This guy is slowly becoming washed up. None of this article actually covers what is the purpose of HBM in agentic AI systems? What is the pJ/bit for non HBM DRAM is. How much BW does different LLM models need dense vs MoE and plot this across active parameters. If he starts doing analysis from first principles, he might realize why there may still be upside for memory and storage stocks from here. He also seems to have stopped posting his account update this time, coz optical stocks are getting whacked and this guy is all in into optics. This guy is going to get cooked with that portfolio in the next 12 months coz optical isn’t 2027 story.
You set the trap in the history section and then walked into it on purpose, which is the fun part. "This time is different" is the phrase you flag at every Micron peak, and the agentic re-rating thesis is that same line in a better suit. Telling them apart comes down to being precise about where the cyclicality actually lives.
It was never really a demand story. Every prior cycle had real, growing demand too. the cycles came from supply racing to meet it and overshooting, and your own roast names the culprit each time: Samsung doesn't cut. So "agentic demand is permanent" can be entirely true and still not break the cycle, because what breaks the cycle is supply discipline, and that's the variable that has never held.
Which makes the price/book to price/earnings re-rating less a bet on demand durability than a bet on supply discipline finally sticking. The bull case quietly swaps "is the demand real" for "will the three of them stop knifing each other this time," and the section you just wrote is a long answer to that second question. Structural demand on top of undisciplined supply is just a longer cycle with a bigger top, which is close to where you land anyway. The re-rating needs the supply side to behave, and that's the one actor your own archive keeps convicting.
The problem with Samsung is that it's a Chaebol with lots of governance issues. Many have tried to split it up, unsuccessfully. It is doomed to trade at a heavy discount forever. Yes, it will be the most profitable firm this and next year, but don't hope the multiple is going to reach TSMC or even Micron levels.
All the technical points aside, the section on MU's earnings call history should be required reading for analysts to understand why you ALWAYS independently verify what management says.
if u are inquiring about samsung, I'd be cool to know if anyone has an idea why the preferred shares are trading at an ever increasing discount to the commons despite this being historically unusual. None of the usual explanation seem all that conclusive. anyway, entertaining piece as always.
In Korea, preferred shares do not have voting rights, in exchange for preferred rights to dividends and higher dividend (typically 1% higher). So the lower value reflects the lack of control preference shares have compared to common shares.
But it's a relic of the past - none of the newer, younger companies have this class of preferred shares. I think it became popular in the 80s and 90s (before Asian financial crisis), as retail investors hunted for extra dividend yield.
Oh I’m well aware, but the discount used to be small or even a premium to the commons. Now it’s a very, very sizable one. Ideally Samsung would just retire the preferreds or at least include them in the stock buyback.
And as a former Samsung preferreds owner, this made me so mad that I ended up selling Samsung for Hynix and micron.
HBM has already been replaced by LPDDR5x in inference and this actually furthur increased the shortage in DRAM market. Fun fact: Kioxia already developped SSD with Optical IO as a lab project, so CXL with Optical IO must be under development in somewhere else.
Hbm hasn't been replaced. You use both because hbm is more expensive to stack and issues with temperatures, so you have SOCAM, stacked lpddr5x surrounding the CPU for a kv cash offload when the hbm gets full.
Once those temperatures get fixed (sk hynix IHMB for example) then you can stack them higher
I appreciate the framework here, and the PHY analysis is spot-on—the HBM4 qualification chaos at 11 Gbps proves bump capacitance is a real architectural limit. But calling HBM a "mistake" underweights the switching costs already baked into the ecosystem.
Optically-disaggregated LPDDR requires more than Nvidia solving photonics in 2-3 years. Hyperscalers need to rearchitect inference stacks, memory controllers get redesigned, and thermal/power budgets revalidated across every accelerator generation. That's not a 7-year transition—it's 12-15 years before 90% displacement, especially when HBM4E and hybrid-bonded HBM5 buy incumbents another 4-5 years.
My read: HBM is a local maximum, not a mistake. It's the best solution given today's packaging and interconnect constraints, even if it's not the global optimum. The question isn't whether optical disaggregation is better—it clearly is—but whether the transition happens fast enough to justify "mistake" versus "best bad option we had for a decade."
Why stack the DRAM chips on top of each other, so you need TSVs?
Rotate them 90 degrees, so the edge (the part that wires connect to if you’re using wire bonding) touches a line of contacts on a controller/aggregator base chip with multiple such lines.
All the density of HBM without the TSVs. This is what the winners are going to do…
HBM can still be a huge winner here, but assuming customers will always accept high costs and supply limits forever is a big ask. Big buyers usually look for a way around bottlenecks.
This is a wonderful article imo. Very reflected and spot on. Short term demand for HBM is real and it will intensify but in order to sustain current valuation levels, demand must convert into durable economics in the long-run as well. There is not a single company in the world that managed to build its economic moat on technological leadership and kept it for decades. Technology will evolve, competition and supply will normalize and so will HBM stocks - exactly as you said.
7747/9747 if you have access to HKEX on IBKR, can buy 50k for 100k notional. I assume you’ve already ruled out $EWY and $DRAM for whatever reasons, but throwing them out there in case you haven’t.
This guy is slowly becoming washed up. None of this article actually covers what is the purpose of HBM in agentic AI systems? What is the pJ/bit for non HBM DRAM is. How much BW does different LLM models need dense vs MoE and plot this across active parameters. If he starts doing analysis from first principles, he might realize why there may still be upside for memory and storage stocks from here. He also seems to have stopped posting his account update this time, coz optical stocks are getting whacked and this guy is all in into optics. This guy is going to get cooked with that portfolio in the next 12 months coz optical isn’t 2027 story.
