# The Great Memory Trade

Source: https://kermankohli.substack.com/p/the-great-memory-trade
Kiwi News: https://news.kiwistand.com/stories/The-Great-Memory-Trade?index=0x69f1b8cefe11018a5771db21f5989f5c4b8e61953bb21a91b02c54a72787882f9e6fb6ad
Submitted by rvolz.eth on 2026-04-29 UTC, 1 upvote, 1 comment

## Summary

The author, who says they have their entire net worth in SK Hynix, Sandisk and Micron as of April 2026, argues that memory is the most bottlenecked and most underpriced component of the AI trade. Memory bandwidth and capacity limit which AI models can run and how much context they can use, and the author says HBM is required for every NVIDIA chip sold. The author cites SK Hynix's expected 2027 profit of $150 billion against a $540 billion market cap, and Micron's expected FY2027 profit of around $100 billion against a $500 billion market cap. The author also rejects the view that memory is a cyclical commodity, pointing to long-term agreements and CEOs saying memory is sold out until 2027 at the earliest.

## Article

It’s now been just under a year since I’ve been investing in semiconductors actively. My first article around “We don’t have enough compute” was basically spot in in August 2025. RAM shortages kicked in around September 2025 and semis have been parabolic since then.

I had just started migrating my capital to semiconductors from crypto at the time. Virtually all the names I listed in my Machine Economy article have doubled since November last year.

I do feel relatively confident that what I write about, directionally speaking, materialises fairly accurately in the coming months/years. Since I’ve published those articles I’ve refined my strategy even more and as of April 2026 have my entire net-worth in three stocks:

SK Hynix, Sandisk, Micron.

I am all-in and am willing to commit to this publicly. I trade options around other semiconductor names but there’s nothing I love more than I have conviction on than memory.

These are all stocks that are up 250%-1,000% in the past year alone. I wish I had purchased more earlier but that’s okay, I believe we’re still early.

Most people look at all the AI tickers and think they have missed the boat or are unsure what to buy or allocate into. My singular answer here is memory by-far. I wasn’t as sure about this when I first wrote about memory but as I have:

Been tuned into the earnings calls for the major semiconductor companies

Buying my own hardware and seeing where the bottleneck is

Building lots of software with AI and understanding the capabilities of these models

My experience keeps telling me abundantly that memory is the biggest play that the Street does not understand.

To understand the memory trade you have to start at the technology first and foremost. In the AI race, it’s the models that we need to understand. The layperson thinks you need a powerful computer/GPU to be able to do all the computations and that’s what is important.

That is partially true but what is forgotten is that the scale of these computations has to be kind of “held in the computers’ head”, like how a human does mental math and carries digits over in their head.

In order to do this you need two things:

Memory capacity (raw GBs)

Memory bandwidth

The first people don’t really understand, the second even less so.

When you think about AI models they typically come in classes of parameters. I created this graphic below that gives you context about the size of current AI models.

A core premise here is basically that your more capable models will be more parameters. Efficiencies mean that more smaller models become more capable but that means the larger parameter models become even more capable. In the AI world, being memory poor means you can’t:

a) Run the models you want b) Be limited with how much context you can provide them

Both a) + b) are extremely important.

If you want to be capable in the AI world, it’s less about how many calculations you can do at once but rather how many calculations can you do at-once, coherently towards a singular outcome.

To make it clear: higher parameter models require more memory.

Without getting into quantisation (a story for another time, assume FP16 for now), you can roughly infer that for every parameter you need 2GB of VRAM (GPU memory) per 1B parameters. So for something like GPT 5.4 to run on consumer hardware you’d need 2TB of VRAM (which is so insanely expensive or borderline impossible to get).

Some like like Qwen3.5 come in smaller variants like a 35B parameter model (which I am running right now) that takes 70GB of VRAM and only leaves me with 24GB of VRAM on my $10,000 GPU running in my $15,000 machine.

The RTX 6000 Pro Max card is the largest VRAM configuration available on the market. If you want to run more capable models like GLM 5.1 or Kimi K2.5 you need at least 400-600GB of VRAM which requires a cluster of GPUs (about 6-8) which is where you get into your multi-hundred thousand dollar machines ($150k - $500k).

Also you have to remember that your AI sessions are just as important as the model. When coding, 128k context windows are nice but fill up fairly quickly. Ideally you want 256k context windows if you’re doing anything serious I find. With certain optimisations you can get away with less etc but the premise basically holds true: the more you want to use these models to do more cool things YOU NEED MORE MEMORY.

Hopefully by now you understand that AI is a game of memory. GPUs without memory are functionally useless (for inference that is). If you want the AI build out you can’t ignore memory.

I used to believe that public markets are efficient but coming into semiconductors, it feels like people are asleep at the wheel. I’d actually correct it and say mentally disabled at the wheel.

If we roughly break down all sub-sectors of the AI trade you’d have something that looks like the following categories:

Chip designers (NVIDIA, AMD)

Fabs (TSMC, Intel)

Equipment (ASML)

Memory (Micron, Hynix, Samsung)

Optics (Lumentum, Coherence)

Energy (Bloom + trad energy cos)

Cooling (Vertiv, Eaton)

What’s funny about each of these categories is that they’re predicated on the same underlying demand cycle. You can’t for example, have a ton of Nvidia GPUs being sold without a ton of memory being sold. All of these companies should theoretically have the same underlying demand assumptions being made.

Currently a company like Vertiv trades at a forward PE ratio of 50. Investors are willing to pay 50 times earnings because they believe that AI demand is so strong and it is worth paying that multiple for the future.

