SK hynix Commits 54 Trillion Won to Yongin Y2 and Cheongju M17 Fabs

A tower crane above a high-rise under construction, illustrating a report on SK hynix's two new memory fabs and their build schedule. Photo by Arman on Pexels

BUSINESS · 17 Aug 2026 —

SEOUL, 17 AUG 2026 — SK hynix has approved 54 trillion won, about US$38 billion, for two new memory fabs. The first cleanroom opens in December 2028 and the second in June 2029, which is the number that matters more than the investment.

Memory prices have roughly doubled this year. The supply that answers them is three years away, and anyone buying servers, phones or cloud capacity between now and then is buying into the gap.

The commitment

35.2tn wonYongin Y2, for DRAM including high-bandwidth memory

19.1tn wonCheongju M17, for NAND including enterprise SSDs

Feb 2027M17 groundbreaking; first cleanroom December 2028

Jul 2027Y2 groundbreaking; first cleanroom June 2029

The board approved the two projects on 7 August. Yongin Y2 covers about 1.13 million square metres and is aimed at high-bandwidth memory and next-generation DRAM. Cheongju M17 covers about 680,000 square metres and is aimed at NAND, specifically the enterprise solid-state drives and key-value cache storage that inference workloads consume.

Investment completes in April 2031 for M17 and October 2031 for Y2. Both fabs are part of a larger programme the company announced in June, which earmarks 600 trillion won for the Yongin cluster, 100 trillion for Cheongju, and another 400 trillion for a new hub in the southwest. That programme's completion date was also pulled forward, from 2045 to 2033.

Three years is the story

A fab is not a factory that can be brought forward by wanting it sooner. Land, power, water, cleanrooms and tool installation run on their own schedule, and the schedule here is explicit: ground broken in early and mid-2027, first cleanrooms at the end of 2028 and in mid-2029, investment complete in 2031.

Set that against what prices have already done. Conventional DRAM contract prices rose by a record 90 to 95 per cent quarter on quarter in the first quarter of this year and by a further 58 to 63 per cent in the second. NAND rose about 60 per cent in the first quarter and 70 to 75 per cent in the second.

Nothing SK hynix approved this month touches any of that. The decision is a statement about 2029, made in a market that has been short since 2025, and the interval between those two dates is the whole problem with memory as an industry. Capacity is decided years before it is needed and arrives years after it was wanted.

The spike is moderating, but not where it hurts

TrendForce expects conventional DRAM contract prices to rise 13 to 18 per cent in the third quarter and NAND to rise 10 to 15 per cent. After two quarters of double-digit and triple-digit moves, that reads like relief.

Read the composition instead. The moderation is coming from the consumer side, where buyers have simply stopped being able to pay. TrendForce reports that PC and smartphone customers have hit their affordability limits. This is visible in falling retail PC shipments despite rising prices, conservative production plans from handset makers, and softening demand for memory cards, USB drives, and client SSDs.

Server demand is doing none of that. Cloud providers have secured supply through long-term agreements, keeping the server segment tight. The price signal, then, is not that supply has improved — it is that one group of buyers has been priced out while another has not.

That is a redistribution of the shortage rather than an end to it, and it lands unevenly. A hyperscaler with a long-term agreement is insulated. A regional cloud operator buying at contract prices is not, and neither is a handset assembler in this region working to a fixed bill of materials.

What this has already done here

The consequences have been arriving for months, in places that did not look like memory stories.

We reported that OVHcloud raised prices by up to 87 per cent and named the memory shortage as the reason, and that the Pixel 11 arrived more expensive on the same pressure. Samsung posted a record memory quarter alongside its first mobile loss, which is the same squeeze visible on both sides of one company's accounts.

For buyers in Southeast Asia the practical read is that device and cloud pricing has a component now that has nothing to do with the vendor's margin, and that component is not going away on the strength of an announcement about 2029. Procurement planning that assumes memory costs revert next year is planning against a schedule that has been published and says otherwise.

Why the split between DRAM and NAND matters

The decision to approve one fab for DRAM and another for NAND shows where the company sees AI demand landing.

High-bandwidth memory sits next to the accelerator and has been the visible bottleneck. The NAND commitment is the less obvious half. Inference at scale reads and caches enormous volumes of data, and the enterprise drives that hold it have become a critical demand line of their own, not just an afterthought to the GPU.

That is the same argument Kioxia made in taking an optical SSD to customer trials, and it is why storage keeps appearing in stories that look like they should be about compute. It is also worth noting who else is close to this decision: an investment vehicle financed by SK hynix is now Kioxia's largest shareholder, which puts one company's capital on both sides of a NAND capacity race.

What we could not establish

The wafer capacity of either fab. Floor area is disclosed and output is not, and floor area does not translate to wafers without knowing the process nodes and the tool set, so it is not possible to say how much supply this adds or what share of the market it represents.

The announcement leaves several questions open. It does not specify how the 54 trillion won will be funded — whether from cash flow, debt, or state support. It is also unclear if the larger 1,100 trillion won programme is a firm commitment, what share of Y2's output will be HBM versus conventional DRAM, how the accelerated 2033 target affects funding, or if any customers have pre-committed to the new capacity.

