The six layers of the AI stack

Transcript

The different layers of an AI startup explained. What is not every AI company is building the same thing. Some build the model, some build the infrastructure, some build tools on top. And yes, some are basically rappers. Let's break the stack down. At the bottom you have the compute layer, GPUs, data centers, networking, power cooling, and video core weave AWS, Google Cloud, and Microsoft live heavily around this layer. Without compute, open AI, anthropic, Google DeepMind, XAI, Meta, DeepSeek. They take huge amounts of data and compute, train the models that expose them through APIs or products. The next layer is infrastructure for AI applications. Vector databases, observability, evaluation, model routing, inference, guardrails, and agent tooling. These companies are not training the giant models. They help developers use them reliably in production. Then you have the application layer where most AI startups live, coding tools, video generation, customer support, legal AI, sales, healthcare design. The model is not the product by itself. The startup combines it with workflow, UI data, distribution, integrations, and domain knowledge. Now this is where people start using the word wrapper. A wrapper is a thin interface around somebody else's model API. Take a prompt, send it to GPT, display the answer, and that is the entire product. The problem is obvious. If open AI or anthropic adds that feature tomorrow, what is left? But not every company using someone else's model is a wrapper. Stripe uses banks, but stripe is not a bank wrapper. The question is, what are you adding on top for proprietary data? A unique workflow, deep integrations, a strong agent harness, a network effect, a brand, distribution, or years of domain knowledge. That is what turns an API call into a real company. Then there is vertical AI, go in extremely deep into one industry, law firms, construction, insurance, radiology, logistics. Now the mode is not the model. It's understanding the workflow better than anyone, having industry-specific data, or integrating into systems that are painful for outsiders, and finally, the distribution layer. This one gets underestimated. The best AI product doesn't matter if nobody discovers it. Some companies win because they already own the customer relationship. Microsoft has office, Google has search and workspace. Adobe has creative cloud. Salesforce has enterprise customers, so they can put AI directly inside workflows people already use. So the stack looks like this. Compute at the bottom, foundation models, infrastructure, applications, vertical products, and distribution deciding who reaches the customer. The higher you go, the easier it is to start a company. But also the easier to copy, the lower you go, the harder and more expensive it is to enter, but the more defensible it becomes. So when someone says that startup is just a rapper, because using someone else's model does not make you a rapper, having nothing valuable beyond the API call does.

Transcribed automatically from the video, so small errors are possible.