What are your thoughts on the fact that HBM takes at least 3x the clean room space as DRAM, and therefore that HBM capacity growth will cannibalize DRAM capacity?
That is true, and it’s the reason DRAM supply is so tight right now!
However, when manufacturers jump from 60% to 85% HBM yield, their effective output roughly doubles on the exact same clean room footprint.
Combine that massive yield increase with the new factory capacity coming online between 2027-2029, and total effective supply can still outgrow demand despite HBM taking up more physical space.
Yes, but if you assume DRAM yield is already at 95%+, then any conversion of clean room space from DRAM to HBM would reduce overall memory supply, even if you assume HBM eventually gets to equivalent yields as DRAM.
We aren't just restricted to existing clean rooms though. Samsung, SK Hynix, and Micron all have significant new factory capacity scheduled to ramp up between 2027 and 2029.
Right now, because HBM yields are low, manufacturers have to over-allocate and eat into extra legacy wafers just to hit their HBM volume targets. Once HBM yields hit 85%+, they get significantly more HBM output per wafer. This means they can fulfill AI demand without needing to continually eat into legacy DRAM clean room space at such an aggressive rate.
There is a “risk” that each generation of GPUs (usually 2yrs cycle) suck up 1.5x or more memory for same compute GW. Thats your main risk scenario vs your thesis in my view.
Nice work. It’s always good to challenge the prevailing narrative.
There’s a couple of things I question though. Those HBM yields don’t look right. Rule of thumb HBM3e takes 3x die of DDR5. HBM4 will be closer to 4x. TSV will remain the killer as we increase stacks.
The other one is your DRAM bit growth. This year customer forecasts are way north of 50%+ and same again next year. Supply side will struggle to reach 25%.
I don’t disagree with you, memory is still a cyclical business, just a much better one 😁
My gut is that, in time, demand drops off for a while and this is much more likely than supply catching up.
Thank you for reading and for the comment. It’s exactly the kind of debate I was hoping to start.
You make a fair point regarding the die penalty for HBM3e and HBM4, but my yield assumptions are focused on the improvement curve of sellable output over a multi-year time period, rather than the absolute die size. Even factoring in the die penalty, if a manufacturer like Samsung improves their HBM yield from 60% in 2026 to 85% by 2028, that creates a 2.4x effective output multiplier on their existing lines.
TSVs are the bottleneck today, but history shows us that 18-24 months of debugging by world-class engineering teams usually resolves these manufacturing issues.
You are right about growth regarding the immediate term as well. That is the exact reason that conditions are so tight right now and why the market is pricing in peak margins. But my thesis is focused on the 2028+ window. By that point, compounding demand at 50% becomes devastating to hyperscaler CapEx budgets. In my 2028 base case, I model total bit demand growth moderating to a healthy, but more sustainable, 18% YoY.
Meanwhile, effective bit supply catches up to 24% YoY due to yield improvements and new capacity coming online. It’s less about supply matching perpetual 50% growth, and more about demand eventually decelerating to a pace that supply can overtake.
If demand drops off for a while as buyers start demanding actual ROI on their spend, the cycle turns even faster. Whether scarcity ends because supply grows or because demand dries up, the outcome for the sector is exactly the same. Memory prices decline, and margins revert to the mean.
I agree it's a much better, more consolidated business now than it was a decade ago - but it's still a cyclical one at the end of the day. Thanks again for the great comment!
Great and reflected view! I'd like to add one argument that is actually getting stronger every time we see tech stocks explode. The question is:
Do we feel comfortable predicting technological progress 3 years from now? And most of the time a small switch causes demand to shift and stocks to crash.
That's why moat matters here - it is the switching cost for customers.
Thanks for reading👍👍 You're right that predicting tech three years out is tough, which is exactly why pricing these stocks for permanent peak margins is so dangerous!
Switching costs definitely matter in the short term for custom HBM. However, there are big limits to that moat. Big tech will use their massive scale to force standardisation and second-sourcing over time, understanding the risk of relying on a single supplier for literally anything.
A HBM moat also won't protect the entire business if the much larger DRAM and NAND markets enter an oversupply.
So short-term I can see these businesses continuing to perform exceptionally well. I think investors are just too quick to forget about the inherent cyclicality of the industry.
Thanks again for the work. Would have been easier to understand if you show the interim table between your bit demand table vs the final table, as I am not sure how you determined those margins and P/B. Also something is wrong on your dram / gb.. it cannot be so low (like really low), I suggest to check that, and dont forget to include NAND on the last table, as its 1/3 of the TAM. Kudos !!
Really appreciate the great feedback! I’ve just updated the article to include the industry margins I used for my model - hopefully that explains the path from demand to margin compression a bit better.
Also you're right about the DRAM prices. I was using gigabit pricing for DRAM and gigabyte pricing for HBM😅. I've learned my lesson about rushing articles! I’ve fixed the entire table now.
