Every Memory Type Explained in 11 Minutes

[00:00:00]Two chips, same silicon, same factory, same wafer. One went into a GPU in a data center. The other was supposed to end up in your computer. Between September and December of last year, a single 16 gigabit memory chip, the building block inside every stick of RAM on Earth, went from $6.84 to $27.20, nearly four times the price in a single quarter. Nobody redesigned it. Nothing about making it got harder.

[00:00:24]And it didn't stop there. By the middle of this year, that same chip was pushing past $60. So, the easy answer is inflation or a shortage. It's neither.

[00:00:34]Here's the thing. Your computer doesn't have one kind of memory. It has seven.

[00:00:38]Seven layers stacked in a pyramid from a chip inside your processor that answers in picoseconds all the way down to a technology from the 1950s that's still holding half the internet. And right now, a bidding war over one single floor of that pyramid is quietly repricing everything you own. So, let's climb all seven, fastest to slowest, every floor.

[00:01:00]And by the end, you'll know exactly which floor got bought and exactly why you got the bill. Level one, top of the pyramid, the registers. These live inside the CPU core, not near it, inside it. When your processor adds two numbers together, those numbers are sitting in registers. Access time? Around 300 picoseconds. That's 0.0003 0.03 seconds. In that window, light travels about 10 cm. So, your register is faster than light crossing your desk.

[00:01:30]Sit with that for a second. And here's what nobody tells you. A modern CPU has almost none of them, a few hundred kilobytes total, not gigabytes, kilobytes. Why so few? Because at this speed, every single bit is monstrously expensive in silicon area, in power, in heat. So, the registers end up being the smallest, fastest, most valuable real estate in the whole machine. And they're the part nobody ever thinks about, the forgotten floor, which raises the obvious question, if registers are this fast, why not just build everything out of them? Hold that thought. It's basically the whole video. Level two, SRAM, static RAM. This is your cache, L1, L2, L3. The waiting room right outside the CPU. L1 answers in about a nanosecond. L3, maybe 10 to 40. Still absurdly fast. And this is where we finally answer the register question, because SRAM shows you exactly why fast costs so much. To store one single bit in SRAM, you need six transistors. Six, wired into a tiny circuit that holds its value as long as it has power. That's what static means. It just sits there, stable, no upkeep. Six transistors per bit is brutally expensive. That's the whole reason your CPU has kilobytes of registers, a few megabytes of cache, and not a gigabyte. You literally can't afford it. The die would be the size of a dinner plate and cost more than your car. Now, cache is a rabbit hole all on its own. How the CPU guesses what you'll need next, and what happens when it guesses wrong. That's its own video coming soon. For now, just hold on to the number. Six transistors, one bit.

[00:03:05]Because on the very next floor, that number drops to one, and that one change breaks everything open. Level three, DRAM. This is RAM, the sticks you push into your motherboard. DDR5. And this is the floor the whole video is really about. Remember six transistors for a bit of SRAM? DRAM stores a bit with one transistor and one tiny capacitor.

[00:03:26]That's it. One switch, one little bucket of charge, one bit, almost no silicon.

[00:03:31]Which is why you can afford gigabytes of it instead of megabytes. But there's a catch, and honestly, it's a beautiful one. That little capacitor leaks. The charge drains out in milliseconds. So the memory controller has to keep stopping to rewrite every single bit, thousands of times a second, just to keep your RAM from forgetting itself.

[00:03:50]That constant refresh is why DRAM is slower than SRAM. 50 to 100 nanoseconds instead of one. You trade speed for density. Cheaper, bigger, a little slower. And that's the deal on every single floor of this pyramid. And this is the floor that just caught fire. That same 16 gigabit chip went from $6.84 in September to $27.20 by December.

[00:04:11]That's a 2 GB chip, so call it about 1360 per GB, the number you'll watch climb on the pyramid. And contract prices for PC memory, the real ones, not the spot market, didn't just double.

[00:04:22]They kept climbing quarter after quarter. Here's the strange part.

[00:04:26]Nobody's buying more RAM sticks than usual. So, who's eating all the DRAM?

[00:04:29]The answer is the next floor down. And it's not even inside your computer.

[00:04:33]Level four, HBM, high bandwidth memory.

[00:04:36]And this is the villain of the story, not because it's evil, but because it's hungry. HBM is just DRAM. Same one transistor, one capacitor bit. But instead of laying the chips out flat on a stick, you stack them. Eight, 12, 16 dies piled straight up, wired down through the silicon, and bolted millimeters away from a GPU. Why bother?

[00:04:55]Bandwidth. A normal RAM stick moves data down a road a few lanes wide. HBM moves it down a thousand lane superhighway, terabytes per second. And an AI accelerator training a model is starving for exactly that. The GPU can do the math no problem. The bottleneck is feeding it data fast enough. HBM is the feeding tube. And here's the murder weapon. HBM gets made in the same factories, on the same wafers, as your PC's RAM. One HBM stack eats roughly three to four times the wafer space of a normal DDR5 chip, and with worse yields.

[00:05:27]So, every time Samsung or SK Hynix or Micron flips a production line over to HBM for AI, they are quietly deleting three or four ordinary memory chips.

