Reading the cake

You've been through the layers one at a time. The real value of the framework comes from looking at the cake as a whole, because the layers move at different speeds. A new app can launch in a week and a new model in months, but new chip factories take years and new power plants even longer.

When demand pours in from the top, the slowest layer sets the pace for everything above it.

The order in the animation is only an illustration, but the pattern is real. In the last few years the constraint has shifted between chips, advanced memory, grid connections and skilled labor, sometimes all at once.

Jobs: electricians, not just engineers

Because the bottom three layers are physical, Huang argues the AI boom is also a boom in skilled trades. At Davos he listed electricians, plumbers, steelworkers, construction workers and network technicians as people the buildout needs. It's an unusual thing for a tech CEO to emphasize, and it follows directly from the cake: you can't write software to lay a transmission line.

Is it a bubble?

The obvious question about trillions of dollars of investment is whether it will pay off. The cake gives you a way to think about it.

• Huang's case: it isn't a bubble, because the demand is real and still outrunning supply. At Davos he pointed out that renting GPUs is hard, and that rental prices were rising even for chips two generations old. In a bubble, he argued, old equipment would be cheap and idle. BlackRock's Larry Fink, on stage with him, went further and said the world is underinvesting.
• The skeptic's case: the bottom three layers are being built for demand that is expected, and building takes years. If applications don't earn enough to pay for all of it, the layers below end up with more capacity than anyone needs, just as the telecom industry laid far more fiber in the late 1990s than the internet used at the time.

Both can be partly true. The cake helps you see where to look: the top layer has to earn enough to pay for everything beneath it. Watch application revenue, not just spending on chips and data centers.

Every country's question

The cake also explains why governments care. If AI becomes essential infrastructure, like electricity, then a country without its own layers depends on someone else's. At Davos, Huang urged countries to "develop your AI," building local models and infrastructure on top of open models, and argued that AI could help close the technology gap between rich and developing countries. Energy and skilled workers, the bottom of the cake, become national advantages.

Who's serving the cake?

One last habit. The five-layer cake is a genuinely useful map, but it was drawn by the CEO of the company at the center of layer 2, whose products fill layer 3. A map that shows AI as a huge, physical, long-term buildout is also a map that says "keep buying chips." That doesn't make it wrong. It means you should use it as a way to organize the facts, not as proof of any forecast.

Your cheat sheet

Next time you read an AI headline, ask which layer is this about?

• Power deals, grid connections, nuclear plants: Layer 1, energy.
• New GPUs, chip plants, memory shortages, export controls: Layer 2, chips.
• Data-center campuses, cooling, "AI factories": Layer 3, infrastructure.
• New model releases, open vs closed, benchmark records: Layer 4, models.
• Products, customers, revenue, agents at work: Layer 5, applications.

Then ask the follow-up: what does this do to the layers above and below? That second question is the whole point of thinking in layers.