What to watch

Groundbreaking dates are the first test, and February 2027 comes first. Fab schedules slip routinely, and a delay announced before construction starts would say the company had read the demand differently within six months of committing.

Then watch whether Samsung and Micron respond with their own approvals. Three suppliers building at once is how every previous memory cycle turned from shortage into glut, and the timing of those decisions relative to this one is the leading indicator that matters.

Finally, watch the fourth-quarter contract prices. If consumer segments keep moderating while server stays tight, the two-tier market becomes structural rather than a phase, and buyers here will be negotiating in the tier they cannot leave.

Eva Chin’s Verdict Business

Thirty-eight billion dollars is the headline and the dates are the story. First cleanroom December 2028 at Cheongju, June 2029 at Yongin, investment complete in 2031. Memory prices have roughly doubled this year — DRAM contract prices up ninety to ninety-five per cent in the first quarter and another fifty-eight to sixty-three in the second, NAND up about sixty then seventy to seventy-five — and nothing approved this month touches any of it. That interval is the memory industry's permanent condition: capacity gets decided years before it is needed and arrives years after it was wanted. Now the part I would read carefully, because it is being reported as relief. Third-quarter forecasts are thirteen to eighteen per cent on DRAM and ten to fifteen on NAND, which is a sharp deceleration. But look at where it comes from. Consumer buyers — PCs, handsets, memory cards, client SSDs — have hit their affordability limit and stopped absorbing increases. Server demand has not moderated at all, because the cloud providers locked supply in under long-term agreements. So the price signal is not saying supply improved. It is saying one group of buyers got priced out and another did not. That is the shortage being redistributed, and in this region it lands on whoever buys at contract prices without a long-term agreement — regional cloud operators, handset assemblers working to a fixed bill of materials. The other detail worth holding is that SK hynix approved a fab for each half of the business. High-bandwidth memory next to the accelerator is the visible bottleneck; the NAND commitment says inference storage has become its own demand line rather than an afterthought behind the GPU. Kioxia is making the same argument with optical drives — and a vehicle financed by SK hynix is now Kioxia's largest shareholder, which puts one company's money on both sides of that race.

Sources & cross-checks

Primary: SK hynix — for the total of 54 trillion won split as 35.2 trillion won for Yongin Y2 (DRAM, including high-bandwidth memory and next-generation products) and 19.1 trillion won for Cheongju M17 (NAND, including enterprise SSDs and key-value cache storage); groundbreaking in July 2027 for Y2 and February 2027 for M17; first cleanrooms in June 2029 and December 2028; investment completion in October 2031 and April 2031; floor areas of about 1,130,000 m² and 680,000 m²; and the June 2026 mid-to-long-term programme of 600 trillion won for Yongin and 100 trillion won for Cheongju with target completion accelerated from 2045 to 2033.
Also primary: TrendForce, 3 July 2026 — for third-quarter contract price forecasts of 13–18 per cent on conventional DRAM and 10–15 per cent on NAND Flash; the attribution of the moderation to consumer markets reaching their affordability limit and to a higher comparison base; and the segment detail on rising PC retail prices against falling shipments, conservative handset production planning, weak television and set-top box DRAM, sluggish demand for USB drives and memory cards, and OEM inventory constraints on client SSDs.
Also reported: The Korea Herald and SiliconANGLE — for the board approval on 7 August 2026 and the dollar equivalent of roughly US$38 billion. ⚠ The prior-quarter price moves cited here (DRAM up 90–95 per cent in 1Q26 and 58–63 per cent in 2Q26; NAND up about 60 per cent then 70–75 per cent) are drawn from reporting of earlier TrendForce surveys that we did not read in the original.
Our earlier reporting: The consequences already visible are OVHcloud's price rise of up to 87 per cent, the Pixel 11 price increase and Samsung's record memory quarter alongside its first mobile loss. The storage-demand argument is Kioxia's optical SSD trials, and the ownership overlap is the SK hynix-financed vehicle now top of Kioxia's register.
What could not be established: ⚠ The wafer capacity of either fab. Floor area is disclosed and output is not, and the two do not convert without the process nodes and tool set. Also unestablished: how the 54 trillion won is funded between cash flow, debt and any state support; whether the June programme's larger totals are committed or indicative; the share of Y2 output intended for high-bandwidth memory rather than conventional DRAM; whether the accelerated 2033 target changed the funding profile; and whether any customer has pre-committed to the capacity.
Method and limits: Compiled 17 August 2026. The argument that the cleanroom dates matter more than the investment figure, the reading of third-quarter moderation as consumer buyers being priced out rather than supply improving, the description of the result as a redistribution of the shortage that lands on buyers without long-term agreements, and the interpretation of a NAND fab approval as evidence that inference storage has become its own demand line are RECATOOLS analysis rather than claims by SK hynix or TrendForce. This article reports corporate and market disclosure and is not investment advice.