Two forces are reshaping the memory market over the next few years. Neither shows up in the headlines. And you can’t judge a single memory stock without understanding both.

That’s today’s focus.

The memory shortage nobody’s watching

Everyone knows about the HBM shortage. Few are watching the one bubbling underneath it.

Decades of boom-bust cycles have taught memory makers one lesson the hard way: when you get the chance to invest big, spend it on the product with the fattest margin. That’s your insurance policy when the cycle turns. They’re following that lesson right now. And it’s quietly starving NAND (flash) memory.

Just look at where the new memory capacity is going:

  • SK Hynix built its M15X fab in Cheongju for DRAM and HBM—on a site long earmarked for NAND—and pulled the opening forward twice to get it running sooner. It’s also fast-tracking a substantial portion of its (much larger) Yongin Semiconductor Cluster buildout while focusing almost entirely on DRAM/HBM.
  • Samsung is reportedly shifting the focus of its ramping P4 fab to DRAM. And the company’s P5 fab is prioritized for HBM/next-gen DRAM.
  • Micron’s new Boise fab being built now and its planned megafab in Clay, New York are both dedicated to leading-edge DRAM (including HBM).

From what I could find, the only capex going to NAND is from Sandisk and Kioxia retooling existing facilities for denser SSDs, Micron expanding its NAND fab in Singapore—where wafers aren’t expected until the second half of 2028, and SK Hynix recently approving about $13.4 billion for NAND investment at its M17 fab in Cheongju targeted for late 2028.

The obvious rebuttal here is China’s YMTC. It’s now the third-largest NAND maker in the world by bits shipped, and it’s reportedly bringing a new fab online by the end of this year with a long-term target of 500,000 wafers a month.

That’s real supply. But it’s cheap consumer flash, not the high-layer SSDs AI data centers are fighting over. And US sanctions keep YMTC out of most Western data centers regardless of what it builds. So while YMTC will absolutely affect the pressure on the low end of the NAND market, it does little if anything to relieve the specific shortage I’m talking about.

The point is that most of the money going into adding memory capacity is aimed away from NAND, right as demand takes off.

For example, in 2026, for the first time ever, data centers became the largest consumer of NAND flash, passing smartphones.

Why? Because of what we talked about in Part 1: NAND flash serves as AI’s staging area. Checkpointing during training. Keeping GPUs fed with datasets. And holding the overflow when the faster memory fills up from increasingly long context windows.

Which means NAND could stay short a lot longer than DRAM does. The current DRAM shortage is widely expected to run at least through 2028. SK Hynix’s CEO told Reuters in July that customer demand will exceed the company’s supply capacity “even beyond 2030”—and he was talking about memory across the board, not just DRAM.

If that’s right for memory broadly, and NAND is the piece getting the least new capacity, I don’t see how things start to balance in the flash market before the 2030s.

I should tell you that global market intelligence firm TrendForce disagrees with me on this. In July they forecast NAND supply catching up to demand in the second half of 2027, mostly because process migrations add bits without needing new buildings, and because phone demand keeps falling.

They could be right. Squeezing more layers onto the same wafer is real supply, and it’s cheaper and faster than pouring concrete.

But I still think I’m right. Process migrations get harder every generation, and they don’t add cleanroom space, which is the actual binding constraint right now. Plus the demand side of their model leans on phones and laptops staying weak. If AI inference keeps eating flash the way it has been recently, and I fully expect it to, a soft smartphone market won’t be enough to balance anything.

Bottom line: I think NAND stays tight at least as long as DRAM does, and probably longer. The capital that would fix it is being spent elsewhere.

Look at what NAND prices have already done. In the second quarter of this year, TrendForce had NAND contract prices rising 70% to 75% from the first quarter—outpacing DRAM’s 58% to 63%. And remember, those are three-month numbers, not year-over-year. Contract prices climbing 70% in a single quarter is blisteringly fast. Across the first half of 2026, contract prices more than doubled.

Now, TrendForce sees the third quarter cooling off to something like 10% to 15% for NAND. Don’t read too much into that. When you’ve just doubled off a doubling, the math alone slows you down. A 10% to 15% quarter over quarter increase on top of the first half of this year is still an extraordinary number in a business used to seeing prices fall most years.

