Earning $330,000 a month for 18 months: Unveiling the business model of "AI shelling": Is it a flash in the pan or a hidden goldmine?

In the public opinion arena of the technology industry, one can often hear a somewhat contemptuous comment: "Isn't this just a wrapped AI?"

For developers who are racking their brains to build new things, this comment sounds particularly harsh. It's like a bucket of cold water poured directly on the newly ignited flame of innovation. The subtext behind this comment is extremely damaging: It means a lack of core technology, means simply building blocks on the foundation of giants, and means vulnerability that could be toppled at any time.

However, the counter - arguments are equally loud and their logic is so solid that it's irrefutable.

Aravind Srinivas, the CEO of Perplexity, once said bluntly: "Everything in the world is a wrapper. OpenAI wraps around NVIDIA's computing power and Azure's cloud services; Netflix wraps around AWS's infrastructure; even Salesforce, with a market value of up to $320 billion, is ultimately just an advanced wrapper for the Oracle database."

His words do hit the nail on the head. But before delving into this war of words about "definition", it's necessary to clarify what exactly this "AI Wrapper" that the public talks about is.

In short, this is often a product form labeled as "opportunistic". Developers don't train large - scale foundational models themselves. Instead, they directly call ready - made APIs (such as OpenAI's interfaces), and then overlay a thin user interface on top to provide a specific function. The development process usually involves very little complex underlying technology research.

The most typical example is the applications that allow users to "converse with PDFs". In the early days of ChatGPT, when the official didn't offer a direct document - handling function, these applications quickly became popular. Users only need to upload an obscure paper or report, and the AI can quickly generate a summary and answer related questions. This demand is real and urgent, and the solution seems very effective.

However, the problem lies precisely in the fact that this "effectiveness" comes too easily.

The debate about AI wrappers often overlooks a more grand proposition. Wrapping itself is not a sin. The real dividing line is: Is the constructed product a "function" that can be easily replaced at any time, or a "product" that can stand firm?

Some wrapped applications are destined to be just passing fads. Once the giants wake up and integrate these functions into their own ecological suites, they will wither away quickly. But there are also some applications that can take root and grow into towering trees in the cracks between giants. The label of "wrapper" actually masks the real core that deserves attention: Is it a function or a product? How broad is the market segment it's in?

A short - lived function or an enduring product?

Let's first examine the early example of applications that allow users to converse with PDFs.

This type of tool solves a very specific and narrow problem: having trouble understanding documents and seeking AI assistance. It doesn't create new documents, doesn't involve editing existing content, and usually doesn't record users' reading habits to optimize the subsequent experience. It's essentially a one - time tool that users use and then leave, lacking stickiness.

Strictly speaking, it can only be defined as a "capability", not a complete end - to - end solution. It should exist as a button in a document reader or a plug - in in a flagship office software.

This is where the danger lies. When foundational model builders like OpenAI, Anthropic, or Google decide to directly integrate this "capability" into their systems natively, those independent wrapped tools instantly lose their foundation for existence. This is a typical "functional" fate - easy to be replicated, lacking a business closed - loop, and having no moat or long - term defense.

However, the business world is never short of exceptions. Even if they will eventually be swallowed by the platforms, during the time window when the giants are too busy to act, these functional applications can still create amazing wealth and become interesting independent business cases.

The data is enough to show why developers are flocking to this: PDF.ai once had a monthly recurring revenue (MRR) of $500,000; PhotoAI had $77,000; Chatbase had about $70,000; InteriorAI had $53,000. And Jenni AI is even more astonishing. In just 18 months, its monthly recurring revenue soared from $2,000 to $333,000.

This wealth is indeed tempting, but this business model is more like picking up gold nuggets during a gold rush rather than mining a gold mine. Once the surface - level gold nuggets are picked up, the business will end.

Surviving beside the giants' beds

Some wrapped applications are developed thick enough to even evolve into real products, targeting a huge market. At this time, what they face is no longer the mockery of "being a wrapper", but real survival threats.

There are two major obstacles here: One is the control of model access rights, and the other is the monopoly of distribution channels.

Let's first look at model access rights. The code assistant field is the most typical battlefield.

Tools like Cursor have actually taken the concept of "wrapper" to the extreme. It doesn't just simply call APIs. Instead, it deeply integrates AI into the integrated development environment (IDE). It can read the entire codebase, edit files, generate code, roll back changes, and even run coding agents. To some extent, it has completely reshaped the developer experience in the AI era.

This market can support huge imagination. Among the top five technology giants with the highest market value globally, software developers account for about 30% of the total number of employees. Even if development tools can only slightly improve productivity, the value released will be in the billions of dollars. This makes this field a must - fight area for model builders and giants with distribution channels.

However, the Achilles' heel of Cursor - like tools is that they rely heavily on external sources. They have to rely on the model interfaces of OpenAI, Anthropic, and Google to survive until open - source models or self - developed models can match the quality of cutting - edge closed - source models.

The developer forums are filled with complaints from paying users about "rate limits". In actual development projects, developers often encounter the situation where their Claude quotas are used up. Even if users prefer Cursor's interface design and interaction logic, in order to advance the project, they have to switch to the tools provided by Claude official (and pay high fees to avoid the limits). The interface may be better, but access to the model often plays a decisive role.

This dependence is not only about quotas but also about strategic survival. Sam Altman, the CEO of OpenAI, once put forward a famous view: The correct strategy should assume that the models will continue to improve.

