Application Layer

Software is usually thought of in layers that build on one another. At the very bottom sit computers and cables, at the very top sits what a human being sees in front of them. This topmost level is called the Application Layer. This is where the actual application runs: the app on the phone, the website in the browser, the program on the laptop. Everything below works invisibly in the background, making sure data arrives and computing power is available. The term has two meanings that should be kept apart: an older one from networking technology and a newer one from the AI industry.

Why investors argue over the topmost layer

In the AI economy, the Application Layer has become a contested term. At the bottom sit the companies that build large language models, i.e. programs that understand and generate text. These include OpenAI, Google, or Anthropic. Above them sit thousands of smaller companies that buy these models and turn them into products. A program that summarizes doctors' letters does not train its own model. It sends the text to someone else’s model and builds the interface, security, and billing around it.

The central question is: where does the money end up in the end? Training models costs billions, and only a few companies can afford it. Building an application on top of one is comparatively cheap. That’s why so many companies spring up in the Application Layer so quickly. And that is precisely their problem: whoever can start easily also faces easy competition.

Skeptics dismissively call such companies a “thin wrapper” around someone else’s model. Their objection: if the model provider builds the same feature in itself, the product becomes worthless. Supporters counter that the value doesn’t lie in the model but in access to customers and in the data of a specific field. Both sides have examples that seem to prove them right. The question remains unresolved so far.

What happens between user input and the model

A product at this layer doesn’t simply pass an input along. It reshapes it first. Typically, a company’s own documents are searched and appended to the query so the model has access to concrete facts. On top of that come fixed instructions the user never sees, for example regarding the desired tone. Only this whole package is then sent to the model.

On the way back, the same thing happens again. The answer is checked, formatted, and sometimes verified by a second model pass. Meanwhile, the software keeps track of how much computing power was used, since that determines the bill. Part of the work simply consists of keeping several models interchangeable. If one provider gets too expensive, you switch.

The older meaning from networking technology works on the same principle. There, a model with seven layers describes how data travels across the internet. Layer 7 is the application layer, where protocols like HTTP for websites or SMTP for emails operate. Below that, other layers take care of addresses and cables. Here too the same rule applies: the upper layer doesn’t need to know how the lower one does its job.

How to recognize application companies in everyday life

Most AI tools people know are Application Layer products. A translation service in the browser, a coding assistant in the editor, software that pre-sorts job applications. All of them rely on models they didn’t build themselves. A good test is to ask whether the company trains its own model. If the answer is no, it operates at this layer.

In business news, the term usually comes up in connection with funding rounds. It’s then said that money is increasingly flowing “from the model to the application.” This means investors are betting less on new model builders and more on companies with paying customers. A common misconception here is to consider the Application Layer technically undemanding. Data protection, reliable outputs, and billing are difficult tasks – they’re just less visible than a new model.