Sunday, August 30th 2026

SK hynix CEO Says Memory Shortage Will Last Through 2030

SK hynix expects the current memory shortage to persist through the end of 2030, with no clear signs of a slowdown yet, according to CEO Kwak Noh-Jung. Speaking after the company's groundbreaking ceremony for its new Indiana packaging facility, Kwak said any eventual slowdown would likely be the result of a gradual easing rather than the sharp price crashes memory markets have previously seen, per Reuters. SK hynix had previously flagged tight supply lasting through 2028 for commodity DRAM, so this latest comment pushes that prediction out further.

Kwak's reasoning is that memory products have moved away from being purely commodity parts. AI-driven demand has pushed customers toward customized DRAM and HBM configurations, which makes future demand easier for manufacturers to forecast and less prone to the abrupt oversupply swings that caused previous crashes.
The Indiana facility itself won't add wafer capacity, since it's a packaging and testing site rather than a fab. Wafers will still be produced in South Korea and shipped to the US for HBM4E packaging, with cleanroom operations starting in the second half of 2028 and volume production targeted for the third quarter of 2029. For consumers, the takeaway is that DRAM and NAND prices are unlikely to meaningfully ease anytime soon, with PC, GPU, and console memory costs expected to stay elevated well into the next decade.
Source: Reuters
Add your own comment

56 Comments on SK hynix CEO Says Memory Shortage Will Last Through 2030

#2
yfn_ratchet
We've had this same announcement from memory maker exeuctives like five times by now, yeah? And the number kept getting bigger. Everyone involved in this farce just keeps spouting bigger numbers hoping that they'll somehow manifest by their say-so. Meanwhile anyone not nose-deep in their own sewage is gleefully waiting for the other shoe to drop.
Posted on Reply
#4
john_
Probably 2032+
Posted on Reply
#5
ncrs
Just remember that for each GB of HBM destined for hallucination machines they could've produced at least* 3 GB of DDR5 according to Micron:

(source: STH coverage)
* at least since newer generations of HBM will consume even more...
Posted on Reply
#6
zenlessyank
I don't mind riding my current rig for 3.5 years. Not one bit.

Maybe a new innovation will make it out by then anyway.

I wouldn't mind a nice 28 core chip with 256 Gigs of HBM slapped on there.
Posted on Reply
#7
twin_savage
ncrsJust remember that for each GB of HBM destined for hallucination machines they could've produced at least* 3 GB of DDR5 according to Micron:
The way I look at this is that for every gigabyte of HBM, it would take (by this slide's estimates) 32GBs DDR5 to achieve the same throughput.

If I have 1000 mostly memory bound simulations I need to complete and they each take 1 hour on the HBM machine, they will take 32 hours on the DDR5 machine. I'll require 6 HBM machines to complete my workload in a week, but I'll require 192 DDR5 machines to complete that same workload in a week.
HBM is actually reducing the amount of memory in aggregate I require.
Posted on Reply
#8
remixedcat
scam altman needs chucked into nutty putty cave
Posted on Reply
#9
ncrs
twin_savageThe way I look at this is that for every gigabyte of HBM, it would take (by this slide's estimates) 32GBs DDR5 to achieve the same throughput.

If I have 1000 mostly memory bound simulations I need to complete and they each take 1 hour on the HBM machine, they will take 32 hours on the DDR5 machine. I'll require 6 HBM machines to complete my workload in a week, but I'll require 192 DDR5 machines to complete that same workload in a week.
HBM is actually reducing the amount of memory in aggregate I require.
My point is that for the silicon constraints, and that is what is limiting in this "shortage", HBM uses at least 3 times as much of it as DDR5 does.
Your PC won't use HBM so the entire bandwidth argument, albeit correct, is beside the point in the context of current consumer memory pricing.
Posted on Reply
#10
TheinsanegamerN
Well yeah, until the banks start calling on their credit, the AI companies are gonna keep pumping as much as they can. Bubbles can last for a long time, the first warnings about the 2008 subprime crisis were heard in 2004, and that was after over a decade of lending created a clear pattern. But since the banks are riding on this paying off, they will be willing to let that credit sit for years if necessary. To this day, they are still sitting on millions of cars REPO'd since 2020 with no interest in selling them, clogging auction lots and tow yards. As they say, when you owe the bank $1 its your problem, when you owe them $1 million its their problem. How about when you owe hundreds of billions?

