For decades, the computing world has relied on DRAM as the backbone of memory systems, but a perfect storm of AI demand and manufacturing constraints is pushing its limits. Now, a visionary figure in tech has sketched out a radical alternative: a fiber-optic loop that could replace traditional RAM by continuously recycling data at near-light speed.
John Carmack, best known for pioneering 3D graphics in games like Doom, has long been a thinker ahead of his time. In a recent post, he outlined how fiber-optic cables—already used in telecommunications—could be repurposed to create a memory system that doesn’t degrade over time. Unlike DRAM, which loses data when power is cut, this system would maintain data integrity through light pulses circulating in a closed loop.
The idea hinges on a simple but transformative concept: instead of storing data in volatile memory chips, encode it as light signals in fiber, then read and rewrite it repeatedly. This could eliminate the need for DRAM entirely in certain AI workloads, where memory access patterns are predictable and repetitive.
Why this matters
Today’s AI models guzzle memory, and DRAM production can’t keep up. Carmack’s proposal isn’t just theoretical—it taps into existing fiber technology, which is already deployed globally. If realized, it could reduce hardware costs, improve energy efficiency, and sidestep the looming DRAM shortage that threatens to stall AI progress.
But challenges remain. Fiber-based memory would require custom hardware to interface with processors, and the technology would need to prove its reliability at scale. Still, Carmack’s suggestion underscores a broader truth: the next era of computing may not rely on incremental upgrades but on entirely new architectures.
What’s next?
While Carmack’s idea is still in the conceptual stage, it reflects a growing recognition in the industry that traditional memory solutions are unsustainable. Researchers and hardware manufacturers are already exploring alternatives like storage-class memory and in-memory computing. If fiber loops gain traction, they could become a cornerstone of next-generation AI systems—offering a glimpse of a future where data moves not just faster, but fundamentally differently.
