AI Industry Structure, Layer 4: AI Models

The real “brain” that generates an answer when you type a question into ChatGPT is the AI model. This layer, Layer 4 of the AI industry structure, is the substance that actually fills the shell of Layer 5’s applications. Its precise technical name is a foundation model. In this article, we’ll cover how this model gets built, why it costs so much, how it differs from open-source models, and what investors should watch for in this layer.

Revisiting AI Models

There’s a reason it’s called a “foundation.” Once you’ve trained this one model, you can layer functions like translation, summarization, coding, and image generation on top of it with no additional training at all, or only a tiny amount. In the past, you had to build a separate translation model and a separate summarization model; now, this one AI model handles all of those roles.

So how does this “brain” actually get built?

How a Model Gets Built

1Pre-training

This is the stage where the model reads internet-scale volumes of text — books, web pages, papers — and absorbs the patterns of language and knowledge wholesale. Thousands to tens of thousands of GPUs run non-stop for months during this stage. Most of the total training cost (typically 90%+) happens right here.

2Fine-Tuning

A model that’s only been through pre-training honestly borders on incoherent rambling. It just predicts the next word plausibly — it doesn’t answer the way a person actually wants. So it gets trained one more time on question-and-answer example data written by humans, teaching it “this is what a good answer looks like.”

3RLHF (Reinforcement Learning From Human Feedback)

When people rank the model’s various answers by saying “this answer is better,” the model learns that preference and gradually starts answering in ways people like more. This is also the stage responsible for polishing ChatGPT to answer in a particularly “friendly and safe” way. Because it requires ongoing human involvement, the cost is relatively low, but it demands a lot of time and manpower.

Why It Costs So Much

You’ve probably heard that training a single top-tier AI model can cost hundreds of millions to billions of dollars. That’s because three major costs stack on top of each other.

That’s why this layer is close to a “capital game” that not just anyone can enter. In fact, the number of companies in the world that own a truly top-tier model can be counted on one hand.

How It Differs From Open-Source Models

GPT, Claude, and Gemini are closed models — they don’t publish the model’s blueprint (its weights). On the other hand, there are also open-source models like Meta’s Llama, which publish their weights so anyone can download and run them directly.

Meta isn’t giving Llama away out of charity. Instead of charging directly, it’s a strategy to dominate the ecosystem itself by getting developers worldwide to build services on top of its model. A large share of the vertical apps in Layer 5 (Applications) also choose open-source models to save money while running them directly on their own servers.

What Investors Should Watch For

The most common question about this layer is “do these companies actually have a real moat?” Competitors frequently catch up in model performance within a few months, and open-source models are closing the gap fast too. That’s why investors should also weigh these factors.

In the end, that means “who can keep pouring in massive amounts of money, and hold onto customers in proportion to what they’ve spent” matters more than “who builds the smartest model.”

Closing Thoughts

AI models look like the most glamorous part of the AI industry structure, but they’re also by far the most expensive layer. And to cover this massive training cost, companies ultimately have to rent enormous amounts of infrastructure and cloud (Layer 3) underneath.

In the next article, we’ll take a close look at Layer 3, infrastructure and cloud. We’ll cover what business AWS, Azure, and Google Cloud actually run, and why even securing cooling and power has become an industry-wide issue.