Nobody Talks About the Hardware

These days I’ve found myself reading about hardware shortage and how RAM, disk, and NAND manufacturers have all their production for 2026 already sold to AI/Data center companies.

It had me thinking about the strange position compute finds itself in right now.

GPUs weren’t designed for AI. They were designed for graphics as in pushing pixels, rendering triangles. It turned out that the math for doing that at scale overlaps just enough with neural network math. The entire current AI boom runs on repurposed hardware.

AI compute today looks like early computing in some ways (expensive, power-hungry, room-sized) but with a key difference: the general-purpose phase is already behind us. GPUs served that role for over a decade, but now the industry is moving fast toward specialized silicon: Google’s TPUs, Meta’s MTIA, Amazon’s Trainium, Microsoft’s Maia. Even OpenAI is building its own chips. Every hyperscaler has concluded that off-the-shelf isn’t good enough anymore.

And then there are bets and promises. Wafer-scale chips that fit an entire compute cluster on a single piece of silicon. Designs that try to mimic how brains actually work. Optical computing, replacing electrons with light. And the long promised quantum computing. These aren’t products yet, they’re hypotheses about what comes next.

The pattern in computing history is clear: when you hit a physical wall something different emerges. We’re hitting those walls now. Current GPU scaling is running into the limits of what silicon can physically do and we’re solving it by just throwing more hardware and building bigger computing centers. Vacuum tubes hit the same kind of wall. Too hot, too big, too power-hungry, and the answer wasn’t more tubes. It was the transistor, which made everything that followed possible.

Yet almost nobody is talking about it. The conversation is all AI and agents, as if the hardware underneath were a solved problem. It isn’t.