A Blackwell superchip (one Grace CPU paired with two Blackwell GPUs) costs Nvidia something like $13,000 to $18,000 to build. More than half that cost, by some estimates over 60%, is memory.
The processors, the “brains” of AI, are now cheaper than the memory bolted to them.
Just a few years ago, almost no one could have seen this coming.
For decades, memory was the dog of the tech world. Nobody cared whose DRAM was inside their laptop or data center any more than they cared which farm grew the wheat in their bread. Prices swung wildly. Memory companies made a fortune for a year or two then lost a fortune the next year or two.
That world is gone.
Memory and storge stocks have gone gangbusters over the past year or so. They’ve recently pulled back from their 2026 highs, but the 12-month gains are still staggering:
- Western Digital (WDC) up about 600%
- Seagate Technology (STX) up about 500%
- Micron (MU) up about 900%
- Sandisk (SNDK) up about 3,000%
Many investors look at these gains and assume there’s no upside left in memory stocks. But if you understand the dynamics of what’s happening, you can see there’s still a huge opportunity remaining.
My goal with this Memory Investing Series is to help you uncover that opportunity. I’ll also reveal my favorite stocks in the space along the way.
I’ve had conversations with sophisticated investors who simply don’t know how the memory business works and what they should be looking for when it comes to potential investments.
There’s absolutely nothing wrong with that. But because the next five years will be decided by things happening two and three layers down from where most folks are looking, and because you can’t judge those things if you don’t know what the layers are, it makes sense to start this series by bringing things back to the basics and mapping the ecosystem.
That’s what we’ll do today in Part I. What the pieces are, what they do, who makes them, and the terminology you need.
Let’s start with why any of this matters in the first place…
To lean on an analogy I’ve been using for a couple years: Picture an AI system as a kitchen. The processor (GPU, TPU, XPU, etc.) is the chef. Memory is the pantry.
If the pantry is across the parking lot, your chef spends the day jogging back and forth for ingredients. It doesn’t matter how fast he can chop and cook; the pantry’s a major bottleneck.
That’s basically what’s happening inside AI data centers. According to memory giant SK Hynix, more than 90% of the time it takes an AI model to answer you is spent moving data back and forth between the processor and memory, not computing anything. Sixty-thousand-dollar superchips sitting idle, eating electricity, burning cash, waiting for the pantry.
The industry calls this the “memory wall.”
The memory wall exists because compute’s performance gains have been significantly outpacing those of memory for at least two decades. Research published by the Institute of Electrical and Electronics Engineers (IEEE)—the entity that creates many global technical standards—revealed that processing power has been tripling every two years for the last 20 years, while memory performance has improved just about 1.4X every two years.
That gap became the bottleneck.
A simple lesson I’ve learned and leaned on time again in my 25 years as an analyst and investor in disruptive megatrends: own the bottleneck(s).
In the first phase of the AI buildout, the bottleneck was GPUs, and Nvidia became the most valuable company on earth. We’ve been invested in Nvidia (NVDA) since September 2020 in one of my paid advisories, Disruption Investor.
In this phase, one of the biggest bottlenecks is memory. Three years ago, about 75% of AI infrastructure spending went to GPUs. Today, it’s closer to half, and falling, because the money is flowing into everything that feeds them.
When you’re spending north of $700 billion a year on data centers—and that’s just the four big US hyperscalers (Google, Amazon, Microsoft, and Meta)—a shift like that is enormous.
Meeting the four families of memory
In 1956, IBM researchers in California built the first computer with a hard disk drive (HDD), the 305 RAMAC. The drive was the size of two refrigerators, held 50 spinning magnetic platters, and stored about 5 megabytes. That’s enough to store about one of today’s high-quality photos taken with your smartphone.
Hard drives are what the industry calls non-volatile memory. When you turn off the power, the data’s still there. HDDs won the storage business for the next half-century. To this day nothing stores a bit more cheaply.
In the 1980s, a Japanese engineer working at Toshiba, Fujio Masuoka, invented flash memory. **NAND **is a specific type of flash memory, but it’s so dominant the terms NAND and flash are often used interchangeably today. It’s what solid-state drives (SSDs) are made from.
