Semiwiki on LinkedIn Semiwiki on Facebook Semiwiki on Twitter Register / Log in 
Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/threads/dram-hbm-supply-demand-balance.25208/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/EmailDomainReplace] => 1000010
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2031070
            [XFI] => 1060170
        )

    [wordpress] => /var/www/html
)

DRAM/HBM supply-demand balance

O

Ole Hoejlund

Member
I see a lot of bold claims about future supply and demand growth in the DRAM/HBM market, including confident predictions about when supply will catch up with demand. I understand why people are trying to answer that question, I would very much like to know the answer myself. But most public analysis I come across, whether on Seeking Alpha or elsewhere, seems to be guesswork built on a few isolated facts.

Are there any serious analyses that actually do the hard work on the supply side: bit-supply modelling based on fab-by-fab capacity, expansion plans, HBM versus conventional DRAM allocation, node progression, yields, equipment constraints, packaging constraints, and so on?

And are there similarly thorough demand-side analyses that model AI accelerator growth, server memory content, HBM stack requirements, conventional DRAM demand, drone demand, and the likely elasticity or substitution effects?

My current gut feeling is that most analysts are underestimating demand growth and overestimating how quickly bit supply can expand, especially once HBM mix, yield, advanced packaging, node-transition complexity and equipment constraints are taken seriously. I would not be surprised if we do not see anything resembling real balance before 2029 at the earliest.

But that is only a hunch, so I would be very interested in either:

- links to genuinely rigorous work on this
- quick takes on the supply-demand balance from this forum of experts with high quality gut feelings
 
The link you provides has good current information. The three volume DRAM players plus CXMT. Nanya also is mentioned.

One thing that limits growth is fundamental. In the past, DRAM could avoid big capital investment because bit growth due to shrinks provided more output in existing facilities. Those days are done, basically. There is no way to easily shrink DRAM below 10nm.

So, what comes next is very speculative and explosive, basically an inflection point in the industry. Like 3D NAND was 13 years ago in the NAND flash space.
 
benb said:
The link you provides has good current information. The three volume DRAM players plus CXMT. Nanya also is mentioned.

One thing that limits growth is fundamental. In the past, DRAM could avoid big capital investment because bit growth due to shrinks provided more output in existing facilities. Those days are done, basically. There is no way to easily shrink DRAM below 10nm.

So, what comes next is very speculative and explosive, basically an inflection point in the industry. Like 3D NAND was 13 years ago in the NAND flash space.
Click to expand...
I’m on the same page as you, benb. We’ve had an upward demand shock at possibly the worst time imaginable: prior years was low CAPEX and the Industry was just starting to transition to a paradigm without the density lever to balance things out.

As you say, the only way forward is a massive wspm capacity build out, which the supply chain have a hard time delivering at the magnitude demanded without the traditional density scaling. The result is massive price increases to try to rebalance demand and supply.

And my personal take on this: What we’re are seeing is not solely demand destruction to make demand meet supply. Much is deferred demand, as a lot of the new unserved demand is price inelastic (AI, data centers, auto, robotics, drones) while much of the unserved old demand is short term price elastic short-term, but medium/long-term inelastic (post-pone the replacement of you smartphone 1-2 years and the willingness to pay rises with the frustration with your outdated phone.
 
@Ole Hoejlund

I do reports and presentations on this all the time. Some comments I typically make: (call me to discuss)

1) No one really saw this coming. In August 2025, most analysts were predicting moderate growth for at least 6 months to a year. Then companies went all in with their Capex forecast on AI servers in September 2025
2) There is not going to be a material change in supply until late 2027. and even then Memory companies are correctly saying to keep the shortage forever.
3) At some point, all companies will "Capitulate" and accept that they will not get the memory they want and come up with a new plan. That is in progress now, but it is a slow path. then we have a "new normal", efficiency, shared memory architecture,
4) All companies are currently telling memory companies that they are willing to pay an insane amount for hardware. At some point this could change if said companies dont make money. hyperscalers have tons of money and income (AI companies have money but limited income)
5) In all supply chains when you are really short, it looks like the shortfall will last 5 years. Then in a few months, that changes to 4 months. you do not ever know real demand until you are not short. so the "we are short until 2030" is nonsense... BTW the supply agreements are tied to market price and are not rigidly enforceable (unless you think its a good idea to sue your largest customer) and cover 30% of bits shipped.

