Every file in The Memory Atlas — the 3D map of the AI memory supply chain — as plain, fast pages. Same facts, same journey, no WebGL required. Read it here on a phone; open any file in 3D when you're at a desk.
Data as of 13 September 2026
Welcome. Every layer, bottleneck and company in the AI memory chain is a node on this graph — the gold row is the money. Glide around with WASD or the arrow keys, click anything to read it, or keep hitting Next ▶ (or press N) and the atlas walks you from zero to expert, stop by stop.
AI is bottlenecked by memory, not math. To write each word, a GPU re-reads the entire model from memory — and sits idle ~1,200× waiting on it. That one fact is what this whole map is about.
One glass tank = one GPU's memory. The blue block is the model; every chat adds a green slab that grows with the conversation. Drag the sliders below — when the tank overflows, you're buying whole extra GPUs.
This is the thing filling those tanks. The gold cube is the word being written; the beams look back at every stored key/value pair. Toggle the cache off below to see the waste it prevents.
Fast, tiny and costly at the top; vast, slow and cheap at the bottom. Slab width = capacity, particle speed = bandwidth, ◆ = a genuine supply chokepoint. The next stops walk every layer.
If you remember one stop, make it this one. Just three firms make HBM and it is sold out through 2027 — gated from below by TSMC's CoWoS packaging, ASML's EUV monopoly and Ajinomoto's 95% grip on ABF. The bottleneck is packaging, not the chip. Tap "Who makes it" in the panel.
New in HBM4: the bottom die of every stack is now a logic chip made in a foundry. SK Hynix uses TSMC; Samsung makes its own. It decided who won Rubin's allocation — and gave TSMC a toll on memory.
HBM is the famous part; the price of ordinary DRAM is what moves SK Hynix, Samsung and Micron's shares. In 2026 it spiked — because HBM is eating the same wafers. This is the tape to watch.
When a model outgrows one GPU's memory, CXL lets many servers borrow from one shared pool. Toggle Expansion / Pooling / Sharing below and watch the memory move.
NVLink vs UALink vs CXL — whoever wins decides which chipmakers get the design wins for years. Astera Labs is the pure-play that wins whichever way it breaks.
Copper dies after ~1 metre and runs hot; light travels far and carries ~10× the data. At rack scale, optics is taking over — watch both lanes run.
Silicon cannot make light. Every optical link starts with an indium-phosphide laser — fewer than five firms make them at volume, and NVIDIA locked up ~$4B of Lumentum & Coherent supply through 2027.
Now the money. These are the names that structurally define the map — chokepoint monopolies, scale players and pure-plays — and exactly why each one matters.
The biggest risk on the map, in both directions: China builds most of the optical modules and is closing on commodity DRAM — but has no EUV, no HBM3E, no laser, no DSP. And it holds the minerals the West's lasers need.
You've walked the whole chain: the wall → the cache → HBM → the chokepoints → CXL → photonics → the money. Click any node to go deeper — and click a company to see every layer it touches. The floor is yours.