Layer 5: Applications

The top of the cake is the part you touch: applications. A chatbot is one, but so is a tool that screens drug candidates, a copilot that drafts contracts, a robotaxi, a warehouse robot or a coding agent. Huang's essay is blunt about why this layer matters most: it's where the economic value is created. Every layer below is a cost until an application turns it into something people will pay for.

Demand flows downhill

Here's the most useful idea in the whole framework. Value is created at the top, but demand flows down. Every time an application gets more users, or starts doing longer and harder tasks, it pulls on every layer underneath.

Walk it through once:

• A legal copilot doubles its customers. Now it needs twice as many tokens from its model.
• The model provider needs more capacity, so it rents or builds more AI factory space.
• The factory operator orders more chips.
• More chips running flat out need more power, which means new grid connections and new power contracts.

That chain explains a lot of AI news. When app usage jumps, you hear about it weeks later as chip orders, and years later as power-plant deals.

Why applications are getting hungrier

It isn't just more people using AI. Each use is getting bigger:

• Reasoning models think before they answer, which can mean many times more tokens for one reply.
• Agents don't answer once. They work in a loop, calling tools and checking results, sometimes for minutes or hours on a single task. (The hands-on Build an AI Agent course shows that loop.)
• Physical AI, like robots and self-driving cars, runs models continuously, many times a second.

All three multiply the tokens per task, and so they multiply the demand on every layer below.

Does AI take jobs, or make them?

Huang's answer, at Davos and in his essay, is that the application layer mostly changes jobs rather than removing them. His favorite example is radiology: AI got good at reading scans years ago, yet demand for radiologists has kept growing, because the AI handles routine reading and the doctors spend more time on judgment and patients. His summary line is "Productivity creates capacity. Capacity creates growth."

That's one view, and a hopeful one. Other economists and researchers are less sure, especially for jobs made mostly of tasks that AI does well. What the cake does make clear is that the buildout underneath the apps creates a lot of physical work, which is where the last chapter picks up.