You set the trap in the history section and then walked into it on purpose, which is the fun part. "This time is different" is the phrase you flag at every Micron peak, and the agentic re-rating thesis is that same line in a better suit. Telling them apart comes down to being precise about where the cyclicality actually lives.
It was never really a demand story. Every prior cycle had real, growing demand too. the cycles came from supply racing to meet it and overshooting, and your own roast names the culprit each time: Samsung doesn't cut. So "agentic demand is permanent" can be entirely true and still not break the cycle, because what breaks the cycle is supply discipline, and that's the variable that has never held.
Which makes the price/book to price/earnings re-rating less a bet on demand durability than a bet on supply discipline finally sticking. The bull case quietly swaps "is the demand real" for "will the three of them stop knifing each other this time," and the section you just wrote is a long answer to that second question. Structural demand on top of undisciplined supply is just a longer cycle with a bigger top, which is close to where you land anyway. The re-rating needs the supply side to behave, and that's the one actor your own archive keeps convicting.
The problem with Samsung is that it's a Chaebol with lots of governance issues. Many have tried to split it up, unsuccessfully. It is doomed to trade at a heavy discount forever. Yes, it will be the most profitable firm this and next year, but don't hope the multiple is going to reach TSMC or even Micron levels.
All the technical points aside, the section on MU's earnings call history should be required reading for analysts to understand why you ALWAYS independently verify what management says.
if u are inquiring about samsung, I'd be cool to know if anyone has an idea why the preferred shares are trading at an ever increasing discount to the commons despite this being historically unusual. None of the usual explanation seem all that conclusive. anyway, entertaining piece as always.
That would be very good to know indeed, I can only guess because they are not all in on HBM and not as much of a Memory pure-Play than the others?
he's not talking about samsung relative to others. he's talking about samsung preferred vs samsung common shares
In Korea, preferred shares do not have voting rights, in exchange for preferred rights to dividends and higher dividend (typically 1% higher). So the lower value reflects the lack of control preference shares have compared to common shares.
But it's a relic of the past - none of the newer, younger companies have this class of preferred shares. I think it became popular in the 80s and 90s (before Asian financial crisis), as retail investors hunted for extra dividend yield.
Oh I’m well aware, but the discount used to be small or even a premium to the commons. Now it’s a very, very sizable one. Ideally Samsung would just retire the preferreds or at least include them in the stock buyback.
And as a former Samsung preferreds owner, this made me so mad that I ended up selling Samsung for Hynix and micron.
Etf
Etf only buys common shares and inflows this year is crazy.
Also local retail prefers common
Lets see will samsung pay the bonus in preference shares as well as common
slowly becoming M. Burry of Memory stocks :)
HBM has already been replaced by LPDDR5x in inference and this actually furthur increased the shortage in DRAM market. Fun fact: Kioxia already developped SSD with Optical IO as a lab project, so CXL with Optical IO must be under development in somewhere else.
Hbm hasn't been replaced. You use both because hbm is more expensive to stack and issues with temperatures, so you have SOCAM, stacked lpddr5x surrounding the CPU for a kv cash offload when the hbm gets full.
Once those temperatures get fixed (sk hynix IHMB for example) then you can stack them higher
I appreciate the framework here, and the PHY analysis is spot-on—the HBM4 qualification chaos at 11 Gbps proves bump capacitance is a real architectural limit. But calling HBM a "mistake" underweights the switching costs already baked into the ecosystem.
Optically-disaggregated LPDDR requires more than Nvidia solving photonics in 2-3 years. Hyperscalers need to rearchitect inference stacks, memory controllers get redesigned, and thermal/power budgets revalidated across every accelerator generation. That's not a 7-year transition—it's 12-15 years before 90% displacement, especially when HBM4E and hybrid-bonded HBM5 buy incumbents another 4-5 years.
My read: HBM is a local maximum, not a mistake. It's the best solution given today's packaging and interconnect constraints, even if it's not the global optimum. The question isn't whether optical disaggregation is better—it clearly is—but whether the transition happens fast enough to justify "mistake" versus "best bad option we had for a decade."
So you wrote an article refuting HBM, but ended up buying Samsung and Koru?
I’m not under NDA.
Why stack the DRAM chips on top of each other, so you need TSVs?
Rotate them 90 degrees, so the edge (the part that wires connect to if you’re using wire bonding) touches a line of contacts on a controller/aggregator base chip with multiple such lines.
All the density of HBM without the TSVs. This is what the winners are going to do…
really ignorant and feel even more ignorant after listening and reading. But my view is $Dram etf is best way to go. Well at least for me. Thanks
Good read.
HBM can still be a huge winner here, but assuming customers will always accept high costs and supply limits forever is a big ask. Big buyers usually look for a way around bottlenecks.
You still holding your $SMTC?
This is a wonderful article imo. Very reflected and spot on. Short term demand for HBM is real and it will intensify but in order to sustain current valuation levels, demand must convert into durable economics in the long-run as well. There is not a single company in the world that managed to build its economic moat on technological leadership and kept it for decades. Technology will evolve, competition and supply will normalize and so will HBM stocks - exactly as you said.
EWY LEAPs as an alternative?
RE: Positions
7747/9747 if you have access to HKEX on IBKR, can buy 50k for 100k notional. I assume you’ve already ruled out $EWY and $DRAM for whatever reasons, but throwing them out there in case you haven’t.
Neither you or the COATUE guy has ever heard of KSM. It works across VMs too, it is even indicated for that.
That will be used even more going forward. This shortage will not be solved only by the way of fabs.