It's not like SK has been sitting around either. They've been pushing tens of billions of investment into new facilities to expand production. While Micron is catching up SK has been the most proactive about expanding for the future since 2024.
ncrsMy point is that for the silicon constraints, and that is what is limiting in this "shortage", HBM uses at least 3 times as much of it as DDR5 does.
Your PC won't use HBM so the entire bandwidth argument, albeit correct, is beside the point in the context of current consumer memory pricing.
And his point is that to match that throughput, you would need 32GB of DRAM on a very wide bus, the 3GB made on the same assembly line wouldnt cut it, so to meet AI demands they would have to sacrifice 10x the production if they were making DDR instead of HBM, making the shortage worse.
Posted on Reply
#11
twin_savage
ncrsYour PC won't use HBM so the entire bandwidth argument, albeit correct, is beside the point in the context of current consumer memory pricing.
This is very true, but the consumer and enterprise/AI markets are all pulling from the same foundry capacity, so if the enterprise/AI workload can reduce it's overall memory capacity usage by ~10 times by switching to HBM from DDR5, that is net positive even if on a per-GB basis HBM uses 3 times the wafer capacity as DDR5.
Posted on Reply
#12
Hecate91
CEO whos part of the AI bubble and circular investment says the shortage goes on as long as they and the cartel wants it to, what a surprise.
It was 2028 before, but as long as they keep getting off with price fixing they have no reason to supply the market with DDR5, even though consumer DRAM was admitted to be more profitable.
Also SK Hynix expanding since 2024 is meaningless, especially in the AI bubble. It'll take them years to finish, and unless the AI bubble pops there's nothing to suggest any expansion won't simply go to more HBM for scam altman and his friends.
Posted on Reply
#13
starfals
Why stop at 30? Why not 2040? ;p Its fine, humanity was fine pre consoles, phones, computers, smart machines and stuff. I remember not even having power for weeks and weeks. :D
Bring back 1960 XD Woohoo!!
Posted on Reply
#14
phints
Buy nothing. Upgrade nothing.
Posted on Reply
#15
mb194dc
I can only imagine the economic wasteland they'll be globally by then.
Posted on Reply
#16
ncrs
twin_savageThis is very true, but the consumer and enterprise/AI markets are all pulling from the same foundry capacity, so if the enterprise/AI workload can reduce it's overall memory capacity usage by ~10 times by switching to HBM from DDR5, that is net positive even if on a per-GB basis HBM uses 3 times the wafer capacity as DDR5.
The problem here is the assumption that an equivalent DDR5 system could even have been constructed. You'll run into physical signalling, power and cooling constraints long before achieving the bandwidth equivalence. Theoretically you could use many smaller DDR5 dies to meet the bandwidth requirement at similar capacity to HBM. In that case HBM would still be using 3x more silicon per GB.

Next generations are going to use even more due to increasing the number of TSVs required so the area of HBM memory dies has to grow too. Additionally it's going to tap into non-memory high end processes for the base die - either by moving the HBM controller from XPU to the stack (NVHBM) and/or putting some compute directly onto the base die. Not to mention increasing the number of stacked dies.
AI/ML is bandwidth and capacity constrained, so they wouldn't reduce capacity in favor of bandwidth. They want both.
Posted on Reply
#17
azrael
While I wouldn't call the RAM crisis outright *planned* scarcity (I'm being very generous here) I would at the very least call it an immensely *convenient* scarcity. Also, that CEO's name (you know which part) is a joke in and of itself.
Posted on Reply
#18
twin_savage
ncrsThe problem here is the assumption that an equivalent DDR5 system could even have been constructed. You'll run into physical signalling, power and cooling constraints long before achieving the bandwidth equivalence. Theoretically you could use many smaller DDR5 dies to meet the bandwidth requirement at similar capacity to HBM. In that case HBM would still be using 3x more silicon per GB.
As a data point, I've got a system with both HBM2e and DDR5 on it and in this specific implementation I'm getting about 5 times the memory bandwidth as DDR with HBM:


The problem with trying to go to a very wide bus on "normal" DDR5 DRAM chips is the physical spacing between them will quickly skew the data/clocks too much without tons of redrivers jacking up the complexity/price. I know you said "die" and not "package" so I'm not exactly addressing what you said, but I wanted to throw that out there.

