper AI” companies have been growing. This wasn’t because you did anything right; it was because “the tide was rising.” This tide is the dividend of large model popularization, what economists call a “Beta return.”
And Beta returns are borrowed leverage. When the tide goes out, you see who’s been swimming naked.
The value that truly belongs to you, the part of you that cannot be replaced, is called an “Alpha return.”
What was “SmartWrite’s” Alpha? The answer: almost zero.
It had no proprietary data, no deep industry know-how, no network effects, and no brand loyalty. Its entire “value” was built on the premise that “OpenAI won’t do this itself.”
And OpenAI will absolutely do it itself.
This is not speculation; it is the fundamental logic of business. Any platform, after establishing a foothold, will extend upward to consume the downstream market for value-added services. Microsoft did it, Apple did it, Google did it, and OpenAI will do it too.
Because platforms also need to grow. And the fastest path to growth is to “eat someone else’s lunch.”
So the question is: in this era of “platform-eats-all,” can ordinary people still start businesses? Can they still invest? Are there still opportunities?
The answer is: yes.
But the opportunities are not in “wrappers.” They are in the corners that “platforms cannot cover.”
In those corners lie the true “antifragile assets” of the AGI era.
What are those assets?
Chapter Three: Cognitive Ascension: Re-evaluating Antifragile Assets in the AGI Era
I. From “Wrapping” to “Deep Diving”: What Constitutes a Real Moat?
We have dissected the balance sheet of the “wrapper AI,” seen through the nature of its “toxic assets,” and understood how “sunk costs” drag a founder into the abyss. But knowing “what is wrong” is not enough. We need to know “what is right.”
In the AGI era, under the shadow of the “platform-eats-all” paradigm, what kind of assets are truly “antifragile”? What kind of companies can survive, and even thrive, in the face of a crushing blow from evolutionary forces?
This question reminds me of a counter-intuitive case.
In 2007, Apple released the first iPhone. Developers worldwide were ecstatic—a new platform, a new ecosystem, a new opportunity.
But the first wave of “gold rushers” quickly encountered the exact same problem as today’s “wrapper AI” startups. They developed apps on the iOS platform: flashlights, calculators, weather widgets. These apps were popular, with high download counts, but in subsequent iOS updates, Apple “built-in” these features one by one.
Flashlight? Built into iOS 4. Calculator? Always there. Weather? Always there. The developers who only made “flashlight” apps lost their market overnight. But one company survived, and it thrived. Its name was WhatsApp.
Why? Because WhatsApp wasn’t building a “feature”; it was building a “network.” When Apple introduced iMessage in iOS 5, it theoretically “covered for free” WhatsApp’s functionality. But WhatsApp’s user base didn’t churn; it grew.
Why? Because WhatsApp users had formed a network. Your friends were on WhatsApp, your group chats were on WhatsApp, the photos, videos, and voice messages you shared were on WhatsApp. Want to switch to iMessage? You could, but the cost was telling every single one of your friends, “Stop contacting me on WhatsApp; find me on iMessage instead.”
That cost is the “switching cost.” And that switching cost is derived from “network effects.” Network effects are the first form of a moat. But they are not the only form. There is another kind of moat, deeper, thicker, and harder for a platform to “build in.” It is the “exclusive asset.”
What is an exclusive asset? It’s something the “platform can’t take, even if it wants to, and can’t copy, even if it tries.” It could be a dataset, an algorithm, a brand, a form of trust, or a deep-seated industry “know-how.” In the AGI era, amidst the ruins of collapsing “wrapper AI” companies, two types of truly “antifragile assets” are emerging.
The first is “exclusive, high-friction proprietary data.” The second is “deep user trust and industry know-how.”
II. Asset 1: Exclusive, “High-Friction Proprietary Data”
What is “high-friction proprietary data”? Let’s break it down.
“Proprietary data” is data that is unique to you, that others cannot access. It could be your user behavior data, your accumulated industry case studies, or your patents, copyrights, and trade secrets.
“High-friction” means the cost for others to acquire this data is extremely high. This cost can be legal (e.g., privacy regulations), technical (e.g., data silos), commercial (e.g., exclusive licenses), or historical (e.g., twenty years of accumulated experience).
The core characteristic of “high-friction proprietary data” is that large models cannot generate it out of thin air, and platform owners cannot easily replicate it.
Let me share a case I witnessed firsthand.
In 2015, I was still the general manager of the credit approval department at the head office. A company specializing in “medical imaging AI” came to us for a loan. Their technical team was strong, full of PhDs from top universities who had published papers at top conferences. Their product used AI to identify early-stage lung cancer in CT scans.
I asked them a question at the time: “Where does your data come from?” The founder replied, “We obtained 100,000 anonymized CT scan images from three top-tier hospitals, with the lesion locations annotated.”
I asked again, “Is this data exclusively yours?” The founder paused for a moment and said, “It’s not exclusive. The three hospitals also licensed it to two other companies.”
I said, “Then your core asset is not the AI model; it’s the data. And your data is not exclusive. Therefore, I’m sorry, but I cannot approve this loan.”
