AI's demand for memory is pretty difficult to really get across because there's a lot of complex factors, but whatever you can imagine is the demand, it's higher than that.

You can look at pre and post AI to get a slightly better picture, but then the numbers don't look terrible and so the demand isn't as clear.

2020-2023 primary customers were smartphones, laptops, PC. Data centers were eating about 32% of the global market for RAM. Monolithic DDR4/DDR5 was the main product and High Bandwidth Memory (HBM) was about 8%. Total memory set being sold was like 16GB kits to 64GB kits, obviously server kits were going out, just the majority was those mostly for PCs.

2025 hits and the primary customer is AI Data Centers. To put it at scale, you have literally everything that uses memory (and I mean literally every fucking thing on this planet) and AI Data Centers. And the break between those two bins are 30% and 70%. AI data centers are consuming more than twice the memory of literally everything combined that uses RAM that isn't an AI data center.

The primary RAM being made now is HBM, which is way more complex. 23% of all the wafers that will be used to make integrated circuits will be HBM RAM. And by wafers, I mean all the chips that will be made this year, lock, stock, and barrel. If you randomly picked up a wafer out of a fab you have a almost 1 in 4 chance to pick up RAM. And finally the average kit going out is 1TB to 2TB kits, which is a lot more than the old 16GB to 64GB kits.

Now I mention HBM because it eats more wafer, that's because unlike DDR4/5 RAM, HBM RAM is a three-dimensional circuit. 12 to 16 layers of silicon is stacked on top of each other. So HBM consumes about 300% more silicon than other memory (not every layer is one-to-one in size). So you don't just have one fab making chips, you have several fabs making the layers.

The next thing is that building fabs is complex. I hate trying to explain the complexity, but you can't do it overnight. Usually you have to build these things over the course of five years. Just to give you some idea of how technical the construction is. If you had a road within 500 feet of a chip fabricator sitting on a regular concrete floor, the car driving on the road would create enough shakiness in the Earth to cause the chip fabricator to bounce around too much. So when they build the place that have to literally isolate the small earth quakes humans walking around inside the place cause. This requires very complex floor building. And this is just the floor, not to mention how clean the place has to be kept, isolated as much as possible atmosphere, literally specific sections are under vacuum. It's massively complex to build ONE of these.

The complexity comes with a price tag. Average cost to build one memory making factory is around $15B to $20B. It's serious cash, but even if you have 5 years and $20B, there's a specific bottleneck. ASML. ASML is the only company on the entire face of the Earth that makes the chip making machines. They've indicated that if you ordered a machine today, you can expect it roughly 1½ to 2 years from now. That's how many people have put in an order for the machines to make memory.

So all that aside, there's one more bottleneck. HBM has to be stacked in layers, there are very few people on this planet that can do that, and they have years long backlog. And even then, most times the stacking fails. About 30% to 50% of all HBM is trashed because the layers fell apart. And the people who stack are entirely different people than the layer makers. But they're the same people that take that DDR4/5 wafer and cap it into that little black rectangle you see on your sticks of memory. So they have pretty much ~100% of their employees doing nothing but stacking layers of memory together.

Another thing is economic prioritization, HBM is about 500% more than DDR4/5's price tag per GB. A fab producing wafers of DDR4/5 is making about $x.xx. A fab producing a couple of the layers for HBM is making about 500% × $x.xx on average (it's complicated because of the layers), even with the stacking issues. And the profit margin on HBM is 70% versus DDR4/5 before AI which was fingernail thin. SK Hynix was actually taking a loss on production of DDR5 at about -1.6%. So going from -1.6% to 70% profit has created a crowding out effect. Not to mention that since there was a bit of a bleeding out period after COVID, some literally stopped making RAM. Which has made the issue even worse.

The last thing before I run out of characters is the AI growth. AI needs about 300% more memory every ten months. That's how fast these models are growing. That's caused a panic buying and also caused a rushing to fulfill. The industry is losing it's collective mind because the money to be made is big and so lots think it can't last and trying to get their cut before the gravy train derails.