LPDDR looks like it's going the route that I think you're suggesting, they've adopted x128 width packages in the new iterations of the spec as opposed to DDR5's x4, x8 or x16 width packages. It wouldn't surprise me if we start seeing more high performance computers start using it despite it's LP moniker.
ncrsNext generations are going to use even more due to increasing the number of TSVs required so the area of HBM memory dies has to grow too. Additionally it's going to tap into non-memory high end processes for the base die - either by moving the HBM controller from XPU to the stack (NVHBM) and/or putting some compute directly onto the base die. Not to mention increasing the number of stacked dies.
AI/ML is bandwidth and capacity constrained, so they wouldn't reduce capacity in favor of bandwidth. They want both.
Z-angled memory was supposed to be one of the technologies that addressed the extra space traditional HBM memory uses for it's stacking... I'm not sure how successful it will be though.
Posted on Reply
#19
Shakallia
Translation : he'll do everything he can so that it lasts as long as possible.
Posted on Reply
#20
Red Hood
The dram shortage will continue cause we said so - dram cartel
Posted on Reply
#21
Totally
All those bigheaded knowitalls from the other thread who had insider knowledge and crystall balls saying prices and supply were going to normalize within the year, linking charts to show us, the ignorant and unaware of how RAM pricing is cyclical. Where are you all at? Come tell us how Sk Hynix is wrong and they don't know what they are talking about.
Posted on Reply
#22
wNotyarD
Next year he'll say it'll last through 2035.
Posted on Reply
#23
TheinsanegamerN
Hecate91CEO whos part of the AI bubble and circular investment says the shortage goes on as long as they and the cartel wants it to, what a surprise.
It was 2028 before, but as long as they keep getting off with price fixing they have no reason to supply the market with DDR5, even though consumer DRAM was admitted to be more profitable.
Also SK Hynix expanding since 2024 is meaningless, especially in the AI bubble. It'll take them years to finish, and unless the AI bubble pops there's nothing to suggest any expansion won't simply go to more HBM for scam altman and his friends.
Yep, nothing you can do, you are DOOOOOOOMED. May as well sell everything you own and go join the Amish!
TotallyAll those bigheaded knowitalls from the other thread who had insider knowledge and crystall balls saying prices and supply were going to normalize within the year, linking charts to show us, the ignorant and unaware of how RAM pricing is cyclical. Where are you all at? Come tell us how Sk Hynix is wrong and they don't know what they are talking about.
So is anything they posted on those charts wrong? Do the words of a CEO who directly benefits from high pricing offset that evidence in your mind?
Posted on Reply
#24
Hecate91
TheinsanegamerNYep, nothing you can do, you are DOOOOOOOMED. May as well sell everything you own and go join the Amish!
And there's the whataboutisms, lol.
There is nothing anyone can do unless investors stop huffing cope and realize they're throwing money into a furnace.
Unless enthusiasts plan to empty their wallet instead of voting with it, we're stuck with the memory we have for at least the next 4-5 years. Also while hoping nothing fails in the meantime, as companies find any excuses to refuse an RMA.
TheinsanegamerNSo is anything they posted on those charts wrong? Do the words of a CEO who directly benefits from high pricing offset that evidence in your mind?
I'd love to see some evidence from people who said memory pricing would be back to normal by 2028. So far that isn't going to happen when the big 3 are colluding with each other.
Posted on Reply
#25
TheinsanegamerN
Hecate91And there's the whataboutisms, lol.
Literally not a whataboutism, but hey, dont let that stop you!
Hecate91There is nothing anyone can do unless investors stop huffing cope and realize they're throwing money into a furnace.
Unless enthusiasts plan to empty their wallet instead of voting with it, we're stuck with the memory we have for at least the next 4-5 years. Also while hoping nothing fails in the meantime, as companies find any excuses to refuse an RMA.
So...you DONT want investors investing in expanding memory production?

You do realize if production isnt expanded, this shortage is gonna last a LOT longer, right?
Hecate91I'd love to see some evidence from people who said memory pricing would be back to normal by 2028. So far that isn't going to happen when the big 3 are colluding with each other.
I'd love to see a counter argument that isnt just "nah none of that matters, the collusion man, we're never gonna buy RAM again maaaan". Considering it is 2026, you cant say that those who claimed the market would normalize by 2028 are wrong. If you could, then you should be buying futures contracts instead of crying on forums.
Posted on Reply
Add your own comment
Nov 20th, 2026 12:53 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts