The Stack Is Real. The Source Matters.

NVIDIA published a piece this week laying out a "five-layer cake" model of AI infrastructure: energy → chips → infrastructure → models → applications. Worth reading. Worth reading carefully.

The thing it gets genuinely right is the physical substrate correction. Most professionals — and most AI commentary — treat AI as a software phenomenon. Models, apps, prompts. The piece makes the case that intelligence generated in real time has a physical cost: electrons, heat, cooling, land. Every token produced consumes power. That's not a detail. It corrects a real blind spot, and it's the kind of grounding that's been missing from boardroom AI conversations.

The thing it doesn't earn is in the section on electricity and the internet. The analogy is used to imply that building the infrastructure layer is where the strategic importance lies — and, by extension, where the value accrues. But that's not what the electricity era showed us. The companies that built the power grid didn't become the dominant firms of the electric age. The infrastructure commoditised. Value migrated to the applications built on top of it. The piece doesn't address this, and it should, because it's the obvious objection.

The reason it doesn't address it is visible in the byline. This was published by NVIDIA's official account. NVIDIA sells chips — layer two in their own model. The framing that makes infrastructure essential and trillion-dollar is precisely the framing that maximises demand for what NVIDIA manufactures. That's not a reason to dismiss the framework. It's a reason to notice that one of the layers in the cake was described by the company that profits most if you believe in it.

The five-layer model: useful. Extract it. The investment thesis quietly inside it: not theirs to draw without declaring the interest.