Kinetic Random Access Memory stores data as continuous light circulating in fiber — not as charge held in silicon. A new physical substrate for AI inference.
ARC Institute of Knowware is building a memory substrate that does not store data — it keeps it in motion. Modern AI has hit a wall that is not a compute problem: inference is memory-bound, and silicon memory is bound by charge, refresh, and the fab. K‑RAM steps outside that constraint entirely. By holding information as light circulating through a coil of single-mode fiber, we relocate the memory wall instead of optimizing within it — built on settled physics, commodity glass, and a sovereign supply chain.
Data enters as light and stays as light — circulating, compressed, and regenerated without ever returning to the electronic domain. Six stages, one continuous loop.
Information arrives as light. Model weights are written onto the substrate across many wavelengths at once through dense wavelength-division multiplexing — hundreds of independent channels sharing a single strand of glass. From the instant data enters, it never converts back to an electronic signal.
The light enters a coil of single-mode fiber and simply keeps moving, lap after lap, at the speed of light in glass. There is no cell to charge and nothing to refresh. Distance becomes capacity: the longer the loop, the more data it holds. Memory here is a property of motion, not of matter at rest.
Before it circulates, each weight matrix passes through a layered compression stack. A frequency transform exposes its underlying structure; that structure is fitted to a continuous generating function and then folded into a compact ternary network. We do not store the data so much as store the recipe that reproduces it.
On demand, the generating network regenerates the exact layer that was folded into it, sealed bit-perfect against the original. Reconstruction is lossless by construction — the substrate returns precisely what was committed, on the loop’s next pass, exactly when the model asks for it.
Kilometers of circulating light only remain legible if the glass holds perfectly still. A thermally controlled vessel keeps the fiber in phase, so a signal that began its journey arrives intact however many laps later it is read. Coherence is what turns light in motion into reliable memory.
The whole system — vessel, thermal control, and optical switch — lands in a standard rack and speaks to the cluster over a single fiber. A memory-starved deployment becomes a compute-bound one. No new silicon, no fab queue, no architecture rewrite. Glass and light, dropped in beside the GPUs already there.
K‑RAM is not a new material or a new transistor. It is a set of techniques applied to physics that has been understood for half a century — light, glass, and the mathematics of folding structure into almost nothing.
A weight matrix is folded into a smooth cubic generating function rather than kept as a dense grid of values. The function is tiny; what it can unfold back into is not. This is how a frontier-scale layer collapses into something a loop of light can carry.
The generating networks are quantized to three states — {-1, 0, +1}. Base-three is the most information-dense way to encode structure with discrete symbols, and it leaves the regenerating network almost weightless while keeping reconstruction exact.
Frontier model weights are not random — they are constitutionally regular, built from structure that repeats at depth. We measure that regularity and exploit it, storing the rule a layer obeys instead of every number the rule produces.
Light circulating through kilometers of fiber is exquisitely sensitive to the temperature of the glass. Precision thermal control holds the substrate in phase, turning a delicate optical effect into a memory you can trust across years of continuous operation.
A single fiber carries hundreds of independent channels at once, each on its own color of light. Capacity scales by adding wavelengths, not by shrinking features — bandwidth grows along an axis Moore’s Law never touched.
The speed of light in glass is a constant, not a trade secret. You cannot engineer around it, iterate past it, or acquire the company that found it. The substrate is commodity fiber, made the same way since 1970. The advantage is knowing what to do with settled physics.
When memory is light in glass instead of charge in silicon, the consequences are not incremental. They are structural — and they all arrive together.
Storage scales with the length of the loop, so it grows by adding glass — not by buying the most advanced chips on the market or waiting on a fab. The entire weight set of a frontier model can live permanently in working memory, with nothing to shard and nothing to swap.
Light in a waveguide is a near-lossless way to move and keep information. There is no array of cells to power and refresh billions of times a second. The substrate spends energy to stay coherent — not to remember — which is a different and far smaller bill.
There is no fab dependency, no export-controlled process, no supply chain routed through three countries. The substrate is commodity optical fiber and the medium is light. Built in Calgary, Alberta, K‑RAM is memory infrastructure a nation can actually own.
For operators relocating the inference memory wall, and for partners building the substrate underneath the next decade of AI.
Get in Touch →