Do you have a source for more info? That seems odd to me - LLMs are particularly resilient to bitflips, so if anything the demand would pull on DDR5 in general, not on ECC specifically.
>There is plenty of DDR5, its not even very expensive,
My 2x16GB of DDR5-6000 that I bought for 250€ is now 650€. 2x32GB that cost 200€ last July now goes for close to 1250€. Even my shitty ass DDR4-3000 sticks from over ten years ago would be selling for close to 300€.
DDR5 is horribly expensive, the only variants that are not as expensive are low frequencies or low amounts.
No projections/information/analysis. The statement from someone with a vested interest in this demand lasting an extended period of time (including the statement that demand will continue past 2030).
I remember in the 90's there was some sort of shortage reportedly due to fires. In the end it was found the shortage was due to collusion by the manufacturers.
After the first two times, do they even need to collude?
Can't they count on each other doing the profitable thing? Especially when they PR it:
> Furthermore, focused investment in HBM is creating a virtuous cycle that improves profitability in the general-purpose memory market. As resources are diverted to HBM, the supply-demand balance for general DRAM is improving.
They know many things I don't. I can easily assume they have a specific reason to make a public statement that is within their interests. Without any reports/prospectus/booked contracts, it's just marketing/spin.
Whom exactly? Korea and Taiwan are already saturated. EU may take 5 yearrs just to approve plans formnew factory. US is interested in super fast super duper AIG, ddr5 is way too boring.
DDR5 might be boring, but the high prices are strangling the tech market, including AI. It can’t keep this up much longer. Someone needs to step up and make more. I’m surprised the USA hasn’t offered money to TSMC, Intel, or Micron to expand manufacturing under the CHIPS act or whatever to help ease this shortage.
Given the possibility of setting up new memory fab plants takes 10 years, given that data companies are spending money on power plants based on power consumption forecasts, but are not spending on memory fab plants, I take it that data companies are seeing the current memory makers as sufficient for their needs.
Meaning if the avg consumer is going to buy memory, they are going to have to fork a whole lot more money than they did before, because they will be paying the price of profit generating data companies.
The next phase of improvements will have to be in consumer software to be more memory efficient. Gone are the days of electron apps and other memory intensive systems of programming./
That's like 3 and a half years away? I can see that - fabs don't get built overnight and bubbles aren't under any obligation to pop on a timeline. Wild that 2030 is not that far away. Sometimes it feels like we never left 2020.
Clearly it’s the fabrication of RAM/HBM chips. But if the margin is so high, why can’t other fabs like CPU fabs be used? I assume they’re not as CPU prices are going down. Or are they radically different and not easily reconfigured?
Honestly surprised by the degree of skepticism about this on HN.
The rack assemblers like Dell, HPE, Supermicro, and Lenovo have 18 to 24 month backlogs on deliveries at the moment, and even the neoclouds have 6-9 month backlogs to onboard customers onto their SaaS platforms, and 12-24 month backlogs for delivering bare metal. All those require memory.
Also, 2030 is barely 3 years away from now as 2027 is around the corner. Yes we are ALL old now.
Using priors from 2015-2019 today is the equivalent of thinking your experience in 1995-1999 should inform decisions made in 2008-2013.
Those skeptical are anticipating a possible large drop in demand before that backlog clears. ie. a change in market conditions that means that many of the orders in that backlog would be cancelled or belong to bankrupt entities.
Many words have already been spilled as to whether or not that might happen, so I'll just shrug rather than adding to them.
I think that a lot of people miss key aspects of the AI boom. Even discounting the skeptics, and the bubblers, crashers, etc.
There are a bunch of things that happen in parallel to the AI boom:
a) everyone and their mother is building out capacity. For each GB of VRAM installed, you generally want at least 1 GB of RAM, if not more. So just take nvda's numbers and go from there. And that's just for the GPU servers. But you also need ancillary services around GPU farms. You need some compute for filtering, preprocessing, you need data storage, and so on. And they're all RAM hungry processes.
b) RL is giving lots of capabilities improvements for AI, but it needs lots and lots of parallel runners for scenarios. Those runners need RAM. So every bit of non-GPU capacity is also being used, and more installed.
c) With increasing capabilities comes increasing usage. Everyone is deploying "agents" and those need to run somewhere, and they take RAM (especially the badly coded ones). There've also been reports of labs / big players hoovering up every mac available, for "agentic computer control" / training / whatever. That's on top of people buying them to run "24/7" stuff a la claw/hermes/etc.
