Wrappers Ate the Entry Point

Companies that should have died are growing larger each quarter

Silicon Valley's conclusion in 2024 was categorical. AI wrappers—companies that simply call OpenAI or Anthropic models and attach a UI on the front end—would soon die. The logic was that model providers would absorb those functions or the next model would erase their raison d'être. Among venture capitalists, "thin wrapper" was practically an insult.

But the landscape now in 2026 is the exact opposite. Code generation tool Cursor has reached annual revenue in the hundreds of millions of dollars without building its own model. Video editing's Descript, customer support's Intercom Fin, and legal document's Harvey all borrow others' models. Those who should have disappeared by conventional wisdom are growing larger each quarter.

At this point, the conventional wisdom should be questioned. What was wrong wasn't the companies, but the direction of reading the value chain.

What the "thin" diagnosis saw and missed

The criticism that wrappers are thin accurately identified one thing. The function of model calling itself is easy to replicate. Anyone can call the same GPT with a single API key. There is virtually no technical moat.

However, this diagnosis missed something bigger. The fact that where users begin their work—that entry point—cannot be replicated. If the window developers open to write code is Cursor, then regardless of the model, that traffic passes through Cursor. Models can be swapped out, but entry points cannot. Because users' fingers are already there.

A second structure overlaps here. Workflow lock-in. The legal documents accumulated in Harvey, the codebase context learned by Cursor, the consultation history built up in Intercom—these remain with the wrapper, not the model. Even when model companies release smarter models, where that data and work context will go is ultimately the company holding the entry point. Beneath the seemingly thin surface, data gravity was accumulating.

So the real question is not "is the technology thin" but "where do margins come from." Model companies are driven into per-token cost competition, while those holding entry points charge monthly subscriptions per user. Token prices fall but subscription fees stay the same. That difference becomes the wrapper's margin entirely. An asymmetry where the cheaper models get, the better the entry point's profitability becomes.

There are counterarguments. Why can't model companies build entry points themselves? Indeed, OpenAI is pulling coding and search into ChatGPT. But model companies cannot own workflows for every vertical domain. Legal, medical, video, semiconductor design each demand different data permissions, regulations, and work practices. Model companies are busy defending horizontal platforms, and vertical entry points are too numerous and fragmented. Rather than gaps big tech can't eat, wrapper territory consists of gaps with little reason to eat.

Capital is moving from models to entry points

The flow of money shows this realignment. In 2023, venture capital concentrated on foundation models, with tens of billions of dollars going into model training. From 2025, the center of gravity shifted. Model training became so capital-intensive it became a game for a few big tech companies, with Microsoft, Google, Amazon, and Nvidia effectively oligopolizing the infrastructure layer.

So venture capital flows to the upper layer—applications and entry points. Models become commodities, and who captures users above them becomes the new battlefield. Infrastructure domination is a finished game, and entry point domination is a game just beginning. Capital bets on games that have started, not games already won.

This structure in one sentence: Big tech takes infrastructure, startups take entry points, and the model layer in between sees margins thinned by price wars. Both ends of the value chain are crushing the middle.

Which layer do Korean companies stand on

At this point, we must mark the coordinates of Busan and Seoul. Most Korean companies are currently either coveting the model layer in this chain or remaining positioned as infrastructure customers.

Naver and Kakao build their own models, but are in the middle of the global token price war. More dangerous are the countless Korean SaaS companies that merely add a chatbot saying "we've added AI too." This is not owning an entry point but lending one function to someone else's entry point. The window where users begin work is still a foreign tool.

Let's ask the question again. In this competition, are Korean companies suppliers, customers, or standard setters? In semiconductors and displays, they are suppliers. That is strong. But in software entry points, they are mostly customers. They buy foreign models, code with foreign development tools, and upload to foreign clouds. In the chain's profitable segments, Korean software companies' seats are vacant.

The opportunity lies exactly where conventional wisdom didn't look. Not models, but owning entry points for specific industries tailored to Korean language, Korean regulations, and Korean work practices. Korean healthcare, Korean legal, Korean manufacturing floor workflows are vertical gaps that foreign big tech has little reason to eat. If data gravity accumulates there, whatever the model changes to, traffic passes through that entry point.

The cost of watching while Silicon Valley moves capital to the entry point layer is not simple opportunity loss. If foreign platforms occupy entry points first, Korean companies' next cost structure becomes tied to subscription fees they set. Even as token prices fall, the money we pay does not decrease. That difference flows forever as the margin of whoever holds the entry point. The moment we read why wrappers didn't fail as someone else's story, we are already sitting in the customer seat of that chain.