SK Hynix is spending $38 billion to build two new chip fabs

[Alfonso Maruccia]

The big picture: Foundry corporations have repeatedly stated that the chip supply chain crisis is not going to improve anytime soon. Adding more manufacturing plants could help partially ease the industry's pain, which is exactly what SK Hynix plans to do over the next few years. Let's just hope the AI bubble doesn't burst in the meantime.

SK Hynix just announced a massive new investment to strengthen its manufacturing capacity for both DRAM and NAND flash products. The South Korean corporation, one of the world's three largest memory manufacturers alongside Samsung and Micron, is going to spend 54 trillion won ($38 billion) to build the new "Y2" and "M17" chip fabs in Yongin and Cheongju, respectively.

The new plants are part of a larger investment strategy totaling 700 trillion won. Yongin Y2 is the second of four new fabs planned for the Yongin Semiconductor Cluster. The plant will increase SK Hynix's DRAM production capacity, with construction expected to begin in July 2027.

The first cleanroom in Y2 is expected to be ready by June 2029 and will house the equipment required to produce high-bandwidth memory solutions and "next-generation" DRAM chips. Meanwhile, construction of the Cheongju M17 plant is expected to begin in February 2027. The facility is scheduled to open its first cleanroom in December 2028, with production primarily focused on NAND Flash chips.

SK Hynix acknowledges that HBM memory is now one of the most sought-after chip formats in the world. Nvidia, AMD, and other vendors are struggling to secure enough capacity to keep developing and producing new GPUs and AI accelerators for data center hyperscalers. Analysts say that new data center plans are still on paper for the most part, while Big Tech companies continue to buy up every memory chip they can find.

SK Hynix also said that demand for NAND flash is surging rapidly. The focus is once again on hyperscalers and enterprise SSDs, which are being deployed to further expand the capacity and capabilities of AI services. NAND chips are used as cache storage in AI inference workloads, with agentic AI and physical AI applications such as robots expected to drive demand even higher.

Commenting on the newly announced construction plans, an unnamed SK Hynix official said that the company aims to become "a key partner in AI infrastructure, contributing to the stability of the global AI semiconductor supply chain."

Memory manufacturers have long been criticized for their alleged unwillingness to build new manufacturing capacity, with SK Hynix predicting that the supply chain crisis will likely peak between 2027 and 2030. The new Y2 and M17 plants will certainly increase the company's production capacity, although they should come online just as the market finally begins to settle down.

Permalink to story:

SK Hynix is spending $38 billion to build two new chip fabs

 

[ScottSoapbox]

I hate when greedy corporations take their profits and spend them on expanding supply for their customers. This will bring prices back down and then what will I complain about online?

 

[JMGuy]

I hate when greedy corporations take their profits and spend them on expanding supply for their customers. This will bring prices back down and then what will I complain about online?

Lol. The passage of time perhaps? Otherwise, you’ve certainly illustrated, rather neatly, how online public discourse routinely starves out substantive engagement with reality

 

[Anton Longshot]

I hate when greedy corporations take their profits and spend them on expanding supply for their customers. This will bring prices back down and then what will I complain about online?

Looks like you won't profit from the lower prices should those happen. They're exclusively catering to their AI server customers. Unless you're one of those customers of course.

 

[Endymio]

Memory manufacturers have long been criticized for their alleged unwillingness to build new manufacturing capacity

@Alfonso Maruccia continues his campaign of inventing facts to support a narrative. No reputable industry observer has "long criticized" memory makers for failing to build new capacity, because they've been spending huge sums to expand capacity since 2024, when the 2023 DRAM price crash ended. The problem is that these fabs take several years to come online, and demand continues to increase faster than supply.

Looks like you won't profit from the lower prices should those happen. They're exclusively catering to their AI server customers.

Who told you something so absurd? More importantly: why did you you believe it?

 

[RudyBob]

USA USA

SK Hynix’s major investments and activities in the United States include a $3.87 billion advanced packaging plant in Indiana, a $10 billion dedicated AI investment arm, and a historic $26.5 billion U.S. stock market debut.

Manufacturing and Infrastructure

Indiana Advanced Facility: Committing $3.87 billion to build an advanced chip-packaging and research facility for high-bandwidth memory (HBM) chips, with production targeted to begin in 2028.