What happened next? In 2018, another, better-funded company paid a hefty sum to sign exclusive agreements with those three hospitals, obtaining the “exclusive rights” to all the data. The company that came to me, lacking data, couldn’t iterate its AI model and was eventually eliminated from the market.
This case teaches us that in data-driven industries, the “model” is the tip of the iceberg, while the “data” is the submerged foundation. The thicker the foundation, the more stable the iceberg. What does true “high-friction proprietary data” look like?
I’ve seen one truly impressive example. It was an AI company working in “industrial quality inspection.” Their clients were some of the world’s top semiconductor manufacturers.
Where did their data come from? From the clients’ production lines. Every wafer, every chip, generates a massive amount of “defect images” and “test data” during the manufacturing process. This data is the client’s core trade secret and would absolutely not be shared with a second company.
This creates a “high-friction” barrier. First, the data is “exclusive”—only this company can see it. Second, the data is “high-value”—it’s directly related to chip yield and cost. Third, the data is “dynamically updated”—as the production line runs, data is generated, and the AI model iterates.
This company’s business model is not “selling an AI model” but “evolving with the client.” The client can’t leave them, not because their technology is so superior, but because their “data flywheel” is deeply integrated with the client’s business processes, forming a symbiotic system where “I am in you, and you are in me.”
That is a true “antifragile asset.” No matter how powerful a large model becomes, can it generate these “production line secrets” out of thin air? No. No matter how formidable a platform owner is, can they bypass the client and directly obtain this “exclusive data”? No.
So, if you are an entrepreneur or an investor evaluating an AI project, don’t ask, “How good is its model?” Ask, “Where does its data come from?”
If the answer is “from public data,” then it’s a “wrapper”—because anyone can get public data. If the answer is “from an exclusively partnered client,” then it’s a “deep diver”—because it has dived into depths others cannot reach.
III. Asset 2: Deep “User Trust and Industry Know-How”
Now, let’s move to the second type of “antifragile asset.” This asset is deeper than data, more enduring than technology, and harder than any model.
It’s called “trust.”
Sound abstract? Let me make it concrete with a true story.
In 2018, I met the CEO of a company that makes enterprise risk management software, Mr. Zhou. His company wasn’t large, just a few dozen people with annual revenues of a few tens of millions. But his client renewal rate was consistently above 98%.
When I visited him, I asked, “Mr. Zhou, what is your core competency?”
He didn’t show me code or patents. He told me a story.
“Last year, a bank’s credit department approached us about implementing an AI risk management system. We competed against two other tech companies. One was founded by a tech guru from Silicon Valley, whose model had extremely high accuracy. The other was a team from a tech giant like BAT, whose product had an excellent user experience.”
“Who did the bank choose in the end?” I asked.
“They chose the tech guru’s company,” Mr. Zhou smiled. “But they abandoned it after just three months. Then they came back to us.”
“Why?”
“Because while the tech guru’s model was accurate, it couldn’t understand the ‘human element.’ A bank’s credit officer needs more than just a ‘loan’ or ‘no loan’ conclusion; they need to know ‘why.’ They need the AI to tell them where the risks lie, what the logic is, what evidence supports it. If the AI just says ‘risk too high, deny,’ the officer can’t explain that to the client.”
“So what’s different about your product?”
“In our product, we’ve embedded the ‘experience’ of the old risk management experts on our team, who have been in the field for decades, into decision trees and explanation templates. When our AI says ‘deny,’ it tells the officer: ‘Because this company’s inventory turnover has declined for three consecutive quarters, exceeding the industry average by 20%, there may be a risk of inventory backlog. It is recommended to further verify inventory details and aging structure.'”
“With this conclusion, the officer can communicate with the client and report to their superiors. They feel that this AI is not a ‘black box,’ but their ‘advisor.'”
Do you see it now? This company’s core asset is not the lines of code or the model, but the “decades of risk management experience”—the “know-how.”
And this “know-how” has two characteristics that large models cannot replicate.
First, it is “tacit knowledge.” What is tacit knowledge? It’s the things you can understand but cannot articulate. A risk officer with thirty years of experience can glance at a financial statement and immediately spot a problem. If you ask him, “How did you see that?” he can’t explain it. It’s just a “feeling.” This feeling is an “intuition” sedimented in his bones after reviewing tens of thousands of statements and falling into countless traps. Large models cannot learn this because it’s not in any textbook or public dataset; it exists only in human “experience.”
Second, it is an “asset of trust.” This company has served banks for over a decade, helping them avoid countless bad loans. The banks’ trust in it wasn’t built in a day; it was earned through one successful case after another, one avoided pitfall at a time. This trust is a “friend of time.” A large model can calculate a conclusion in a second, but it cannot win a user’s trust in a second.
This is the refraction of the “human prism” in business.
When the objective force of “evolution” (the iteration of large models) strikes with devastating power, companies with only “technology” but no “trust” are like rootless duckweed, swept away in a single wave. But companies with “deep industry know-how” and “user trust” are like great trees rooted in rock—the stronger the wind, the deeper the roots.
Because their customers are not buying a “function”; they are buying “peace of mind.”