You don’t get a run like that unless demand is dwarfing supply.

Put the price action together with how long this undersupply could run, and I owe you a correction from a previous Grow or Die piece…

A correction I owe you on Sandisk

In mid-January 2026, I wrote a two-part essay on Sandisk (SNDK)—a NAND flash pure-play—in which I said I wouldn’t buy the stock at that time even though I thought the company was doing some impressive things.

My basic reasoning was that HBM had effectively transformed DRAM from a commodity into a more customized, contracted, premium product, and that until something did the same for NAND, flash would remain a boom-bust commodity and Sandisk would eventually get taken down with it. I said I’d want to see High Bandwidth Flash (HBF)—a tech with the same configuration as HBM but using flash instead, which I’ll come back to—actually change the market before I bought Sandisk.

I’ve changed my mind. And I want to briefly hit on why because it isn’t about HBF.

It’s that I underestimated how long a shortage can run when nobody’s building the thing that’s short.

The structural argument I made in January was about the destination. What I missed is that the road to get there is years longer than I assumed because the entire industry is allocating capital away from NAND to chase HBM. The same cleanrooms, the same equipment budget, and much of the same engineering talent, are needed for both NAND and HBM. And all those things are being pulled to the HBM side of the building so to speak.

Sandisk’s most recent quarter (fiscal Q4 2026) shows us what this looks like in practice: revenue grew 372% year-over-year, and 51% sequentially to $8.97 billion, while gross profit margin hit 84.6%.

For comparison, consider that Sandisk’s average quarterly gross profit margin in 2025 was about 32%. That’s not a typo. We’re talking about gross margin expanding more than 5,000 basis points in under a year. I’ve been doing this a long time and I can’t think of another example of a business repricing that hard, that fast. Meanwhile, data center revenue jumped 103% sequentially to $2.98 billion.

Worth mentioning real quick: That insane gross margin around 85% is a shortage artifact. It will come down when supply and demand start to balance out. So don’t expect it to continue forever.

I still wouldn’t say Sandisk is a low-risk stock by any means. And I can promise you it will be volatile in the near term. But considering what we see playing out in its market, I think the stock could be a big winner over the next three years—even taking into account its huge run over the past year or so that I talked about in Part 1.

I’ll have more to say about this later on in our Memory Investing Series when I dig into specific stocks.

One quick word on HBF, since I made it the thesis-changing condition back in January. It’s advancing. Sandisk and SK Hynix put out the first industry technical specs for it in early August. But I’m no longer waiting on it. HBF was basically my answer to “what turns NAND into a premium product capable of essentially forcing the market to change structurally and become less cyclical?” It turns out the shortage itself will have the same effect for the next several years, whether or not HBF ever ships in volume. I now view HBF as upside to the thesis instead of the thesis itself.

That’s the first force: capital fleeing NAND while demand for it climbs. It’s a timing opportunity, and timing opportunities have expiration dates.

The second force is different. It’s slower, it’s structural, and it changes what these companies are actually worth, permanently.

Memory is turning into a custom silicon business

This is the big one, and it’s the one I think the market understands the least.

It comes down to one piece of silicon: the base die. Think of it as the control floor at the bottom of an HBM stack. Every byte passes through it.

When you build HBM, you take several DRAM chips, stack them on top of each other like pancakes, drill thousands of tiny vertical wires straight through them (through-silicon vias, TSVs), and then sit the whole tower on one extra chip at the very bottom. That bottom chip is the base die. The DRAM pancakes above it are the core dies, the actual memory cells that store your bits.

The stack then sits next to the GPU on a shared slab of silicon called an interposer. The base die is what actually talks to the GPU. It also routes data up and down the stack through the TSVs like a traffic controller, and handles power management and error correction, among other things.

In other words, the base die is the only part of an HBM stack that behaves more like a processor chip than a memory chip. And that’s where the story is…

If you follow what’s happening at the base die layer, you can see the entire memory business changing character.

Through HBM3E, the memory makers like SK Hynix and Micron built the base die themselves on their own lines using a DRAM process.

