Hello, my two cents - it would have been better if your article shows some concrete quantitative analysis on the projections in 2028 onwards taking into account legacy demand, AI demand, and additional supply, split by the three main products (dram, hbm, nand) and the key assumptions. There are a lot of debates between the “memory is cyclical” and “this time is different” due to change of the role of compute. However I have yet to see someone publish an in depth scenario analysis, and what I see mostly are more general convictions.
Fair enough. I try to write my articles to allow everyone to understand - I think it lowers the readability for most people when things get too technical.
That said I can certainly schedule an article which really gets into the weeds of the industry. I’ll sit on it for a bit and see if I can come up with a way to explain some of the more complex topics to readers who don’t have the time to do hours of research independently. Thanks for the feedback!👍👍
You speak as if it's so easy to improve yields from 50% to 90% 😂 and no, HBM is not indistinguishable from each other and there's a lot of work on custom HBMs being designed with certain clients
I never said it was easy - I said it was an engineering problem, not a physics problem. Micron, Samsung, and SK Hynix employ 15,000+ engineers across all three companies focused on exactly this. When you have that much money and that many PhDs working on one specific process, it seems inevitable. It won't happen overnight, but I predict it will happen by 2028.
You're right that custom HBM exists. But it's a thin layer on top of a commodity base. NVIDIA custom modules, maybe Tesla custom modules, a few others. That's high-margin work for a small percentage of total HBM production. The other 80%+ is indistinguishable. Once yields improve and supply doubles, even the premium custom work faces margin compression from increased competition for those slots.
The bulk of HBM supply is commodity, the current scarcity is engineering-driven. The custom work won't save overall sector margins when the supply shock hits.
Solid take. Something I have been thinking as well within the debate of whether this memory cycle represents a structural shift of just a super-cycle. Everything tells me that so far, the industry is closer to a super-cycle than a structural shift given that none of the big three hold a moat that the other cannot replicate. Great work, keep it up!
Thanks Andy! While Micron has the technological advantage right now, i expect Samsung will hit back at them soon, and hard. It all ends with a race to the bottom at the end of the day😉
Great article Darius. I work at the Sharp edge of non AI computing working for a global Broadcaster with over 500 esxi hosts globally and Supermicro Linux hosts used as a 3rd in-house cdn. All out quotes for servers to replace eol equipment have increased by 50% due to ram costs (256gb-1TB per host) and our hardware vendors are all stating 2028 as the date when the memory shortage will most likely end.
Will look at some of your other posts as this is the first one I’ve seen which paints a realistic picture of the long term for these companies. No direct exposure as didn’t move quickly enough but have exposure through a concentrated Uk investment trust that is owned by an ex Artemis fund manager which is performing extremely well.
Great to hear you’re performing so well in the UK! Feels like the first time in a long time that the UK is actually heading in a better direction. Pretty sure people are realising that domestic investments are in many ways much safer with the US being so politically volatile. Let’s hope it keeps up!!👍👍
The Uk is actually going in the opposite direction due to our current useless socialist government. Should have said the investment trust is based in the Uk but focuses heavily on tech with 25 securities. Trust is Blue Whale growth.
The article applies an old framework to a new regime.
By ignoring the architectural requirements of platforms like Vera Rubin and the memory-heavy nature of Agentic AI reasoning, the bear argument misses the fact that memory is no longer just a passive storage tank - it has become the primary infrastructure bottleneck for next-generation computing.
Even if demand is hitting records, memory CapEx is skyrocketing too. The wave is always dragged back down once supply catches up. Assuming 'this time is different' in a notoriously cyclical industry is a dangerous bet.
Please for example check this video published yesterday. Michael Dell and Jensen Huang expressing their frustration about the memory situation pretty bluntly. Definately no need to read between lines.
The memory market is obviously extremely right right now, which is the exact reason Micron is up 100+% YTD. I think the market is confusing a temporary bottleneck with permanent pricing power though, which is why I argue that my 2028 this pricing power will be gone.