"There are two strategies for building an AI startup. One is to assume that the models won't get better; the other is to assume that the models will continue to evolve at the same pace. It seems that 95% of the people in the world should bet on the latter, but many startups are built based on the former. When we do our jobs well and the model capabilities are upgraded, those companies that bet on the wrong side will be mercilessly crushed."

This kind of crushing is all - round. From knowledge tutoring, healthcare, creative expression, shopping, to writing assistants, legal assistants, and every huge market segment, as long as there is profit to be made, model manufacturers have the motivation to enter the market themselves.

The brutal strangulation of traffic and channels

In addition to the direct competition from model manufacturers, distribution channels are the second sword hanging over the heads of startups. Even if the model builders hold back for the time being, startups still have to face another severe question: Can they build a large enough user base before the giants with existing products and large - scale distribution channels add AI functions?

This is the modern echo of the classic business case of Microsoft Teams vs. Slack. The challenge is to build a loyal customer base before Microsoft embeds Copilot into Excel or PowerPoint, before Google weaves Gemini into Workspace, or before Adobe integrates AI into its creative suite.

An independent AI - wrapped tool for spreadsheets or presentations not only has to overcome functional homogenization but also has to compete against the giants' bundling sales advantages, distribution channel advantages, and users' high switching costs.

This kind of channel competition from giants also applies to other large markets such as healthcare and law. In these fields, regulatory frictions and control over the "System of Record" often favor established enterprises like Epic Systems. For example, a clinical note generator that can't write data into the electronic health record (EHR) will eventually hit Epic's distribution barrier.

Of course, there are always exceptions in business competition.

First, speed itself is a weapon. Tools like Cursor, although lacking control over core dependencies (model access), have such amazing growth rates that they become very attractive acquisition targets. Windsurf got a $2.4 billion acquisition license deal from Google; Gamma reached $50 million in revenue in about a year; Lovable reached $50 million in revenue in just six months; Galileo AI was acquired by Google. Rapid market share often gives enterprises a chance to exit before being crushed.

Second, excellent execution can occasionally overcome structural advantages. Midjourney, with its excellent product quality, convinced Meta to use its services, even though Meta has a much larger budget and distribution capabilities.

Finally, foundational models may abandon some markets due to risk aversion. The regulatory burdens in the healthcare and legal fields, or the potential reputational damage from AI companions and adult content, leave opportunities for operators willing to face extreme regulatory scrutiny or controversy. The opportunities are still huge, but competition (or acquisition) may knock on the door at any time.

The glimmer of light in the cracks: The gold mine for independent developers

Not every market gap will attract the covetousness of model builders or technology giants. In the long - tail part of the business ecosystem, there are a large number of work requirements that are too small for the scale of venture capital but can support a business worth millions of dollars.

These niche markets are a paradise for founders who are careful with their money and pursue lean operations.

Imagine AI applications for astrology, manifestation, or dream interpretation. A dream - interpretation AI that allows users to record their dreams every morning, generate AI videos based on the dreams, maintain a kind of dream diary, and reveal certain psychological patterns over time solves a complete work closed - loop.

Users can, of course, tell their dreams to ChatGPT, and it can even save the history. But a dedicated application can build specific fields (such as recurring characters, places, things, themes, etc.) to capture dreams in a structured way and integrate with sleep - tracking data in a way that a general chatbot can't.

Such a niche market is small enough to avoid the strategic radar of large models but large enough to sustain a highly profitable independent business.

As model builders and traditional giants enter the market, the existing players in this "wrapper" debate face strategic choices. Enterprises that can survive the competition storm of model builders often have two key characteristics.

First, even if they don't own the model, they must have the dominant right over the results.

Applications that have been embedded in users' workflows (such as Gmail/calendar, Sheets, EHR/EMR, Figma) don't need to cultivate new user habits. Building these platforms from scratch is much more difficult than adding AI capabilities to existing platforms. When these applications directly send operations to a proprietary record - keeping system (controlling calendar events, submitting claims, creating purchase orders, etc.), the action of "completion" occurs within the giants' environment. At this time, AI is just an input to the existing workflow, not a replacement.

Second, successful survivors will build proprietary data from customer usage.

Every correction by users, every handling of edge cases, every approval, and all human feedback will be transformed into training data, continuously refining the product over time - this is a valuable asset that cutting - edge general models can't touch.

Although Cursor is not a traditional giant and relies on external models, it plans to compete by capturing developers' behavior patterns. As its CEO Michael Truell said in an interview: Capturing user data and feeding it back to the product is the real sustainable advantage. This dynamic is similar to the search wars in the late 1990s and early 2000s: Only through users' clicks and interaction behaviors can we truly understand users' intentions and optimize the product.

Looking back at this debate about AI wrappers, both critics and defenders have their reasons, but they are also one - sided.

The critics are right. Many wrapped applications without defensiveness will eventually disappear as platform functions swallow them up. The defenders are also right. Every successful software company is essentially "wrapping" certain underlying technologies.

But the real insight often lies between the two.

Even if a new application starts as a "wrapper", as long as it can fit into the actual work scenarios of users, write data into a proprietary record - keeping system, build proprietary data and continuously learn and evolve from usage, or occupy the distribution channel before the giants bundle the function, it will have long - lasting vitality.

More importantly, those "wrapped" products that can quickly iterate and continuously deliver functions to solve users' pain points when competition approaches will be extremely difficult to defeat. It is these characteristics that draw the line between a short - lived "function" and an enduring "product".

Original article link

https://www.wreflection.com/p/wrapping-my-head-around-ai-wrappers

This article is from the WeChat official account "CSDN", author: Nowfal, translated by Wang Qilong, published by 36Kr with authorization.