NAND is also non-volatile, but it has no moving parts. It’s built directly onto silicon the way processors are, so it shrinks and scales the way chips do. It’s also more reliable and much faster than HDDs. But for decades it lost on the metric that mattered most: cost per bit. It wasn’t until the 2000s that SSDs became a real option for businesses, and a few years later for your laptop.
Back at IBM in the mid-1960s Dr. Robert Dennard invented the memory that still runs the world today. The industry called it 1T1C (one transistor, one capacitor). It became DRAM (dynamic random-access memory).
DRAM is volatile memory. Cut the power and the data disappears. That sounds like a flaw, but it’s the whole point. Because it “forgets” engineers could make it really fast.
DRAM is about 1000X faster than an SSD and 100,000X faster than a HDD.
The fourth family worth knowing about is SRAM (static RAM). It’s 10X to 100X faster than DRAM. But that speed comes at a huge cost. So SRAM is typically used sparingly. It accounts for less than 1% of all memory shipped.
That said Cerebras and Groq built entire AI chips around SRAM for lighting fast inference. They work great, but their capacity is tiny. One AMD MI450-series GPU, for example, holds 432 gigabytes of memory, about the same as 864 of Groq’s chips. Groq’s chips are far faster per bit. They just have far fewer bits.
SRAM’s speed was enough for Nvidia to pay about $20 billion in late 2025 for Groq’s assets and much of its team. But it’s a specialist, not a DRAM replacement.
What each type of memory actually does inside an AI data center
The different types of memory aren’t really competitors. They’re a stack. And each one has a job.
SRAM is used only in tiny amounts right on the processor as cache, so this section will focus on the other three families.
HDDs serve as the archive. Hyperscalers have been hoovering up every scrap of data they can find to train their models. Hard drives are the only thing cheap enough per bit to hold it all. About 90% of the data stored in hyperscale data centers sits on spinning disks. And here’s a point Western Digital’s management made that I think is underappreciated: when you run an AI system, the compute and memory get recycled and reused, but the data the AI generates does not. It gets stored. Forever. On HDDs.
NAND flash serves as the staging area. Hard drives are too slow to feed running GPUs, so SSDs sit in between streaming them datasets. They’re also used to save checkpoints while a model trains—so a week’s worth of work isn’t lost if something crashes—and to hold the overflow when the faster memory fills up.
DRAM serves as the workhorse. This is where the real work happens. And within DRAM there’s a hierarchy of its own, which we’ll get to in just a second.
The rule of thumb: Every operation that *can *be done by the fastest memory is done by it. Only when a job outgrows at pool does it spill down to the next tier, and then the next.
The three DRAM kingdoms
DRAM isn’t one product. In 2026 it’s three, and they do different jobs.
DDR5 is the general-purpose DRAM. “DDR” stands for double data rate. It moves data on both its up and down beats, which is where the doubling comes from. The number “5” is just a generation marker, like a model year. DDR5 launched in 2020. It’s about twice as fast as DDR4, and it’s the overwhelming majority of DRAM shipped today. It plugs into slots on a motherboard using modules called **DIMMs **(dual in-line memory modules).
In a data center, DDR5 is the memory pool for the CPUs.
LPDDR is the low-power version of DDR.
It was designed in the mid-2000s for mobile phones, where battery life matters more than raw speed. To save power it gets soldered directly onto the board instead of plugging into a slot.
That sounds like it would be terrible for a data center because if it fails, you’d have to throw out the whole board instead of just swapping out and plugging in a new module.
But Nvidia looked at that weakness and decided to make it a feature…
In 2023 Nvidia launched its first AI “superchip” design, combining an Arm-based CPU and a stack of GPUs. Feeding that CPU with normal DDR5 would have required a tangle of copper wires running out to DIMM slots. This configuration would have consumed too much power and given off too much heat for an AI rack. So Nvidia’s engineers soldered LPDDR right next to the CPU instead.
The catch was exactly what we just talked about. When the soldered memory fails, you can’t replace it. So you lose the CPUs and GPUs. The hyperscalers didn’t like it, but they had no other viable options.
Being the great company that it is, Nvidia worked with the biggest memory companies to solve this problem too. Together they developed a module format called SOCAMM that lets low-power memory slot into a board the way DDR does. It’s modestly higher in terms of power consumption and heat compared to pure soldered LPDDR, but still far better than DDR5 modules. The second generation of this tech (SOCAMM2) will go into Nvidia’s Vera Rubin racks, and this adoption alone is expected to push LPDDR shipments for AI servers close to smartphone volumes in the next year or two.