[00:05:37]Chips that would have gone into laptops, phones, your desktop. The AI boom didn't buy your RAM off a shelf. It bought the factory time before your RAM was ever born. That's why level three exploded because level four ate the assembly line. Quick one, if breakdowns like this are useful to you, you can support the channel on Patreon or by becoming a member. That's it. Back to the pyramid.

[00:05:58]Level five. And we just crossed the biggest cliff in the entire pyramid. I want you to actually feel this one.

[00:06:04]Everything above this line is volatile.

[00:06:06]Cut the power and it's gone. Registers, cache, DRAM, HBM, all of it forgets the instant the electricity stops. Level five is where memory finally learns to remember without power. This is NAND flash, your SSD. But look at what it costs you. Look at the latency. We jumped from 100 nanoseconds to 100 microseconds. That's 1,000 times slower.

[00:06:28]Persistence has a price and the price is speed. NAND stores a bit by trapping electrons in an insulated little pocket, a floating gate. No power needed. The electrons just sit there in the dark.

[00:06:39]But every time you write, you blast electrons through an insulating layer and that layer physically wears out.

[00:06:45]NAND has a hard limit on write cycles.

[00:06:47]Your SSD is dying, slowly, a little bit every single time you hit save and it gets sneakier.

[00:06:53]To cut cost, modern drives cram more bits into each cell, QLC, four bits per cell. Denser, cheaper and slower and less durable. So the manufacturers hide a small, fast, single-bit region, an SLC cache that pretends to be quick while it quietly shuffles your data off into the slow, dense cells behind your back.

[00:07:13]Cheaper, bigger, slower and a little bit of a liar. The pattern holds. Oh, and one more thing. An SSD is closer to 15-20 cents a gigabyte now and climbing for the exact same reason as level three. Keep going down. Level six, the hard drive, the spinning disk and everyone will tell you it's dead. It's not. It's thriving, just not on your desk. It moved into the data center and it did it for one reason, dollars per gigabyte. An SSD is 15-20 cents a gigabyte now. A hard drive is closer to 1 and 1/2, and when you're storing exabytes, that gap isn't small. That gap is billions of dollars, and this thing is gorgeous physically. A platter spinning at 7,200 RPM, a read head flying 2 nm above the surface. Scale that up proportionally, that's a jumbo jet cruising at the height of a few sheets of paper at supersonic speed, and it never once touches the ground. To read your data, the drive literally waits for the right spot to rotate under the head, which is exactly why it's slow, 5 to 10 ms. Compared to your CPU register, that's not a thousand times slower. It's not a million, it's tens of millions of times slower. If a register access is a single heartbeat, a disk seek is a 2-year wait at that scale.

[00:08:26]Slower, cheaper, bigger. Every floor, same trade. And there's still one floor below this one, the oldest technology in the whole building, and it's winning.

[00:08:35]Level seven, bottom of the pyramid, and this is the plot twist because it's the oldest thing here, magnetic tape, technology from the 1950s. You'd assume tape is a museum piece. It isn't. It runs the modern internet's basement. LTO tape, those cartridges hold the cold data, the backups, the archives, the photos you uploaded in 2014 and never opened again. And the hyperscalers, the exact same companies buying up all that HBM, are also the biggest buyers of tape on the planet. Google, AWS, Azure. Their cheapest storage tier, very often it's a robot arm loading a tape cartridge. Why?

[00:09:11]Follow the rule one last time. Tape is the slowest thing in the building. To read a file, a robot physically goes and fetches the cartridge, loads it, and winds the reel to the right spot.

[00:09:22]Seconds to minutes, glacial, but it's well under a cent per gigabyte with compression, closer to half a cent. It burns zero power just sitting on a shelf, and it lasts 30 years. So, look at the extremes. The fastest memory in your machine and the slowest are separated by a factor of a billion in speed and just about the same factor in price. That is not a coincidence. That's the entire law of the pyramid. And now you can see all seven floors of it at once. Here's the whole pyramid on one screen. Read it top to bottom and one rule falls right out of every row. Each level only exists because the one above it is too expensive and the one below it is too slow. That's it. That's the whole design. Speed, cost, size. Pick two and you get a floor. This pyramid has been stable for decades. Engineers treated these trade-offs like the laws of physics, fixed, permanent, done. And then in 2026, something reached in and grabbed the middle. AI didn't want the top of the pyramid, registers are too tiny to matter. And it didn't want the bottom, tape is too slow to train a model on. It wanted the middle floors, DRAM and HBM. The exact levels every other device on Earth also leans on. So it bought them, all of them. And when you buy out the middle of a pyramid, every floor that shares those fabs pays more just to route around the hole. And it's not letting up. The squeeze runs through 2027 into 2028. Memory was always a pyramid, AI just bought the middle floors. The pyramid held steady for decades until AI decided to buy out the middle floors. And if you want to see exactly how that's about to blow up the price of your next PC, which companies, how far it runs, and where all that money actually lands, I pulled the whole thing apart in the next one.

[00:11:02]It's called why DRAM prices are about to explode. Go watch that one next.