AI is a five-layer cake

On stage at the World Economic Forum in Davos in January 2026, talking with BlackRock's Larry Fink, NVIDIA's CEO Jensen Huang explained the AI boom with a picture: AI is a five-layer cake. He repeated it in a March 2026 essay titled "AI Is a 5-Layer Cake", and it's become one of the most quoted ways of explaining the AI industry.

From the bottom up, the five layers are:

1. Energy. The electricity that everything else runs on.
2. Chips. The processors that do AI's math.
3. Infrastructure. The buildings, cooling and networks that turn chips into what Huang calls AI factories.
4. Models. The trained AI systems themselves: language models, but also models for biology, chemistry and robots.
5. Applications. The products people and businesses actually use, where the money is made.

Each layer rests on the one below. A model can't run without infrastructure, which can't work without chips, which can't work without power. Scroll to build the cake, then send one ordinary question down through it.

Why a cake, and not a cloud

Most people meet AI as a chat box on a screen. It feels like pure software, weightless and living "in the cloud." Huang's point is that this is the icing. Underneath it are power plants, chip factories, and buildings full of machines, all of which are slow and expensive to build.

He goes further. In his essay he argues that AI is now essential infrastructure, like electricity and the internet, and that building all five layers is "the largest infrastructure buildout in human history." Whether or not you buy the superlative, the framing explains a lot of headlines that otherwise seem unrelated:

• A tech company signs a deal for a nuclear power plant. (Layer 1)
• A chipmaker announces new factories. (Layer 2)
• A cloud company plans a data center the size of a town. (Layer 3)
• A lab releases a new model that tops the benchmarks. (Layer 4)
• A hospital rolls out an AI tool that drafts patient notes. (Layer 5)

They're all the same story, one layer at a time.

One question, five layers

Take the question in the animation: "Plan me a weekend in Kyoto."

• Applications. An app, maybe a chatbot or a travel site, receives your question.
• Models. It passes your question to a model, which writes the answer one token (roughly a word piece) at a time.
• Infrastructure. The model runs in a data center: a building packed with computers, wired together and cooled.
• Chips. Inside those computers, GPUs do billions of multiplications for every token.
• Energy. All of that math is paid for in electricity, used at the moment you ask.

That last point surprises people. The answer to your question didn't exist anywhere until you asked. It was made on the spot, and making it used power. That idea is where the whole cake starts, and it's the subject of the next chapter.