TL;DR:
- Korea’s June 29 plan puts over $550B toward new memory fabs — a slice of a much larger $3.1T multi-decade Samsung/SK spending plan.
- Most new capacity targets high-margin server DRAM and HBM, not everyday PC/laptop memory.
- Micron is running its own parallel expansion; China’s CXMT is growing fast but not yet server/HBM-ready.
- Real relief isn’t forecast before late 2027-2028 — and some analysts warn of a glut once capacity lands.
- For now, DRAM, NAND, and HBM stay tight, and surplus enterprise memory/storage holds its value.
image credits: samsung.com
On June 29, South Korea’s government stood with Samsung and SK Hynix to unveil the largest industrial investment plan in the country’s history. Reporting on the announcement puts a precise number on the new capacity being built right now: $518 billion for four new memory fabs in southwestern Korea, split two apiece between Samsung and SK Hynix, plus another $52 billion for a dedicated HBM (high-bandwidth memory) packaging hub. That’s roughly $570 billion aimed specifically at new DRAM and NAND capacity.
That $570 billion is the concrete, near-term slice of a much bigger picture, and it’s worth keeping the two straight. Samsung and SK each separately announced their own long-range corporate investment plans the same day, reported to total roughly $1.7 trillion for Samsung and about $1.4 trillion for SK, both stretching to the mid-2030s and covering existing fab operations and non-memory chip lines alongside new capacity. JPMorgan’s own aggregation of the full package, circulated the day of the announcement, puts the combined total at $3.1 trillion. Most of that figure is not new memory capacity specifically — it’s each company’s entire multi-decade spending plan, as estimated by JPMorgan and Korean financial media rather than stated outright by Samsung or SK themselves. The $570 billion above is the part that’s cleanly and specifically new fab construction.
Does doubling Korea’s memory capacity actually change who wins, or how tight supply stays for the rest of us? JPMorgan analyst Jay Kwon called the broader plan the start of a “Mega Investment Era.” The detail in his note is a better place to look than the headline figure.
How the longer corporate plans break down, per JPMorgan’s estimates
JPMorgan’s note lays out how Samsung’s and SK’s full multi-decade totals — not just the new-fab package above — are allocated. These are 10-year-plus plans running through the mid-2030s, and they mix new construction with continued investment in facilities that already exist, so treat the following as estimates rather than official line items.
SK Group’s roughly $710 billion memory allocation (within its broader $1.4 trillion total) breaks down further:
- ~$390 billion for its Yongin fab, accelerating a ramp-up previously expected to stretch into the 2045-2047 window to roughly 2033-2040
- ~$65 billion to expand NAND flash production in Cheongju
- ~$260 billion toward additional semiconductor facilities, potentially in the southeast
Samsung Electronics’ semiconductor-specific allocation runs to roughly $1.36 trillion within Samsung Group’s broader $1.7 trillion total:
- ~$1.07 trillion for continued investment in Samsung’s Yongin semiconductor operations, spanning both existing and new lines
- ~$260 billion, reportedly, for a potential new hub in Gwangju
- ~$36 billion for HBM back-end packaging
Kwon estimates 60-70% of the combined total goes to front-end wafer equipment, 20-30% to infrastructure and cleanroom construction, and the remainder to back-end packaging.
That’s real, broad DRAM and NAND capacity — this isn’t a plan that’s secretly all HBM. But the strategic logic behind it is to win long-term supply contracts with the big cloud providers, which pulls new capacity toward the highest-margin corners of the market: server-grade DRAM and HBM. Commodity memory for ordinary PCs and low-end servers benefits from the rising tide, but it isn’t the point of the exercise. And deep-tech buildouts like this don’t move on political timelines: fabs announced today take years to reach volume no matter how urgent the shortage feels right now.
Who actually controls the market today
Counterpoint Research’s own Q1 2026 tracking gives an approximate snapshot of where things stand before any of this new capacity comes online (different trackers publish slightly different splits, so treat these as directional rather than exact):
| Segment | Samsung | SK Hynix | Micron | China (CXMT/YMTC) | Others |
|---|---|---|---|---|---|
| DRAM | ~38% | ~29% | ~22% | ~8% (CXMT) | — |
| NAND | ~29% | ~18% | ~13% | ~13% (YMTC) | Kioxia ~14%, Western Digital/SanDisk ~13% |
| HBM | ~21% | ~58% | ~21% | not shipping at volume | — |
(NAND shares in particular vary a bit by quarter and by tracker, partly due to reporting changes since Western Digital and SanDisk’s split — the rough shape of the market matters more here than the exact decimal.)
Micron already holds roughly a fifth of both the DRAM and HBM markets, so it’s not a bystander here. CXMT’s DRAM share, while real and growing fast, still trails the top three by a wide margin, with essentially no HBM volume yet.