Supply Chain Strategy: Positioning packaging and R&D operations closer to major U.S. artificial intelligence and tech customers like Nvidia.

 

[Plutoisaplanet]

AND?! How will it be powered??? I thought that was the real news for chip factories! For instance, this was the headline yesterday on TechSpot:

SpaceX and Tesla plan their own natural gas power plants for massive $16.8 billion Texas chip factory

 

[LordVile95]

Lol. The passage of time perhaps? Otherwise, you’ve certainly illustrated, rather neatly, how online public discourse routinely starves out substantive engagement with reality

It won’t bring prices down though as it’ll be producing HBM which hasn’t been in consumer products since Vega

 

[Theinsanegamer]

It won’t bring prices down though as it’ll be producing HBM which hasn’t been in consumer products since Vega

But that HBM is displacing GDDR used in AI GPUs currently.

 

[LordVile95]

But that HBM is displacing GDDR used in AI GPUs currently.

And they can’t make enough HBM so they’re building new lines to make more HBM. If you make 10 HBM. Hips per minute and have a demand of 30 per minute so you move most of your lines to make a total of 15 and then build new lines to make 25 you’re still lower than demand

 

[noel24]

I hate when greedy corporations take their profits and spend them on expanding supply for their customers. This will bring prices back down and then what will I complain about online?

I see It sligthly different from reports I've read. It allegedly started with Open AI buying out huge amounts of DDR5 wafers, while Their DataCenter build-out slowed down, actually taking dozens of month into future. They artificially created supply issue, maybe to boost Their stock value. The problem is, by the time those data centers will stand, and get their electricity supply online, AI machines will move to DDR6 for the best performing chips. What will happen with those warehouses full of uncut DDR5? Who will absorb the losses? Will consumers will be priced out of DDR6, but have the oversupply of dirt chip DDR5? Will AMD and Intel move to DDR6 supported CPUs and have Us struggle for second hand overpriced DDR5 supported CPUs and Motherboards, like They did with original 5x00X3D CPUs? Are We looking at the artificial MiddleAges in personal computing and should really abandon PCs and move to consuming nearly-dead, AI driven, Internet on smartphones and SmartTVs? Is gardening trully the future of My past-time? Does Steve Walton has a plan for that scenario?

 

[Loadedaxe]

I see It sligthly different from reports I've read. It allegedly started with Open AI buying out huge amounts of DDR5 wafers, while Their DataCenter build-out slowed down, actually taking dozens of month into future. They artificially created supply issue, maybe to boost Their stock value. The problem is, by the time those data centers will stand, and get their electricity supply online, AI machines will move to DDR6 for the best performing chips. What will happen with those warehouses full of uncut DDR5? Who will absorb the losses? Will consumers will be priced out of DDR6, but have the oversupply of dirt chip DDR5? Will AMD and Intel move to DDR6 supported CPUs and have Us struggle for second hand overpriced DDR5 supported CPUs and Motherboards, like They did with original 5x00X3D CPUs? Are We looking at the artificial MiddleAges in personal computing and should really abandon PCs and move to consuming nearly-dead, AI driven, Internet on smartphones and SmartTVs? Is gardening trully the future of My past-time? Does Steve Walton has a plan for that scenario?

I think there’s a legitimate concern buried in there, but a few different things are getting mixed together.

You’re right that OpenAI’s projected memory requirements are enormous. OpenAI itself announced that Samsung and SK hynix plan to scale advanced memory production toward 900,000 DRAM wafer starts per month for Stargate. That absolutely contributes to manufacturers believing AI demand will remain huge.

Where I think you go off a little is describing this as OpenAI buying out warehouses of DDR5 wafers to artificially create a shortage or boost its stock value. OpenAI is still privately held, and the 900,000 figure is a projected future requirement/capacity target, not evidence that OpenAI currently owns mountains of unused DDR5 wafers sitting in storage.

