The State of AI Infrastructure
The AI infrastructure race looks like a technology competition. It is actually a physical supply chain competition — one where the decisive variable is not model quality, not capital raised, nor engineering talent, but the gap between when you believed and when you committed.
Where are we in the long cycle?
Before anything else, the right temporal context matters. AI is not a product cycle or an application wave — it is a supercycle catalyst technology, the kind that triggers a 30–50 year restructuring of how the global economy operates. The prior examples are electricity and the internet. What distinguishes a supercycle catalyst is that it reshapes the infrastructure layer of the economy itself, not just the application layer on top.
Every supercycle has phases. Infrastructure accumulation comes first — the period when the physical substrate gets built. Railroads, power grids, fiber networks. This is where we are now. The application layer, the economic diffusion, the moment when the technology becomes invisible — those come later. The most common error in analysis is treating AI as if we are already in the application phase. We are not. The infrastructure is still being laid.
The practical implication is straightforward: value accrues to picks-and-shovels operators during this phase, not to application builders.
NVIDIA's earnings of the magnitude it does from a relatively modest slice of TSMC’s cumulative CapEx is the canonical expression of this dynamic. The GPUs, the HBM memory, and the EUV lithography tools are the steel rails of this era. The trains — the applications that run on top — have not yet arrived at scale.