Micron is running its own expansion, aimed at a different problem
Micron isn’t sitting out the capacity race. Its fiscal 2026 capex is tracking around $27 billion, and every quarter of fiscal 2027 is guided to remain above $10 billion.
It’s also quadrupling DDR4 output at its Manassas, Virginia fab, backed by a $2 billion-plus expansion. But that capacity has a narrow target: automotive, defense, and industrial customers running long product lifecycles. It isn’t meant to restock general enterprise or consumer channels.
New Idaho fabs aren’t expected to begin production before 2027, and startup dates for projects this size commonly slip. That’s a multi-year clock similar to Korea’s, and it’s weighted toward the same AI-adjacent, specialty demand.
China’s capacity is growing fast — but it’s not yet competing where it matters most
CXMT is the fastest-growing DRAM supplier in the world right now. But Chinese memory makers are currently focused on smartphone and domestic-market demand, mostly lower-end consumer applications. Shifting into DDR4 or more advanced DRAM would require redevelopment and requalification, which makes a near-term supply impact unlikely, in that account.
Server buyers demand extensive qualification and reliability testing, and that process takes years, not quarters. CXMT’s rise changes the long-run competitive picture. It isn’t yet a substitute for Korean or US capacity where the current shortage bites hardest: server DRAM and HBM.
So when does supply actually loosen up?
TrendForce’s own reporting has been fairly consistent on timing: current industry forecasts point to capacity additions staying constrained through 2026 and into the first half of 2027, with tight supply persisting into 2028. Pricing is expected to hold in a “T-shaped” pattern at elevated levels until tightness gradually eases — projections, not guarantees, and the exact quarter keeps shifting a bit with each new forecast round.
Here’s one way to understand why that timeline looks the way it does. One detailed bottom-up model from SemiAnalysis tracks planned wafer additions at all four major suppliers through 2028 — Samsung, SK Hynix, Micron, and CXMT combined, including every fab in this announcement and China’s fastest capacity ramp on record. Under that model’s own baseline assumptions, DRAM stays undersupplied by a high-single-digit percentage in 2026, widening to the low-to-mid-teens in 2027. Those specific figures are one firm’s modeled output, not an industry consensus, and they’d shift if AI demand or fab execution comes in faster or slower than assumed — but the direction (undersupply persisting, and worsening before it improves) shows up across multiple forecasters, not just this one.
Part of the explanation isn’t that too little capacity is being built. It’s where that capacity goes. SemiAnalysis’s model projects server DRAM and HBM combined will account for well more than half of the entire DRAM market by the end of 2027, a much larger slice than in past cycles. HBM is also more wafer-hungry than it looks from bit counts alone: various industry estimates put its wafer consumption per gigabyte at roughly three to four times that of ordinary DRAM, depending on the generation and stack height. Every wafer redirected to HBM is one less wafer making commodity memory for a laptop or a low-end server — one mechanism analysts point to for why capacity announcements keep landing years ahead of actual price relief.
There’s a real chance this resolves faster, or slower, than any model assumes. Memory has swung between severe shortage and glut multiple times over the past two decades, usually because a wave of capacity ordered during a shortage lands right as demand cools. If AI infrastructure spending slows before this round of fabs reaches volume, the same capacity built to fix today’s shortage could just as easily produce tomorrow’s oversupply. Neither outcome is a certainty. What forecasters seem to agree on is that relief isn’t arriving in the next few quarters.
What this means for the secondhand and surplus market
The relief windows the industry itself is forecasting run from late 2027 into 2028, and even those estimates are contested and subject to revision. Between now and then, server DRAM, HBM, and the broader DDR4/DDR5 pool that competes with them for fab capacity looks set to stay structurally tight, per current industry forecasts.
That has a direct read-through for anyone holding retired servers or decommissioned memory modules, and many categories of enterprise-grade SSDs: that hardware is worth more today than it was a year ago in most reported pricing, and quite possibly worth more now than it will be once this capacity wave actually lands.
If you’re an IT manager or procurement lead planning a data center refresh or a larger hardware liquidation in the next year, current forecasts argue for selling into today’s market rather than waiting — most industry projections don’t put a price drop before late 2027 at the earliest. If you’re sitting on a smaller stack of surplus DDR4 or DDR5 modules and drives from a decommissioned project, the same logic applies at a smaller scale. It’s also worth watching 2028 closely, since the same dynamics explaining today’s shortage are exactly what several analysts warn could tip into a glut once enough capacity finally lands at once.
Either way, the practical move is going to a bulk RAM buyer rather than listing pulled DIMMs one at a time, especially for a full pallet of retired modules. The same logic applies to drives — a buyer set up to purchase SSDs and hard drives in bulk alongside the memory saves a second round of sourcing and shipping. Given where the timeline sits, that’s worth doing now rather than waiting on a price drop that isn’t on the near-term forecast.