Wrapper

Many AI products don’t do the computing themselves. They send the user’s input to a large provider like OpenAI, Google, or Anthropic. That’s where the actual language model runs, i.e., the program that generates new text from text. The answer comes back and is displayed in the company’s own app. Such a product is called a wrapper, from the English word for packaging or shell. The company builds the packaging, while renting the expensive inner workings.

Why the word is often meant as an accusation

In investor circles, “just a wrapper” is harsh criticism. The idea behind it: whoever owns nothing of their own has nothing to defend. A competitor can rent the same interface and build a similar product within a few weeks. Even more dangerous is the provider itself. If OpenAI builds a feature directly into ChatGPT, the small company’s business model becomes obsolete overnight. In the industry, this risk is half-jokingly called being “killed by the next update.”

Then there’s the margin problem. For every request, the wrapper pays money to the AI provider. These costs grow with every user and can hardly be reduced. A classic software company earns almost pure profit from its ten-thousandth customer. A wrapper loses money on every customer if it sells too cheaply.

Nevertheless, the verdict is often too sweeping. Some wrappers are very successful because they possess something the model provider doesn’t: access to a specific professional group, vetted specialist data, or an integration with software that companies already use. The coding assistant Cursor, for example, uses third-party models and was still valued at billions.

What happens between input and the foreign model

Technically, the connection runs through an interface, known in industry jargon as an API. This is a fixed address on the internet to which you send text and from which you receive text back. The provider charges per token, i.e., per word fragment. About four characters of text correspond to one token. The wrapper pays both for the question sent and for the answer generated.

Between the user and the interface lies the wrapper’s actual work. It usually attaches an invisible piece of additional text to the user’s input, the system prompt. This specifies what role the model should take on and what it must not do. Often, the company’s own documents are also sent along so that the answer matches the company’s data. This procedure is called RAG, meaning answers generated using looked-up sources.

A good wrapper and a bad wrapper differ exactly at this point. A bad one passes the question along virtually unchanged. A good one checks the answer, calls a second model if needed, caches frequent queries, and switches providers if another one is cheaper or better. This independence from any single model is the most important protection against price increases.

Wrappers in everyday life and in financial news

Most AI tools people use without a second thought are wrappers. Writing aids in the browser, chatbots on shop websites, summary features in note-taking apps, AI assistants in video conferences. You rarely recognize them, because the foreign model is never mentioned anywhere. One clue is the phrase “powered by” in the fine print or in the privacy notices.

In financial news, the word usually comes up in the context of valuations and funding rounds. Analysts then ask how much genuine substance lies behind a product. The term is equally important when it comes to stock price movements: when a large provider announces new features, the valuations of smaller providers who sold exactly that sometimes drop.

A common misconception is that wrappers are technically undemanding. Calling the interface takes only a few minutes; turning that into a reliable product takes years. Moreover, the boundary is fluid. Anyone who trains their own small model for partial tasks and only uses third-party models for the rest is no longer considered a pure wrapper.