Every three months, the market receives a fresh reminder of just how early we still are in the AI infrastructure cycle. What began as a narrow story has steadily matured into something far broader, and each quarter of data reinforces that the momentum behind this buildout is not fading — it is spreading.
From GPUs to the Full Factory
The cleanest way to frame the evolution is to think of it in waves. The first wave, which defined the opening couple of years of the AI boom, was overwhelmingly about GPUs. They were the scarce resource, the headline story, and the chokepoint that every hyperscaler was racing to secure. That phase is not over, but it is no longer the whole picture.
The new phase is about the entire factory built around those GPUs. This encompasses the foundries that manufacture the chips, the advanced packaging that assembles them, the networking fabric that links them together, the memory systems that feed them, the CPUs that coordinate workloads, and all the supporting layers in between. The surface area of the opportunity has widened significantly, and with that widening, the bottlenecks themselves are shifting from one layer of the stack to another.
Durability Through Breadth
This rotation of bottlenecks is actually one of the strongest signals that the cycle has real staying power. If the story were only about a handful of dominant names, you would expect the pressure points to remain static. Instead, the fact that constraints keep migrating — from GPUs to packaging, to memory, to networking — is evidence that demand is genuinely broad-based and that the number of beneficiaries is no longer a shortlist. It is now dozens of companies across multiple layers of the ecosystem, and that breadth is exactly what gives the cycle its durability.
The Agent Era and the CPU Renaissance
AI itself is also graduating at a rapid pace. This year is shaping up to be defined by agents — autonomous systems that plan, reason, and execute tasks across software environments. That shift has important knock-on effects for the hardware stack. Because agentic workloads place heavier demands on orchestration, coordination, and general-purpose compute, CPUs are becoming interesting again. After years of being overshadowed by accelerators, they are re-emerging as a focal point of the infrastructure story.
Memory is poised to become a similarly massive component. As models grow, context windows expand, and agents maintain richer state, the volume and bandwidth of memory required scales dramatically. Add to this the many different elements and layers within the CPU ecosystem itself, and the investable universe continues to expand rather than contract.
Demand Still Outrunning Supply
The broader takeaway is that the cycle is nowhere close to cooling off. Recent updates from the companies sitting at the most strategic points in the supply chain — the lithography leaders and the leading-edge foundries — have again underscored that we are still early and that demand continues to exceed supply. When the most capacity-constrained links in the chain are signaling tightness rather than slack, it is difficult to argue that the buildout has peaked. If anything, the evidence points in the opposite direction: the foundation is still being poured, and the structures built on top of it are only beginning to take shape.