Many people worry that storage will turn out like the fiber optics of 2000—because of excessive infrastructure buildout and overcapacity, leading to blowups. But I think storage and fiber optics still have a few fundamental differences at the physical layer:

1) The “linearity” of Maxwell’s equations means optical paths can be multiplexed across many frequencies and won’t interfere with each other; therefore, in theory, the bandwidth of a single optical fiber can be infinitely large.
2) Even at nearby frequencies, fiber optics can further multiply communication efficiency using purely mathematical methods such as modulation and encoding.
But the problem with HBM right now is absolutely not “efficiency”—it’s capacity. A Rubin GPU with 500GB of VRAM needs twice as many HBM chips as a 250GB Blackwell GPU would, and that’s an immutable mathematical fact.
In addition, for every extra bit in VRAM, you need another piece of physical space—so you waste a bit more wafer. This is unlike what exists in the virtual world, where as long as energy conservation is satisfied, you can create photons arbitrarily; simply adding more photons is “easy.”
Therefore, likening storage to fiber optics isn’t appropriate. Even if there is “overcapacity” in the end, its difficulty and timeline will be far stronger than—and come later than—the internet infrastructure bubble of 2000.
Even if researchers develop storage-specific “byte reuse” or context compression algorithms to reduce demand, that would again run into Jevons’ paradox.
The higher the storage efficiency, the more people use it; with demand unchanged, total demand still ends up increasing.
So hard!