More importantly, AI memory demand isn't simply “DDR5 today, DDR6 tomorrow.” Modern AI accelerators depend heavily on HBM, which is specialized DRAM with very different packaging and bandwidth requirements from the DDR5 DIMMs we stick into desktop PCs. SK hynix says its new Yongin fab will produce HBM and other next generation DRAM, while its new Cheongju fab will produce NAND. These aren't $38 billion factories being built solely to crank out today's consumer DDR5 forever.

Could AI demand be overestimated? Absolutely. I've made that argument myself. If data center construction slows substantially or AI investment doesn't produce the expected returns, memory manufacturers could eventually find themselves with too much capacity.

But that's actually the part that should make PC buyers less worried about the “Middle Ages.” Semiconductor memory has historically been highly cyclical. When manufacturers build too much capacity and demand disappoints, you generally get oversupply, falling prices and production cuts...not consumers wandering the wasteland looking for a used AM5 motherboard.

And right now SK hynix is publicly saying almost the opposite...it expects memory demand to exceed supply through 2030, with 2027 potentially being particularly tight. Whether that forecast proves correct is another discussion, but there's currently little evidence that SK hynix thinks it's building billions of dollars worth of factories for memory nobody will want.

So, keep the concern about AI distorting the memory market...there's merit to that.

And gardening is probably safe as a hobby rather than our emergency replacement for PC gaming. **** I may start doing it!!

 

[Endymio]

It won’t bring prices down though as it’ll be producing HBM which hasn’t been in consumer products since Vega

How are you so reliably wrong on all issues? SK Hynix's Yongkin Y2 will produce both HBM and conventional DRAM. And even if it did simply produce HBM, that alone would reduce conventional DRAM prices, as many corporate purchasers are buying it as they're unable to procure HBM.

 

[LordVile95]

How are you so reliably wrong on all issues? SK Hynix's Yongkin Y2 will produce both HBM and conventional DRAM. And even if it did simply produce HBM, that alone would reduce conventional DRAM prices, as many corporate purchasers are buying it as they're unable to procure HBM.

I don’t think you can read.

 

[Watzupken]

Whenever corporation boast about something, people really need to take it with a bucket of salt. It is better to watch what they do versus what they say, because it is usually not in line. For example in this case, when all the memory makers are raking in historical profits, the 38 billion investment don't sound like it is matching the "overwhelming" demand. It seems to me more like they are are carefully expanding because if this AI demand crash all of a sudden, then they will flip to oversupply very quickly.

 

[Endymio]

For example in this case, when all the memory makers are raking in historical profits, the 38 billion investment don't sound like it is matching the "overwhelming" demand.

That $38B is nowhere near all of what's being spent. Samsung and SK Hynix alone are spending $518B (half a trillion dollars) in South Korea alone, and Micron is spending another $250B in the US:

Memory Capacity Production Slowly Expanding To Meet AI Industry Demand

Samsung, SK hynix, Sandisk, Kioxia and Micron report building out new DRAM and NAND flash manufacturing and packaging facilities coming on line in late 2027 or 2028.

www.forbes.com

 

[ScottSoapbox]

I see It sligthly different from reports I've read. It allegedly started with Open AI buying out huge amounts of DDR5 wafers, while Their DataCenter build-out slowed down, actually taking dozens of month into future. They artificially created supply issue, maybe to boost Their stock value. The problem is, by the time those data centers will stand, and get their electricity supply online, AI machines will move to DDR6 for the best performing chips. What will happen with those warehouses full of uncut DDR5? Who will absorb the losses? Will consumers will be priced out of DDR6, but have the oversupply of dirt chip DDR5? Will AMD and Intel move to DDR6 supported CPUs and have Us struggle for second hand overpriced DDR5 supported CPUs and Motherboards, like They did with original 5x00X3D CPUs? Are We looking at the artificial MiddleAges in personal computing and should really abandon PCs and move to consuming nearly-dead, AI driven, Internet on smartphones and SmartTVs? Is gardening trully the future of My past-time? Does Steve Walton has a plan for that scenario?

Memory tech becomes obsolete at a predictable rate so there are already plans on how and when current and future fabs will pivot when the time comes.

The most likely outcome is that consumer hardware will stay on DDR5 longer than the industry would have predicted pre RAM apocalypse, while hyperscalers will move to DDR6 as fast as the tech allows since price to performance is irrelevant until the bubble pops.