SK Hynix and Samsung are committing $590 billion to build the next generation of AI memory. The plan covers four new factories, a packaging center, and fifteen years of next-gen R&D. And it arrives at the worst possible time for buyers of PCs, smartphones and graphics cards.
Key Takeaways
- South Korea’s MOTIE ministry announced a $590 billion combined plan.
- SK Hynix and Samsung together hold close to 80 percent of the global HBM market.
- Memory prices are expected to climb 40 to 50 percent in Q3 2026 and keep rising into 2027.
Have an AI Sum Up This Article
ChatGPTThe Korean bet, in numbers
South Korea’s Ministry of Trade, Industry and Energy laid out a plan that mobilizes more than half a trillion dollars over the next years. Out of the $590 billion headline, 800 trillion won goes to four new factories in the country’s southwest region.
A separate 81 trillion won envelope is reserved for a dedicated packaging center. Packaging is the stage where memory chips are assembled with GPUs and processors. It has become the critical bottleneck for high-bandwidth memory, or HBM, which powers every modern AI data center.
The plan also commits 30 trillion won to next-generation R&D, spread over fifteen years. That long-horizon piece targets the memory architectures that will succeed today’s HBM, as AI models keep growing and context windows keep expanding.
SK Hynix and Samsung are the only two serious suppliers who can run this race. The Korean ministry estimates that the two players together hold close to 80 percent of the global HBM market, with Micron and a few outsiders splitting the rest. South Korea has decided to double down on a segment where it already wears the crown.
This is not the first time Seoul has put AI at the center of its industrial strategy. Samsung moved its entire Korean operations to ChatGPT and Codex last year, a sign that the country is betting on adoption as much as on production. The memory plan extends that logic on the hardware side.
A memory shortage in full swing
The plan’s timing is no accident. AI memory demand is exploding, and production capacity has not kept up. Investment bank Jefferies published projections that have rattled the entire tech ecosystem.
Per Jefferies, memory prices are expected to rise 40 to 50 percent in Q3 2026, then 30 to 40 percent in Q4. The trend continues into 2027 with another 40 to 45 percent hike, before a possible breather in 2028 when 15 to 20 percent of new capacity comes online.
For end buyers, the squeeze is already visible. Apple raised prices on Macs and MacBooks and pointed directly at memory component costs. PC makers, consumer GPU brands and smartphone vendors are expected to follow. Anything that ships with DRAM or NAND inside is going to get more expensive.
Demand priority is shifting the entire allocation. AI data centers come absolutely first, ahead of consumer electronics. Hyperscalers lock in long-term contracts with capped pricing, which leaves consumers to deal with the spot market, which is more volatile and more expensive.
This hardware inflation lands while the US giants are still building out at unprecedented scale. Meta is pouring $145 billion into its Ohio data centers, a sizable share of which goes straight to HBM memory. Each training rack consumes hundreds of gigabytes of HBM, and the queue keeps getting longer.
Also on Horizon:
- Tencent Study: AI Agents Finish 14% of Real Tasks
- Coinbase Picks Chinese AI, AI Bill Drops 50%
- Anthropic Survey: AI Does Half Your Work
Winners, losers, and what comes next
Short term, the winners are obvious. SK Hynix and Samsung are capturing demand that structurally exceeds their capacity. The HBM order books at both companies are sold out through 2027 according to multiple sector analysts, and margins are climbing with the prices.
The losers are also identifiable. PC buyers, RTX graphics card shoppers and high-end smartphone users will absorb the bill. Game studios that equip developers with high-end machines will see their hardware line item swell. Anyone planning to upgrade a Mac or laptop in H2 has every reason to buy now.
Medium term, the question is timing. The four new Korean factories will not reach production for 18 to 24 months. Until then, the squeeze stays. The plan will only ease the market starting in 2028, and only if AI demand stabilizes. Nothing today suggests it will plateau.
Geopolitically, South Korea locks in a strategic asset. Washington has been pushing for two years to diversify suppliers, but no credible alternative has emerged. Micron is scaling up but stays well short of Korean HBM volume. China is excluded from the high-end segment by US export controls.
One unknown remains. If the AI bubble deflates faster than expected, the Korean plan could land in overcapacity exactly when it ramps. The clouds around true model profitability, hyperscaler margins and enterprise adoption could flip the math. But for now, Seoul is making a massive bet that demand will hold.
Follow the story on Horizon.