What are your thoughts on the fact that HBM takes at least 3x the clean room space as DRAM, and therefore that HBM capacity growth will cannibalize DRAM capacity?
That is true, and it’s the reason DRAM supply is so tight right now!
However, when manufacturers jump from 60% to 85% HBM yield, their effective output roughly doubles on the exact same clean room footprint.
Combine that massive yield increase with the new factory capacity coming online between 2027-2029, and total effective supply can still outgrow demand despite HBM taking up more physical space.
Hope this helped👍👍
Yes, but if you assume DRAM yield is already at 95%+, then any conversion of clean room space from DRAM to HBM would reduce overall memory supply, even if you assume HBM eventually gets to equivalent yields as DRAM.
We aren't just restricted to existing clean rooms though. Samsung, SK Hynix, and Micron all have significant new factory capacity scheduled to ramp up between 2027 and 2029.
Right now, because HBM yields are low, manufacturers have to over-allocate and eat into extra legacy wafers just to hit their HBM volume targets. Once HBM yields hit 85%+, they get significantly more HBM output per wafer. This means they can fulfill AI demand without needing to continually eat into legacy DRAM clean room space at such an aggressive rate.
There is a “risk” that each generation of GPUs (usually 2yrs cycle) suck up 1.5x or more memory for same compute GW. Thats your main risk scenario vs your thesis in my view.
Nice work. It’s always good to challenge the prevailing narrative.
There’s a couple of things I question though. Those HBM yields don’t look right. Rule of thumb HBM3e takes 3x die of DDR5. HBM4 will be closer to 4x. TSV will remain the killer as we increase stacks.
The other one is your DRAM bit growth. This year customer forecasts are way north of 50%+ and same again next year. Supply side will struggle to reach 25%.
I don’t disagree with you, memory is still a cyclical business, just a much better one 😁
My gut is that, in time, demand drops off for a while and this is much more likely than supply catching up.
Thank you for reading and for the comment. It’s exactly the kind of debate I was hoping to start.
You make a fair point regarding the die penalty for HBM3e and HBM4, but my yield assumptions are focused on the improvement curve of sellable output over a multi-year time period, rather than the absolute die size. Even factoring in the die penalty, if a manufacturer like Samsung improves their HBM yield from 60% in 2026 to 85% by 2028, that creates a 2.4x effective output multiplier on their existing lines.
TSVs are the bottleneck today, but history shows us that 18-24 months of debugging by world-class engineering teams usually resolves these manufacturing issues.
You are right about growth regarding the immediate term as well. That is the exact reason that conditions are so tight right now and why the market is pricing in peak margins. But my thesis is focused on the 2028+ window. By that point, compounding demand at 50% becomes devastating to hyperscaler CapEx budgets. In my 2028 base case, I model total bit demand growth moderating to a healthy, but more sustainable, 18% YoY.
Meanwhile, effective bit supply catches up to 24% YoY due to yield improvements and new capacity coming online. It’s less about supply matching perpetual 50% growth, and more about demand eventually decelerating to a pace that supply can overtake.
If demand drops off for a while as buyers start demanding actual ROI on their spend, the cycle turns even faster. Whether scarcity ends because supply grows or because demand dries up, the outcome for the sector is exactly the same. Memory prices decline, and margins revert to the mean.
I agree it's a much better, more consolidated business now than it was a decade ago - but it's still a cyclical one at the end of the day. Thanks again for the great comment!
Great and reflected view! I'd like to add one argument that is actually getting stronger every time we see tech stocks explode. The question is:
Do we feel comfortable predicting technological progress 3 years from now? And most of the time a small switch causes demand to shift and stocks to crash.
That's why moat matters here - it is the switching cost for customers.
Thanks for reading👍👍 You're right that predicting tech three years out is tough, which is exactly why pricing these stocks for permanent peak margins is so dangerous!
Switching costs definitely matter in the short term for custom HBM. However, there are big limits to that moat. Big tech will use their massive scale to force standardisation and second-sourcing over time, understanding the risk of relying on a single supplier for literally anything.
A HBM moat also won't protect the entire business if the much larger DRAM and NAND markets enter an oversupply.
So short-term I can see these businesses continuing to perform exceptionally well. I think investors are just too quick to forget about the inherent cyclicality of the industry.
Very well said, couldn't put it any better!
Thanks again for the work. Would have been easier to understand if you show the interim table between your bit demand table vs the final table, as I am not sure how you determined those margins and P/B. Also something is wrong on your dram / gb.. it cannot be so low (like really low), I suggest to check that, and dont forget to include NAND on the last table, as its 1/3 of the TAM. Kudos !!
Really appreciate the great feedback! I’ve just updated the article to include the industry margins I used for my model - hopefully that explains the path from demand to margin compression a bit better.
Also you're right about the DRAM prices. I was using gigabit pricing for DRAM and gigabyte pricing for HBM😅. I've learned my lesson about rushing articles! I’ve fixed the entire table now.
Thanks again!