So the next steps are (IMHO)

see when hyperscalers accept the new normal (2027?).... bit prices grow <15% per quarter
See whether lots of capacity comes on line or whether it stays tight in 2027 (Hope they keep it tight)
At any time listen for hyperscalers to say "we are in digestion phase" or "we have multiple constraints". this could occur at any time, due to pushback on AI datacenters or financial returns.

The real question for me is: Companies have demonstrated they are willing to spend 5x more on memory bit prices than I thought. Can you hold onto that going forward? I would recommend memory companies only add 10% bit capacity per year forever and keep the shortage going unless the hyperscalers take partial ownership of Fabs and split the margins.


there is other data on margins by memory type, % used for DDR vs HBM, impact of Chinese companies, and details on supply agreements. too much to list here.
 
Ole Hoejlund said:
AFAIK, yes. They are planning a shift to 3DRAM in this time window. Though, IDK if that explains the less aggressive Samsung build out.
Click to expand...
4F starts in a couple years, 3D is much further out. One interesting topic, Since memory is so high margin now, companies can take a financial risk on capacity and architecture. They could build a 3D fab, delay the ramp 2 years and still be very profitable. This has never been true in the past.
 
MKWVentures said:
@Ole Hoejlund

I do reports and presentations on this all the time. Some comments I typically make: (call me to discuss)

1) No one really saw this coming. In August 2025, most analysts were predicting moderate growth for at least 6 months to a year. Then companies went all in with their Capex forecast on AI servers in September 2025
2) There is not going to be a material change in supply until late 2027. and even then Memory companies are correctly saying to keep the shortage forever.
3) At some point, all companies will "Capitulate" and accept that they will not get the memory they want and come up with a new plan. That is in progress now, but it is a slow path. then we have a "new normal", efficiency, shared memory architecture,
4) All companies are currently telling memory companies that they are willing to pay an insane amount for hardware. At some point this could change if said companies dont make money. hyperscalers have tons of money and income (AI companies have money but limited income)
5) In all supply chains when you are really short, it looks like the shortfall will last 5 years. Then in a few months, that changes to 4 months. you do not ever know real demand until you are not short. so the "we are short until 2030" is nonsense... BTW the supply agreements are tied to market price and are not rigidly enforceable (unless you think its a good idea to sue your largest customer) and cover 30% of bits shipped.

So the next steps are (IMHO)

see when hyperscalers accept the new normal (2027?).... bit prices grow <15% per quarter
See whether lots of capacity comes on line or whether it stays tight in 2027 (Hope they keep it tight)
At any time listen for hyperscalers to say "we are in digestion phase" or "we have multiple constraints". this could occur at any time, due to pushback on AI datacenters or financial returns.

The real question for me is: Companies have demonstrated they are willing to spend 5x more on memory bit prices than I thought. Can you hold onto that going forward? I would recommend memory companies only add 10% bit capacity per year forever and keep the shortage going unless the hyperscalers take partial ownership of Fabs and split the margins.


there is other data on margins by memory type, % used for DDR vs HBM, impact of Chinese companies, and details on supply agreements. too much to list here.
Click to expand...

This is valuable input, thanks!
 
MKWVentures said:
2) There is not going to be a material change in supply until late 2027. and even then Memory companies are correctly saying to keep the shortage forever.
Click to expand...

MKWVentures said:
BTW the supply agreements are tied to market price and are not rigidly enforceable (unless you think its a good idea to sue your largest customer) and cover 30% of bits shipped.
Click to expand...
From Micron earnings deck:

We are excited to announce that we have now signed 16 strategic customer agreements, or SCAs, which we expect will fundamentally transform our business model

With respect to supply, our customers are recognizing that supply shortages in memory and storage will take considerable time to improve.

