Nvidia Says AI Is a Five-Layer Cake. Waydev Just Took Two Slices.

Nvidia CEO Jensen Huang has a simple way to explain the AI economy: a five-layer cake of energy, chips, infrastructure, models and applications. It is usually discussed by investors and policymakers. It is just as useful for engineering leaders, because software development is one of the applications pulling hardest on every layer beneath it. Here is what the cake means for the AI SDLC, and why we built our own model for it: Waydev SDLC 30B.

AI is not an app. It is an industrial stack.

In his essay “AI Is a 5-Layer Cake” and in Nvidia’s video explaining it, Huang argues that AI is essential infrastructure, like electricity and the internet, not a clever app or a single model. Looked at industrially, it resolves into five layers:

Every agent PR pulls on the whole cake

This is the part engineering leaders rarely think about. When a coding agent runs for an hour, re-reads your repository, calls tools and opens a pull request, that single task travels down the entire stack: a model generating tokens, an AI factory serving them, chips converting power into computation, and an energy grid behind it all. Multiply that by every agent on every team, every day.

That is why AI spend in engineering behaves less like a software license and more like a utility bill. It scales with usage. It depends on which model you pick. And because energy and compute sit at the bottom of the cake, the cost of intelligence will always matter, no matter how fast prices fall per token. The question for engineering leaders is not just “which model is smartest?” It is “which model is right-sized for the work, and where should it run?”

Every engineering org has its own five layers

Zoom into a single engineering organization and the same stack appears, at company scale. Each layer is a decision you are making, whether you realize it or not.

Notice the extra layer on top. Nvidia’s cake ends at applications because, for the industry, applications are where value is created. Inside a company, value is only real once it is measured: what the agents actually shipped, what it cost, and whether it held up in production. Without that sixth layer, you have a very expensive cake and no idea whether anyone enjoyed it.

Why open, vertical models matter for the SDLC

Huang makes a point that is easy to skip: most of the world’s models are free, and open models are what let researchers, startups, enterprises and whole nations take part in advanced AI. He also notes that when open models reach the frontier, they activate demand across every layer of the stack.

For enterprise engineering, open models solve a specific problem. The most valuable context about how your organization builds software, your code, your tickets, your incidents, your standards, is also the context you are least willing to send to a public API. An open base model, specialized for the SDLC and deployed inside your own environment, lets you use that context at the model layer and the infrastructure layer you control.

Our slice of the cake: a model built for the SDLC

That is why Waydev now works at two layers of the cake: applications and models. We built Waydev SDLC 30B, a vertical model for the software development lifecycle, post-trained on an open base, and designed to understand how engineering organizations actually work: delivery, review, quality, adoption and the agents doing more of the work every month.

Mapped back onto Huang’s cake, Waydev SDLC 30B lets an engineering organization own more of its own stack:

Measuring your cake from the top down

Because every application pulls on every layer beneath it, the right metrics start with outcomes and trace back to cost. These are the ones we recommend.

Every company will build its own slice

Huang says every company will use AI and every nation will build it. For engineering organizations, that means more than buying seats on someone else’s models. It means deciding which layers to rent and which to own, right-sizing the model to the work, keeping the most valuable context in-house, and measuring whether the whole cake is producing what it should. The infrastructure buildout is the bottom of the cake. The AI SDLC is where it gets eaten.