The word “process” here basically means the manufacturing recipe. The tooling, materials, and layout rules a fab uses to print transistors on silicon. A DRAM process is optimized for cramming billions of tiny storage capacitors into a small space as cheaply as possible. Density and cost per bit are everything. The transistors it can make are slow and power-hungry by today’s standards. And it offers relatively few layers of metal wiring on top for routing signals around. This is all fine historically speaking because memory chips didn’t have to do much thinking.

The base die’s job was simple enough through HBM3E that it could be manufactured using a DRAM process.

That changed recently with HBM4. The base die now has to do far more work—more speed, more routing, more control—and a DRAM process can’t build it. So HBM4 base dies get built using a logic process. A logic process is basically optimized for the opposite of what a DRAM process is optimized for—transistors that switch superfast at very low voltage, plus over a dozen layers of metal wiring to route extremely complex circuits. It can’t make good storage capacitors, and it’s much more expensive per square millimeter.

Moving to a logic process brought TSMC into the HBM market and made SK Hynix its customer. Micron still builds its own HBM4 base die in house. But it’s already handed TSMC the job for the next generation. TSMC will manufacture both Micron’s standard and custom HBM4E base dies, ramping in 2027.

So the big memory makers (except Samsung, which has its own logic capacity) now need to buy leading-edge logic wafers from TSMC to ship their flagship AI product. That’s a new dependency in businesses that historically owned their entire manufacturing stack.

But turning memory makers into TSMC customers is just a small piece of this story. What really matters: once the base die is a logic chip, HBM stops being anything close to a commodity.

Once the base die is a logic chip, you can put customer-specific circuitry on it and even offload some computation from the GPU onto the memory stack. HBM4E extends this idea further with fully customizable base dies and custom interfaces. That turns HBM from a very complex but still interchangeable part into a truly co-designed, customized one, with all the pricing power and customer switching costs that implies.

We’re now entering the era of custom HBM. Yes, this is real. It’s on the official roadmaps of Samsung, SK Hynix, and Micron. And Marvell has announced it’s co-developing a custom HBM architecture with all three.

The next step after custom HBM is memory-on-logic, which I won’t get into now other than to say it makes memory even less of a commodity.

All this means memory companies are becoming custom silicon companies. Not entirely, and not overnight. But at the top end of their product line, where they make their money, their business is looking a lot less commoditized.

For decades, memory was defined by three things. Standard parts. Prices that swung wildly with supply and demand. And bargaining power sitting with buyers, because they could always walk. That’s why these stocks always earned a low, cyclical multiple.

HBM has already changed things. Multi-year agreements locking in volume and price are now standard. In October 2025, SK Hynix said its entire 2026 DRAM (including HBM) and NAND capacity was sold out. Micron signaled the same for its 2026 HBM about two months later.

But those are contracts about volume and price. They aren’t contracts about the product. When one expires, the customer can move.

Custom HBM changes that. Designing a base die jointly means fixing specs and volume two to three years before shipment, which makes the order book look like a logic chip company’s backlog rather than a commodity producer’s. New sources of revenue show up that never existed in memory, like engineering fees recovering design costs. And the customer lock-in becomes real and physical: switching suppliers now requires a redesign. Executives at Samsung and SK Hynix have suggested custom HBM gains momentum in 2027 and becomes an important pillar of the market around 2028 to 2029.

I’m not saying volatility is going to disappear because of this shift. But I am saying it changes shape, basically from market-driven to order-driven. And order-driven cycles have historically earned higher multiples than unpredictable commodity cycles. We have the whole history of the logic chip industry to confirm that.

One final note: You can probably see now that while the two forces I talked about seem completely different on the surface, they’re really one force with two effects.

The force: HBM “gravity.”

The two effects: It pulls capital away from NAND and pulls DRAM toward custom logic.

That’s all for today. Now that you have a good grasp of the memory market’s basics and some of the big things driving it, the time has come to talk about specific stocks.

Next time, I’ll rank some of my favorite memory stocks and explain why I think they’re poised to outperform (or not).

Thank you for reading. I hope you have a wonderful weekend.