In all of this, RAM is the bottleneck. And for the skeptics, you don't have to take hynix's word for it. Just look at the sales for CPUs / Mobos. They're down, because building a computer now is bottlenecked at RAM.
Same. There is huge, real demand for memory and storage everywhere you look. New capacity takes 3-5 years to come online. Every vendor has completely sold out capacity and is backlogged on top of that. This 5-year estimate is actually pretty realistic to me. I think we are stuck with high prices on these components for at least the next 3 years, minimum.
At a basic level, the steps involved in setting up a fab are extremely serial in nature. While there are ways in which each step has been cost-cut in the past, amdahl's law itself limits it to a minimum of 2-3 years. If you want good yields and do it more properly, you will need 3-5 like GP said. And that is still being optimistic
Additionally, HBM requires 3+ times the capacity of DRAM, so it gets crowded out. They do command 5x margins, so the unit economics is ok. But this means that even new capacity will mostly be used to service HBM demand, and consumer or enterprise DRAM is unlikely to benefit at the same pace - it will benefit somewhat of course.
Who are those ram bottleneck skeptics you’re referring to? I cannot see them. I don’t know anyone who isn’t aware the AI rollout is responsible for years to come of bottleneck of HBM and other hardware
Just see the top of this thread. The theme of the comments is that the SK Hynix, Micron, etc CEOs are lying about this in order to collectively increase prices artificially.
Some of the backlogs apart from the "top level" one are not being viewed that positively from what I have heard. One of my football friends works at Vertiv's India office, soon transferring to Thane, they basically sell cooling infra for these hyperscalers. They have an insane amount of backlogs but because they are a "spoke" component, the backlog is super volatile. Like some power generation related delay or permit issue will randomly de-scale or cancel projects. And its really difficult to calculate risk of all that and maintain the book.
You've been here awhile, this forum is default incredibly skeptical and critical, especially when it comes to AI. It has been to their predictive detriment for many years now. AI bubble conviction is over 2 years now, $300b NVDA profits since. AI tool adoption has completely encompassed any tech savvy office and isn't going anywhere. Broken records.
Aren't they printing unlimited money from this? I'm sure the chip shortage lasts as long as they're capable of extending it without falling to competitors.
New title: SK Hynix CEO is desperate for memory chip shortage to last until 2030 or longer
Fixed it for you.
SK hynix CEO compelled by CEO obligations to paint a rosy and sustainable picture of memory chip demand.
I dont think he's wrong.
Dell is making moves to introduce server grade hardware without ECC memory.
There is plenty of DDR5, its not even very expensive, its DDR5-ECC where the prices go stratospheric.
Do you have a source for more info? That seems odd to me - LLMs are particularly resilient to bitflips, so if anything the demand would pull on DDR5 in general, not on ECC specifically.
DDR-5 is still expensive, at least until China can get more production online.
>There is plenty of DDR5, its not even very expensive,
My 2x16GB of DDR5-6000 that I bought for 250€ is now 650€. 2x32GB that cost 200€ last July now goes for close to 1250€. Even my shitty ass DDR4-3000 sticks from over ten years ago would be selling for close to 300€.
DDR5 is horribly expensive, the only variants that are not as expensive are low frequencies or low amounts.
The point is that he must say this either way. If he expected demand to decline next year do you think he would come out and say that?
Why not? The income statements can’t lie. And if those fall short of expectations then the stock market is also gonna grill you.
Why wouldn't a memory CEO go on tv and say they expect the AI bubble to pop by the end of the year?
No projections/information/analysis. The statement from someone with a vested interest in this demand lasting an extended period of time (including the statement that demand will continue past 2030).
I remember in the 90's there was some sort of shortage reportedly due to fires. In the end it was found the shortage was due to collusion by the manufacturers.
https://share.google/aimode/ewJpwCHUINOHzjBVg
It did happen in the early 2010s though https://www.crn.com/news/storage/231902425/thailand-floods-t...