Not claiming memory isn’t a bottleneck right now - I just don’t think it will be this way forever.
Broadly agree on the cyclical nature, but I kept making gains on the volatility of memory stocks and have no regrets about not buying and holding. The memory scene is bonkers, and frankly the whole market is detached from the building geopolitical pressures and a long etc. — none of which makes me comfortable parking capital in traditional long or mid-term positions right now.
To be fair, once the data centre craze tops out on demand, the rise of robotics and other memory-hungry tech will be waiting. So it's hard to land on "you're absolutely right" or the opposite. My two cents: keep watch, know the risk, size for it. 🦊
All the memory chips in the data centres need to be replaced within 3-5 years time for the compute to be faster and efficient. Though the chipmakers wont show YoY growth as they are showing now but the demand must stay high in future also for them not go into a negative loop.
It’s an interesting one. I’m not sure I agree. Markets are based entirely on sensationalism. I think that’s the entire reason cyclical stocks exist at all.
Spot on. I usually avoid memory stocks because the historical cycles are absolutely brutal. I will admit the AI demand is real. But the reason to step back right now is not about demand dropping; it is purely about these nutty valuations. We just watched these companies rocket 800% in a matter of months. When a highly cyclical stock gets priced for absolute perfection, your margin of safety vanishes entirely. The underlying business might be booming, but at these parabolic prices, sitting at the table just is not worth the risk anymore.
Hello, my two cents - it would have been better if your article shows some concrete quantitative analysis on the projections in 2028 onwards taking into account legacy demand, AI demand, and additional supply, split by the three main products (dram, hbm, nand) and the key assumptions. There are a lot of debates between the “memory is cyclical” and “this time is different” due to change of the role of compute. However I have yet to see someone publish an in depth scenario analysis, and what I see mostly are more general convictions.
Fair enough. I try to write my articles to allow everyone to understand - I think it lowers the readability for most people when things get too technical.
That said I can certainly schedule an article which really gets into the weeds of the industry. I’ll sit on it for a bit and see if I can come up with a way to explain some of the more complex topics to readers who don’t have the time to do hours of research independently. Thanks for the feedback!👍👍
You speak as if it's so easy to improve yields from 50% to 90% 😂 and no, HBM is not indistinguishable from each other and there's a lot of work on custom HBMs being designed with certain clients
I never said it was easy - I said it was an engineering problem, not a physics problem. Micron, Samsung, and SK Hynix employ 15,000+ engineers across all three companies focused on exactly this. When you have that much money and that many PhDs working on one specific process, it seems inevitable. It won't happen overnight, but I predict it will happen by 2028.
You're right that custom HBM exists. But it's a thin layer on top of a commodity base. NVIDIA custom modules, maybe Tesla custom modules, a few others. That's high-margin work for a small percentage of total HBM production. The other 80%+ is indistinguishable. Once yields improve and supply doubles, even the premium custom work faces margin compression from increased competition for those slots.
The bulk of HBM supply is commodity, the current scarcity is engineering-driven. The custom work won't save overall sector margins when the supply shock hits.
Solid take. Something I have been thinking as well within the debate of whether this memory cycle represents a structural shift of just a super-cycle. Everything tells me that so far, the industry is closer to a super-cycle than a structural shift given that none of the big three hold a moat that the other cannot replicate. Great work, keep it up!
Thanks Andy! While Micron has the technological advantage right now, i expect Samsung will hit back at them soon, and hard. It all ends with a race to the bottom at the end of the day😉
Great article Darius. I work at the Sharp edge of non AI computing working for a global Broadcaster with over 500 esxi hosts globally and Supermicro Linux hosts used as a 3rd in-house cdn. All out quotes for servers to replace eol equipment have increased by 50% due to ram costs (256gb-1TB per host) and our hardware vendors are all stating 2028 as the date when the memory shortage will most likely end.