Even as we expect industry supply to improve gradually in 2028, we currently do not have line of sight as to when memory supply will be able to catch up with increasing demand.

FQ3-25 Gross Margin: 84.6%!!!

FQ4-25 guidance GM: 86%; EPS $31/share ( x4 = $124/share, which is the full stock price 12 months ago!!)
 
One key point often overlooked comes from the last DRAM/NAND cycle. The only major player that truly benefited from falling memory prices was Intel, whose core business is CPUs. Lower memory costs stimulated overall system demand, indirectly boosting CPU sales. If you’re expecting Nvidia to play a similar role this time, that assumption may not hold. Nvidia and SK hynix have already formed strong strategic alignment. Meanwhile, Intel’s recent hiring of SK hynix’s former CEO could signal renewed interest in re-entering or reshaping its position in the memory market.
 
MKWVentures said:
4F starts in a couple years, 3D is much further out. One interesting topic, Since memory is so high margin now, companies can take a financial risk on capacity and architecture.
Click to expand...
We need it yesterday!!
MKWVentures said:
They could build a 3D fab, delay the ramp 2 years and still be very profitable. This has never been true in the past.
Click to expand...
So true.

But then like you had said:
MKWVentures said:
5) In all supply chains when you are really short, it looks like the shortfall will last 5 years. Then in a few months, that changes to 4 months. you do not ever know real demand until you are not short. so the "we are short until 2030" is nonsense..
Click to expand...
 
MKWVentures said:
4F starts in a couple years, 3D is much further out. One interesting topic, Since memory is so high margin now, companies can take a financial risk on capacity and architecture. They could build a 3D fab, delay the ramp 2 years and still be very profitable. This has never been true in the past.
Click to expand...
On a more positive note, elevated memory prices can have some stabilizing effects. Higher costs often translate into increased token or service pricing, which can help sustain revenue and protect jobs across the value chain. Similarly, more expensive consumer devices—like smartphones—may not be entirely negative, as they can discourage overconsumption and reduce resource waste. In that sense, pricing pressure can contribute to a more balanced and sustainable market environment.
 
MKWVentures said:
4F starts in a couple years, 3D is much further out. One interesting topic, Since memory is so high margin now, companies can take a financial risk on capacity and architecture. They could build a 3D fab, delay the ramp 2 years and still be very profitable. This has never been true in the past.
Click to expand...
SK hynix was most aggressive or active in promoting 4F2 at the VLSI conference this year. It looks like they're stopping the 6F2 architecture at 1d. The EUV is apparently a big burden, and they say 4F2 relieves this, besides providing electrical benefits. I find it hard to imagine more than 3 nodes or generations of 4F2 though. The word lines and bit lines are still getting too close, and the capacitor dielectrics are getting too thin.

Samsung had shown a 16-layer 3D DRAM at the VLSI conference, with (deliberately?) very few details. CXMT indicated 2030 for 3D DRAM at IMW; Micron indicated some time after ("over the next decade").
 
Ole Hoejlund said:
Me answering me, but I found this recent analysis from Deutsche Bank, which may be of interest to others here: https://www.macrostream.ai/articles/6a3214548f2442d721550211

Headline finding in their analysis: DRAM supply gap will persist until 2030 - maybe longer - due to massive data center demand growth and constrained supply build out.
Click to expand...
benb said:
The link you provides has good current information. The three volume DRAM players plus CXMT. Nanya also is mentioned.

One thing that limits growth is fundamental. In the past, DRAM could avoid big capital investment because bit growth due to shrinks provided more output in existing facilities. Those days are done, basically. There is no way to easily shrink DRAM below 10nm.

So, what comes next is very speculative and explosive, basically an inflection point in the industry. Like 3D NAND was 13 years ago in the NAND flash space.
Click to expand...
The report's sampling is weird. It has JHICC but not Winbond? Is macrostream.ai an AI site?
 
You must log in or register to reply here.
Back
Top