Oh its probably that again
https://www.polygon.com/ram-manufacturers-sued-supply-price-...
Is the thing that’s different now that the colluders are banking on is that we are in a worldwide kleptocracy?
After the first two times, do they even need to collude?
Can't they count on each other doing the profitable thing? Especially when they PR it:
> Furthermore, focused investment in HBM is creating a virtuous cycle that improves profitability in the general-purpose memory market. As resources are diverted to HBM, the supply-demand balance for general DRAM is improving.
January 5, 2026 - SK hynix market outlook.
https://news.skhynix.com/en/2026-market-outlook-focus-on-the...
Yep
But fortunely today we know that there is huge demand due to AI
Assume the person knows something you don't.
They know many things I don't. I can easily assume they have a specific reason to make a public statement that is within their interests. Without any reports/prospectus/booked contracts, it's just marketing/spin.
Tell me when you’re looking for a new bridge, I have the best ones around
He could just have said "lasting forever and ever".
It's ram chips, the bottom of the barrel of semi manufacturing. If you get high prices for long enough, everyone is gonna jump into the market.
especially if they get away unpunished for their price fixing. https://web.archive.org/web/20260701064235/https://www.polyg...
> everyone is gonna jump into the market
Whom exactly? Korea and Taiwan are already saturated. EU may take 5 yearrs just to approve plans formnew factory. US is interested in super fast super duper AIG, ddr5 is way too boring.
Only competitor is mainlaind china!
> EU may take 5 years just to approve plans for a new factory.
Who is going to build a memory chip fab in Europe? Qimonda collapsed into insolvency 20+ years ago and Infineon has long since abandoned the market.
DDR5 might be boring, but the high prices are strangling the tech market, including AI. It can’t keep this up much longer. Someone needs to step up and make more. I’m surprised the USA hasn’t offered money to TSMC, Intel, or Micron to expand manufacturing under the CHIPS act or whatever to help ease this shortage.
As soon as they figure out how to grift money to the Trump family. But they have such a backlog, it might take a while to figure out the details.
> mainlaind china
Exactly. China wil push for it as China factories will supply Chinese AI first.
But for now can't EU regulate the RAM price? They seem to be able to do what they want.
It's great how these days it's so much harder to tell if this is a joke.
Micron makes memory in the USA.
DDR5??
dude, DDR6 was planned for 2027 in 07/2025 [1] No one is going to build factories in 2026 to produce DDR5.
[1] https://www.techpowerup.com/339178/ddr6-memory-arrives-in-20...
Given the possibility of setting up new memory fab plants takes 10 years, given that data companies are spending money on power plants based on power consumption forecasts, but are not spending on memory fab plants, I take it that data companies are seeing the current memory makers as sufficient for their needs.
Meaning if the avg consumer is going to buy memory, they are going to have to fork a whole lot more money than they did before, because they will be paying the price of profit generating data companies.
The next phase of improvements will have to be in consumer software to be more memory efficient. Gone are the days of electron apps and other memory intensive systems of programming./
That's like 3 and a half years away? I can see that - fabs don't get built overnight and bubbles aren't under any obligation to pop on a timeline. Wild that 2030 is not that far away. Sometimes it feels like we never left 2020.
What are the limiting factors here?
Clearly it’s the fabrication of RAM/HBM chips. But if the margin is so high, why can’t other fabs like CPU fabs be used? I assume they’re not as CPU prices are going down. Or are they radically different and not easily reconfigured?
Very different processes. Even setting up a new CPU design in a fab that already makes CPU dies is months of validation, for context.
Given economy of scale, on the long-term can we expect lower than 2019 RAM prices?
Honestly surprised by the degree of skepticism about this on HN.
The rack assemblers like Dell, HPE, Supermicro, and Lenovo have 18 to 24 month backlogs on deliveries at the moment, and even the neoclouds have 6-9 month backlogs to onboard customers onto their SaaS platforms, and 12-24 month backlogs for delivering bare metal. All those require memory.
Also, 2030 is barely 3 years away from now as 2027 is around the corner. Yes we are ALL old now.
Using priors from 2015-2019 today is the equivalent of thinking your experience in 1995-1999 should inform decisions made in 2008-2013.
