What Seoul actually announced
On the last Monday of June, President Lee Jae Myung attached a figure to South Korea's AI ambitions that makes even this year's hyperscaler budgets look modest: a package of government-backed projects worth more than 576 billion dollars, anchored by 800 trillion won, about 518 billion dollars, of new fabrication plants from Samsung and SK Hynix. His framing was explicit: we must secure the core elements of AI faster than any other country.
The concrete pieces: four to six new front-end fabs in the country's southwest, starting on the site of Gwangju's current military airport. Samsung has selected Gwangju and is pulling its separate Yongin cluster forward by about seven years; SK Hynix is still finalising its southwestern site and accelerates its own Yongin plans by roughly twelve years. Around them sit an advanced packaging cluster in the Chungcheong region and 550 trillion won of AI data centre investment planned by 2029.
Why it matters: capacity is a date, not a headline number
Why it matters: the announcement is denominated in won, but the only unit that matters to a buyer is time. The new southwestern fabs are expected to contribute DRAM output by the mid-2030s, and analysts around the announcement do not expect price relief before 2028. Meanwhile the market you buy in today looks like this: second-quarter 2026 contract prices rose about 58 to 63 percent for DRAM and 70 to 75 percent for NAND as capacity shifted to AI servers. A half-trillion-dollar answer that arrives in the 2030s does nothing for a 2027 hardware refresh.
Yes, but: even the 2028 date assumes demand stands still, and this package itself makes sure it will not. The same plan that adds supply also funds 550 trillion won of new AI data centres by 2029, plus three to five more from SK Group alone, and every one of them consumes the exact memory the fabs are meant to free up. Supply and demand were announced in the same press conference. The net price path for the next two years is unchanged: high.
The bottom line for European buyers
The bottom line: treat memory as a dated budget line, not a passing spike. Lock configurations and volumes for 2027 now through framework agreements, standardise on fewer device models, and extend refresh cycles through the peak rather than paying it. A UK or EU company planning a fleet renewal should price it at today's contract levels plus headroom, not at a hoped-for 2027 correction.
There is also a risk-register entry in this news. Europe's own June package, Chips Act 2.0, targets logic, legacy nodes and sovereignty infrastructure; commodity DRAM and NAND are not part of any European plan. After this buildout, two companies on one peninsula will control an even larger share of the memory every European product depends on. You cannot diversify away from Korean memory; you can only contract around its price cycle and name the concentration in your risk register.
Read next: Why Is AI Raising the Price of Hardware You Never Bought? · The First Serious Challenge to Nvidia Is Not a Better GPU. It Is the Inference Layer.