Thanks for sharing, Andrew. Real-world insights like this are incredibly valuable👍👍
Will look at some of your other posts as this is the first one I’ve seen which paints a realistic picture of the long term for these companies. No direct exposure as didn’t move quickly enough but have exposure through a concentrated Uk investment trust that is owned by an ex Artemis fund manager which is performing extremely well.
Great to hear you’re performing so well in the UK! Feels like the first time in a long time that the UK is actually heading in a better direction. Pretty sure people are realising that domestic investments are in many ways much safer with the US being so politically volatile. Let’s hope it keeps up!!👍👍
The Uk is actually going in the opposite direction due to our current useless socialist government. Should have said the investment trust is based in the Uk but focuses heavily on tech with 25 securities. Trust is Blue Whale growth.
The article applies an old framework to a new regime.
By ignoring the architectural requirements of platforms like Vera Rubin and the memory-heavy nature of Agentic AI reasoning, the bear argument misses the fact that memory is no longer just a passive storage tank - it has become the primary infrastructure bottleneck for next-generation computing.
Even if demand is hitting records, memory CapEx is skyrocketing too. The wave is always dragged back down once supply catches up. Assuming 'this time is different' in a notoriously cyclical industry is a dangerous bet.
I think you might be the one assuming too much.
Please for example check this video published yesterday. Michael Dell and Jensen Huang expressing their frustration about the memory situation pretty bluntly. Definately no need to read between lines.
I think they know the situation quite well.
https://youtu.be/oE5lNDhz9oo?is=wJl-nL1IeQNyxYfq
The memory market is obviously extremely right right now, which is the exact reason Micron is up 100+% YTD. I think the market is confusing a temporary bottleneck with permanent pricing power though, which is why I argue that my 2028 this pricing power will be gone.
Not claiming memory isn’t a bottleneck right now - I just don’t think it will be this way forever.
Broadly agree on the cyclical nature, but I kept making gains on the volatility of memory stocks and have no regrets about not buying and holding. The memory scene is bonkers, and frankly the whole market is detached from the building geopolitical pressures and a long etc. — none of which makes me comfortable parking capital in traditional long or mid-term positions right now.
To be fair, once the data centre craze tops out on demand, the rise of robotics and other memory-hungry tech will be waiting. So it's hard to land on "you're absolutely right" or the opposite. My two cents: keep watch, know the risk, size for it. 🦊
Thanks for reading!
bold take!
I try my best 😁
By your logic ASML is also not a physical bottleneck. Gather 15000 engineers for a decade and you get an EUV machine.
1 company with complete dominance over all competitors is very different to 3 all competing tooth and nail to be first.
All the memory chips in the data centres need to be replaced within 3-5 years time for the compute to be faster and efficient. Though the chipmakers wont show YoY growth as they are showing now but the demand must stay high in future also for them not go into a negative loop.
I wouldn't be surprised if the crash happens in 2027.
Only a matter of time!
All this goes down to NetList. Hope things will turn around for them with courts and licenses
I believe you like history but not understand technology at all!
When you don't understand Lithography tool and understand a bit about memory, then you start to build your own world view.
Hi Skyboat! What is it you think I don’t understand?
Totally agree DDI and my post today discusses this topic
Thank you! Glad you enjoyed the post👍👍
If there is a guaranteed crash then markets have priced it in 👍
It’s an interesting one. I’m not sure I agree. Markets are based entirely on sensationalism. I think that’s the entire reason cyclical stocks exist at all.
This is an absolutely superb note.
Thank you! Glad you enjoyed it👍👍
Spot on. I usually avoid memory stocks because the historical cycles are absolutely brutal. I will admit the AI demand is real. But the reason to step back right now is not about demand dropping; it is purely about these nutty valuations. We just watched these companies rocket 800% in a matter of months. When a highly cyclical stock gets priced for absolute perfection, your margin of safety vanishes entirely. The underlying business might be booming, but at these parabolic prices, sitting at the table just is not worth the risk anymore.