Those skeptical are anticipating a possible large drop in demand before that backlog clears. ie. a change in market conditions that means that many of the orders in that backlog would be cancelled or belong to bankrupt entities.
Many words have already been spilled as to whether or not that might happen, so I'll just shrug rather than adding to them.
I think that a lot of people miss key aspects of the AI boom. Even discounting the skeptics, and the bubblers, crashers, etc.
There are a bunch of things that happen in parallel to the AI boom:
a) everyone and their mother is building out capacity. For each GB of VRAM installed, you generally want at least 1 GB of RAM, if not more. So just take nvda's numbers and go from there. And that's just for the GPU servers. But you also need ancillary services around GPU farms. You need some compute for filtering, preprocessing, you need data storage, and so on. And they're all RAM hungry processes.
b) RL is giving lots of capabilities improvements for AI, but it needs lots and lots of parallel runners for scenarios. Those runners need RAM. So every bit of non-GPU capacity is also being used, and more installed.
c) With increasing capabilities comes increasing usage. Everyone is deploying "agents" and those need to run somewhere, and they take RAM (especially the badly coded ones). There've also been reports of labs / big players hoovering up every mac available, for "agentic computer control" / training / whatever. That's on top of people buying them to run "24/7" stuff a la claw/hermes/etc.
In all of this, RAM is the bottleneck. And for the skeptics, you don't have to take hynix's word for it. Just look at the sales for CPUs / Mobos. They're down, because building a computer now is bottlenecked at RAM.
Same. There is huge, real demand for memory and storage everywhere you look. New capacity takes 3-5 years to come online. Every vendor has completely sold out capacity and is backlogged on top of that. This 5-year estimate is actually pretty realistic to me. I think we are stuck with high prices on these components for at least the next 3 years, minimum.
> New capacity takes 3-5 years to come online.
Why should that be set in stone? Surely there are corners that can be cut if there's so much profit on the table.
At a basic level, the steps involved in setting up a fab are extremely serial in nature. While there are ways in which each step has been cost-cut in the past, amdahl's law itself limits it to a minimum of 2-3 years. If you want good yields and do it more properly, you will need 3-5 like GP said. And that is still being optimistic
Additionally, HBM requires 3+ times the capacity of DRAM, so it gets crowded out. They do command 5x margins, so the unit economics is ok. But this means that even new capacity will mostly be used to service HBM demand, and consumer or enterprise DRAM is unlikely to benefit at the same pace - it will benefit somewhat of course.
Agree. Also lots of suppressed demand at current prices. If prices return to anything like pre-AI, demand would grow dramatically.
Who are those ram bottleneck skeptics you’re referring to? I cannot see them. I don’t know anyone who isn’t aware the AI rollout is responsible for years to come of bottleneck of HBM and other hardware
Just see the top of this thread. The theme of the comments is that the SK Hynix, Micron, etc CEOs are lying about this in order to collectively increase prices artificially.
Some of the backlogs apart from the "top level" one are not being viewed that positively from what I have heard. One of my football friends works at Vertiv's India office, soon transferring to Thane, they basically sell cooling infra for these hyperscalers. They have an insane amount of backlogs but because they are a "spoke" component, the backlog is super volatile. Like some power generation related delay or permit issue will randomly de-scale or cancel projects. And its really difficult to calculate risk of all that and maintain the book.
You've been here awhile, this forum is default incredibly skeptical and critical, especially when it comes to AI. It has been to their predictive detriment for many years now. AI bubble conviction is over 2 years now, $300b NVDA profits since. AI tool adoption has completely encompassed any tech savvy office and isn't going anywhere. Broken records.
None of those points disprove a bubble. And it’s not unusual for a bubble to take years to inflate before a correction
Memory chip shortage lasting *until China floods the market with their cheaper but rigged chips.
FTFY. At the rate China's moving on this, it won't be long till they get a stranglehold on the consumer chip market.
What’s a “rigged” chip in this context?
I took it to mean backdoored similar to the NSA's tailored access operations or the Supermicro backdoor several years ago.
It's chips that work just as well, except that they make the original poster sad because china bad.
[dead]
Aren't they printing unlimited money from this? I'm sure the chip shortage lasts as long as they're capable of extending it without falling to competitors.
That's the problem with monopolies.
They do, pretty much, their margins are going up