Carmack's Wild Idea: Using Fiber Lines as AI Memory Cache

Okay so I saw this and had to share because it's classic Carmack thinking outside the box. The legendary programmer behind Doom and Quake just dropped a fascinating idea about using long fiber optic lines as a sort of L2 cache for streaming AI data. And honestly? It's not as crazy as it sounds.

Here's the deal. Current AI systems are absolutely starving for memory bandwidth. We're talking about models that need to shuffle terabytes of data around just to answer your questions about recipe substitutions. DRAM is fast but expensive as hell, and when you need hundreds of gigabytes or more, the costs get stupid real quick. So Carmack's thinking, why not use the speed of light traveling through fiber as a weird kind of memory buffer?

The basic concept works like this: you send data down a really long fiber optic cable, and while it's bouncing around in there at light speed, it's essentially "stored" in transit. Need that data? Time it right and grab it when it comes back around. It's like using the highway itself as a parking lot. Wild, right?

Now before you think this is pure science fiction, remember that engineers have actually done similar tricks before. Back in the day, they used mercury delay lines for memory in early computers. Same principle, different medium. The question is whether it makes any practical sense in 2026 when we've got other options.

My take? It's probably not replacing your GPU's VRAM anytime soon, but for specific AI workloads where you need massive sequential data streams, this could actually work. The latency would be predictable, the cost per bit stored would be way lower than DRAM, and fiber is getting cheaper every year. Classic Carmack move - solving tomorrow's problems with today's infrastructure in ways nobody else would think of.