IV. Future Projections: Who Survives in the “Platform-Eats-All” Era?
Now, let’s put these two types of “antifragile assets” together and see how they will evolve in future competition.
The first asset type: “Exclusive, high-friction proprietary data.” The owners of these assets are typically companies “deeply rooted in an industry.” They don’t compete with the platform; they “co-exist” with it. The platform provides general “cognitive ability,” while they provide specialized “industry data.” The platform is the engine; they are the steering wheel. Neither can function without the other. The moat for these companies will deepen over time. As data accumulates, models iterate, and client dependency grows. When a new competitor tries to enter the industry, it faces not a “technical barrier” but a “data barrier”—without ten years of accumulated data, your model will simply not be as accurate. And this “data barrier” cannot be bought with money.
The second asset type: “Deep user trust and industry know-how.” The owners of these assets are typically “human-machine collaboration” companies. They don’t try to replace humans with AI but to augment humans with AI. Their product is not an “unmanned system” but a “human-in-the-loop decision support system.” Their value lies not in “how fast it calculates” but in “how reassuring its calculations are.” The moat for these companies is “time” and “trust.” Every successful case is a brick; every renewed contract is a coat of paint. The bricks are laid higher and higher, the paint applied thicker and thicker, until it becomes a fortress that no one else can scale.
These two types of assets share a common characteristic: they are “un-build-in-able” by the platform.
A platform can build in a “flashlight” feature because a flashlight is generic. But a platform cannot build in “a specific hospital’s exclusive medical records” because the data is private. A platform can build in a “calculator” because calculation is generic. But a platform cannot build in “thirty years of a veteran risk officer’s experience” because experience is tacit, personal, and a source of “peace of mind.”
This is what I stated in the prologue: “The only two truly ‘antifragile assets’ are exclusive, ‘high-friction proprietary data’ and extremely deep user trust and industry ‘know-how’.”
Now, let’s return to my friend from the beginning of this article, the founder of “SmartWrite.”
His failure was not because he wasn’t hardworking or smart enough. It was because he chose the most “fragile” business model—building a villa on someone else’s foundation, using someone else’s leverage to run his business, and pretending he had a moat with zero-switching-cost users.
If he had chosen to build an “AI medical record analysis system for a specific hospital,” or an “AI case retrieval tool for a specific law firm,” or an “AI credit approval assistant for a specific bank” instead of a “wrapper AI writing assistant,” would the outcome have been different?
The answer is: he most likely would not have ended up at zero.
Because that “exclusive data” and “industry know-how” would have acted as an anchor, securing him firmly in a specific part of the value chain. No matter how powerful the platform became, it could not easily displace him. Because he was providing not a “function,” but “value.” Not “generic,” but “bespoke.” Not “replaceable,” but “irreplaceable.”
This is the “antifragile asset” that entrepreneurs and investors should be chasing in the AGI era.
Conclusion
Let me return to the initial question. My friend asked me on the phone, “The product was clearly great, user growth was fast, why did it all fall apart overnight? What did I do wrong?” Now, I can give him a complete answer. You didn’t do anything wrong. You just did the right thing in the wrong place, in the wrong way. You were right to “catch the trend.” You were wrong to “think the trend was everything.” You thought your core asset was your code and your tens of thousands of users. But in reality, your core asset was your “right to call OpenAI’s API”—and that is not an asset, but a “lease” that can be revoked at any time.
You stood on someone else’s foundation and thought you had built a castle. But when the landlord decided to reclaim the land, your castle was just a pile of rubble.
This is the most fundamental truth of the business world: when you do business on a platform, you are always borrowing someone else’s leverage. And borrowed leverage is always “fragile.” True “antifragility” can only come from “what is uniquely yours”—be it exclusive data or deep trust.
There’s a line from the Tao Te Ching that I particularly like, which can be interpreted as: “The greatest skill appears clumsy, yet its utility is unyielding.”
True wisdom often looks “clumsy”—it doesn’t chase trends, package concepts, or borrow others’ leverage. It is the patience to settle down and accumulate the “hard work”: accumulating data day after day for a decade, building trust drop by drop, and cultivating an industry inch by inch.
This “clumsiness” may seem slow, but its “utility” is “unyielding”—it cannot be easily replaced by a platform, eliminated by a cycle, or crushed by an evolutionary force.
My friend asked me one final question: “So what should I do now?”
I told him: “Forget ‘wrapping,’ go ‘deep diving.’ Go to the corners that large models cannot reach, acquire the data that others cannot obtain, and accumulate the industry experience that can only be earned through time and trust. Remember, what your user needs is not ‘AI,’ but ‘you, using AI to solve their unique problem’.”
This is not a comforting platitude. It is the only answer I am certain of after three decades of observing the business world.
This answer applies to you as well. If you are an entrepreneur, an investor, or simply an ordinary person seeking direction in this era, please remember this:
In the wave of AGI, do not be the one building castles on the sand. Be the navigator who can still set sail after the tide has gone out.
And a navigator’s ship is not built on “someone else’s foundation.” It is forged from “one’s own unique assets.”