Then we go to a company like NVIDIA that holds the keys to the AI kingdom (if you don’t believe this then you need to spend more time with Chat to educate yourself), they trade at a forward P/E ratio of sub 20 as of the time of writing this article.

If NVIDIA does not sell their chips then Vertiv does not have chips to cool. Strange that Mr Market believes so but lets carry on.

For every NVIDIA chip that gets sold, it needs memory. Lots of it and it can only be purchased from 1 of 3 HBM (high bandwidth memory) manufacturers: Micron, SK Hynix and Samsung.

If you believe that NVIDIA’s future is worth 20 times their earnings, you surely wouldn’t think that memory is worth 5 times forward earnings for memory would you? Well actually, the market does.

This is where we come to the crux of our thesis:

Memory is the most bottlenecked component of the AI revolution yet is the most underpriced relative to all other categories.

But why? What prevents memory from being loved?

That’s really it. Average investors are basically LLMs, they have a limited dataset that stays static in-time until someone hard updates the weights (number go up for a long time) and then they update their priors. It’s actually really amusing to witness.

In this case, the LLMs, I mean investors (both will say the same) think that memory is cyclical and will crash in the future like it has in the past. Before refuting that claim lets actually dig in and understand it for a second.

Throughout each of these booms and busts, memory makers go through the following:

Demand explodes to levels that cannot be serviced

Memory companies start charging more due to their increased pricing power

All companies ramp up to increase supply at the same time

Ramping up supply requires CapEx that has on-going costs once built

New fabs come online, market flooded with supply

Profit margins turn negative

Companies struggle to stay alive

Investors become cautious

According to this logic, you could say “memory is a commodity cyclical industry”. However, that’s intellectually lazy. We need to dig deeper. To illustrate this point, we need to look at this chart that breaks down DRAM demand by segment (consumer vs enterprise):

697

Do you see what happened in 2023? Server overtook consumer (mobile). DRAM demand was very closely tied to mobile phone and other consumer electronic sales!

This is true for all prior DRAM cycles as well.

However, it doesn’t end there. DRAM can now be broken down into two product lines: HBM (High Bandwidth Memory) and DDR (Double Data Rate). DDR RAM is substitutable as they’re sticks you can take in and out of a machine. HBM is not. HBM is part of the GPU itself.

Every NVIDIA chip sold requires HBM memory.

By now hopefully you can see where I’m going with this. Before I do, I want to put something in perspective.

Memory for the past 20 years has been for regular computing use-cases that don’t require fast bandwidth (transfer speeds). However high performance computing (AI workloads) are ALL throttled by bandwidth/speed. They are what we call memory-bound after a certain point. This is a critical shift from the past.

Now, to bring it all back together. If you think DRAM/memory prices will crash in 2028 you need to believe that global AI demand will fall off a cliff in 2028. You have to believe that the world wakes up and says:

“You know what, this AI thing is kind of a scam and I’m just going to go back to being a neanderthal doing things by hand”.

If you do believe that then you should probably just stop reading this article and go back to buying Walmart and Coca-Cola. Jokes aside, there is a segment of the market that does believe this because their AI usage is limited to using a chat-bot 5-10 times a day in very domain specific tasks that require a lot of nuance.

However, for people like myself, LLMs have allowed us to operate a much larger surface area with very few people. The value they create is undeniable and regardless of whether we achieve AGI or not, they are worth the CapEx they demand.

Okay so how much upside is left still? Let me lay it out. On a conservative forward looking basis, SK Hynix is looking to make $150 billion dollars of profit in 2027. Bullish estimates go north of $200 billion dollars. So what is the market cap of the stock? $540 billion dollars. That is less than 4 times earnings for the most dominant memory company on the planet.

Lets look at another case: Micron. A stock that Wall Street loves to hate on.

In FY2027, Micron is expected to earn around $100 billion dollars of profit. Their current market cap: $500 billion dollars. Around 5 times earnings. These are EARNINGS and not SALES. To put it in perspective, these numbers ($100b of earnings) are the same amounts that Google/Meta make.

Why? Because “memory is cyclical”.

These 3 words have so much upside in the it’s kind of hard to fathom.

Even if you do believe that AI demand will fall off a cliff in 2028, all these memory companies are signing very strong Long Term Agreements (LTAs) with their customers spanning multiple years. If you go through most earnings calls, all the CEOs are saying the same thing: memory is sold out till 2027 at the earliest but most likely they have to look at 2030 for the next window of capacity.

Memory has already been squeezing everyone (5x increase in spot prices in 12 months) but we have not even begun yet. Memory prices are going to keep rising, memory manufacturers are going to keep raising their prices, Wall Street will keep shouting “memory is cyclical” forcing these companies to be disciplined about bringing new supply which in turn creates a larger shortage.

Even within the 3 companies that make HBM, they all serve different segments of the HBM market and are co-designing with different chip manufacturers.

If you’re short on memory, you also have to believe that robots will not be a thing in the future. Robots are extremely memory sensitive. Image and video generation as well. Long story short: the demand side of memory is still massive. If you’re unsure, just use a bit more AI.

Sandisk is a different bet as they make NAND/Flash storage but the same logic can more or less be extrapolated. They have a hidden advantage of their HBF (high bandwidth flash) they’re co-making with SK Hynix that could be a real game changer. Most people look at Sandisk and think they’ve missed the boat. This is also an incorrect understanding of the stock and the market it operates in.

To close this article I’ll leave a quote that I think is relevant to investing and life:

“The best time to plant a tree was yesterday, the next best time is today”

I hope you learned something from this piece and plant the trees today that will harvest in the future.

## Comments

**rvolz.eth**: > I had just started migrating my capital to semiconductors from crypto at the time.