Memory has always been a brutal and cyclical business. Boom, bust, repeat. Prices spike when supply is tight, everyone builds more capacity, supply floods the market, prices collapse. Anyone who has watched semiconductors for a few decades has seen this storyline a dozen times, and they are right to be skeptical when someone tells them this time AI is different.

So let me tell you exactly why I think this time actually is different, and it comes down to one idea. The demand for AI memory rides an escalator that keeps going up, and the thing that ended every previous memory cycle, demand hitting a ceiling, is nowhere in sight.

The phone plateau

Here is the pattern that killed the last parabolic memory story.

Smartphones drove a massive memory boom. Every new phone carried more RAM than the last, and for years that was a real selling point. People upgraded, memory content per phone climbed, and the makers printed money.

Then it stopped. Not because phones disappeared, but because phones got good enough. At some point a phone had enough memory to do everything a phone does, and adding more stopped mattering to anyone. The spec sheet still listed it, but buyers stopped caring. Memory per phone plateaued because the use case saturated. And once that happened, the memory market went back to being a replacement business, cyclical and commoditized, because there was no more growth in content per device.

That plateau is the heart of why memory has been a boom-bust commodity. Demand has a ceiling, and once you hit it, you are just selling replacements into a saturated market. Ironically, AI has now driven memory prices so high that phone makers are pulling back on it, hesitant to add more even where customers might want it.

This has happened before, more than once

The smartphone wasn't the first time memory rode a demand wave and then fell off it. The personal computer drove the booms before that. Through the 1990s, every new PC needed memory, and as operating systems and software got heavier, each machine needed more of it, so content per device climbed and the makers boomed. Then PCs got good enough, the upgrade reasons thinned out, and that wave flattened too.

Each era had its engine, the PC, then the smartphone, and each engine eventually ran out of road, because there are only so many computers and phones a person needs, and at some point each device had enough memory to do its job. The boom always ended the same way: demand hit a ceiling. That is the pattern AI either breaks or repeats, and the whole question is which.

Why AI doesn't have that ceiling

Here is the thing that makes AI different, and it is simple enough to say in one line. Everyone owns one or two phones. But if agentic AI gets widely adopted, it is going to manage tens or hundreds of tasks for every single person.

Think about what that means. A phone is a fixed thing. One person, one or two devices, a bounded amount of memory. There is a natural ceiling because a human can only hold so many phones. But an AI agent working on your behalf is not bounded by anything physical you own. One person might have ten agents running, then fifty, then hundreds, each handling a task, each holding its own context, each consuming memory while it works. The demand isn't tied to how many devices you can hold in your hand. It is tied to how many things you want done, and that number has no natural limit.

I see this in my own business. The agentic tools I use already consume far more memory per task than a simple chatbot query, because an agent holds a long context, remembers what it did three steps ago, juggles tool calls, and reasons through multi-step problems. And that is today, with adoption barely started. Multiply that by every person running dozens of agents across every task in their life and work, and you are not looking at a market that plateaus. You are looking at one that compounds.

The escalator in the hardware

You can see this same logic in the chips themselves, and the numbers tell the story.

Look at the memory on Nvidia's data center GPUs over the years. The P100 around 2016 carried 16GB. The V100 went to 32GB. The A100 reached 80GB. The H100 stayed at 80GB of faster memory. The H200 jumped to 141GB. The B200 went to 192GB. The next generations push toward 288GB and beyond.

Every single generation carries more memory than the last. And not by a little. The capacity has roughly doubled every couple of years, and the bandwidth has grown even faster.

Why? Because memory is the binding constraint on AI performance. Each new chip has so much more compute power that it would starve without more memory and more bandwidth to feed it. The processor sits idle if the data can't reach it fast enough. So every time Nvidia increases the compute, it is forced to increase the memory right alongside it, or the expensive chip does nothing. Memory growth is chained to compute growth, and compute growth shows no sign of slowing.

This is the opposite of the phone. The phone plateaued because the workload stopped getting harder. You don't need more memory to send a text than you did five years ago. But AI workloads get heavier every year, bigger models, longer context, more reasoning, more autonomy. The thing the chip is being asked to do keeps getting more demanding, so the memory it needs keeps climbing. The escalator goes up because the work going up.

The honest caveat

Now let me give you the other side, because I would not trust this argument if it pretended to have no weakness.

The escalator is not a law of physics. It is a consequence of how AI is currently built. Today, more capability requires more memory. But that relationship could bend if someone figures out how to do more with less, an architectural breakthrough that gets the same AI performance while using memory far more efficiently.

This is not hypothetical. Earlier this month a research firm suggested Nvidia's next-generation platform might use less memory than expected, and memory stocks dropped hard on the fear. Nvidia's CEO refuted it directly and signed a multi-year supply deal with SK Hynix, so that specific scare was false. But the fact that the market flinched tells you the escalator is the thing the whole thesis rests on, and a genuine efficiency breakthrough is the one development that would break it. That is the risk I watch most carefully, more than China, more than new capacity, more than any headline.

So I hold the escalator as the strong base case, supported by years of history and by the basic physics of feeding a processor, while staying honest that it is an architectural consequence and not a guarantee.

Why it matters

If the escalator holds, this is not the old memory cycle, because the thing that ended every previous cycle, demand hitting a ceiling, does not happen here. The use case does not saturate the way phones did. It compounds, because agentic AI scales with the number of tasks people want done, and that number is effectively unlimited.

That is the difference between a commodity that booms and busts and an infrastructure that gets re-rated. The old memory cycle plateaued. This one climbs. And as long as every new chip needs more memory and every person runs more agents, the escalator keeps going up.

I hold a long position in DRAM ETF (ticker: DRAM) and Micron (Finally assigned on a CSP), and I added to it on recent weakness. Not investment advice. Do your own research.

My current MU positions as of June 15, 2026. (Assigned via a CSP $960 strike)

My current DRAM positions as of June 15, 2026.


Keep Reading