HBM (high-bandwidth memory) is the third DRAM kingdom. It’s the one everybody talks about.
If DDR and LPDDR are the pantry, HBM is a cabinet with the most-used ingredients bolted right next to the stove.
Instead of laying memory chips flat around the processor, HBM stacks them vertically 8 to 16 layers high into a DRAM skyscraper, then drills electrical connections straight down through the silicon (through-silicon vias, or TSVs) that let the whole stack behave as one pool of memory.
The stack is then placed right beside the GPU on a slab of silicon called an interposer, which is basically an ultra-fine circuit board made of silicon instead of plastic (and will play a much bigger role in our discussion next time). Because the stack sits next to the GPU rather than on top of it, engineers call this configuration 2.5D rather than 3D.
The result: instead of data crossing centimeters of circuit board through 64 wires, it crosses millimeters through more than a thousand.
HBM can cost 10X more per gigabyte than ordinary memory because:
- It needs TSMC’s advanced packaging (called CoWoS, chip-on-wafer-on-substrate)
- Its yields are relatively poor because it’s so complex
- It eats 3X to 4X as many silicon wafers as normal DRAM for the same number of bits
The industry calls that last one the production trade ratio. It’s important because it’s the reason the entire memory market is experiencing a shortage right now. Every wafer a company points at HBM is three or four wafers it can’t point at the DDR5 in your laptop and the LPDDR5 in your smartphone.
The connective tissue
There’s a layer of specialized tech that helps data move rapidly and cleanly between memory and the processor.
Rambus (RMBS) is a great example. It sells physical “helper” chips that sit on memory modules and keep signals fast and reliable. It also licenses design blueprints (IP, intellectual property) to companies like Nvidia, Micron, SK Hynix, Samsung, Intel, Broadcom, and Qualcomm, collecting royalties when these companies use its designs.
Rambus built the industry’s first HBM4 controller, basically a smart traffic manager that sits on the processor side. And its tech resides in more than a hundred different chips. It also co-invented a new module format called MRDIMM, which interweaves two lanes of memory traffic to approximately double the bandwidth in an existing server slot.
Astera Labs (ALAB) makes switches and signal chips that let processors talk to each other and to pooled memory.
Credo (CRDO) and Marvell (MRVL) occupy neighboring ground. High speed connections, cables, controllers, and the silicon that manages storage drives.
The picks and shovels
Equipment. Nobody makes a memory chip without these companies. ASML sells the EUV lithography machines that print the finest features in leading-edge DRAM. Applied Materials (AMAT) and Lam Research (LRCX) supply the deposition and etch tools. KLA (KLAC), Onto (ONTO), Nova (NVMI) and Camtek (CAMT) handle inspection and measurement. Advantest (ATEYY) and Teradyne (TER) test the finished chips. FormFactor (FORM) makes the probe cards that touch thousands of points on a wafer at once. And BE Semiconductor (BESIY) and Kulicke & Soffa (KLIC) make the bonding tools that stack HBM layers.
Packaging. Amkor (AMKR) and ASE Technology (ASE) are the two big outsourced assembly and test companies, known as OSATs. Cohu (COHU) supplies test handlers and inspection.
Optics. As data travels farther, light replaces copper. Coherent (COHR) and Lumentum (LITE) make the lasers. Fabrinet (FN) and Applied Optoelectronics (AAOI) build the finished modules.
Systems. Companies like Dell and HPE build the servers.
Who actually makes the memory
HDDs are essentially a duopoly plus one: Western Digital (WDC) and Seagate at about 40% market share each, Toshiba with about 20%.
**DRAM **is the tightest oligopoly in tech. SK Hynix, Samsung, and Micron control most of the market, including more than 95% of HBM.
**NAND **is slightly more crowded. Samsung has about 30% market share. SK Hynix (including its Solidigm unit) has about 20%, and Micron, Kioxia, and Sandisk somewhere in the low-to-mid teens each, with China’s YMTC around 10%.
Wrapping up…
Hopefully this gives a decent picture of the broad memory market.
Next time, we’ll start digging into the specific developments driving this market over the next few years… and we’ll start to zero in on how to profit.
Thank you for reading.
