OpenAI’s new Korean partnerships landed first in the Korean and Japanese press, and the tone was blunt: memory is the constraint and Korea is the solution. JoongAng Ilbo framed SK Group’s guidance as a demand shock, quoting internal projections that “오픈AI의 DRAM 웨이퍼 수요는 월 90만장” (OpenAI’s DRAM wafer demand is 900,000 per month), adding that this is “현 HBM 생산능력의 두 배 이상” (more than double current HBM capacity). Asahi Shimbun noted that Samsung and SK Hynix “覚書に署名した” (signed memoranda of understanding), a reminder that these are nonbinding but signal intent and political cover. Korea’s Ministry of Science and ICT was explicit about geography: “수도권 외 대규모 AI 데이터센터 입지 검토” (evaluating sites for large AI data centers outside the Seoul metro). Local media are treating this like an industrial policy moment, not just a tech deal.
Seoul made the point with prices. SK Hynix jumped nearly 10 percent to a record close, Samsung Electronics added around 3.5 percent to a four-year high. On the KOSPI, semiconductors and equipment led; shares tied to data center buildouts and power infrastructure caught a bid, including Samsung C&T and shipyards on talk of floating data centers. Telcos outperformed on expected AI enterprise upsell. The won firmed modestly as foreign funds chased liquidity in mega-cap tech. The ripple was regional: Tokyo’s chip equipment complex rallied on HBM capex read-through; in Taiwan, AI server and packaging proxies gained on expectations of sustained CoWoS and HBM interposer demand. Hong Kong saw selective moves in power and data center landlords. Sentiment shifted from AI software to AI infrastructure, with Korea at the center.
The local case for Korea is straightforward. Talent density, an export-driven chip ecosystem, and vertically integrated conglomerates can compress time-to-build. Samsung and SK Group can pull memory supply, EPC, logistics, and marine engineering under one umbrella. That is precisely what OpenAI’s Stargate needs if it is to scale quickly from today’s patchwork to a standardized hyperscale footprint. As one Korean headline put it, “한국은 AI 허브가 될 조건을 갖췄다” (Korea has the conditions to be an AI hub). Adoption matters too. OpenAI executives say Korea has the most paying ChatGPT subscribers after the US, which aligns with strong enterprise IT spend and a culture of fast rollout. Government is a tailwind, not a brake. That combination makes Korea a build partner, not only a parts supplier.
Under the hood, this is about HBM and the DRAM wafers behind it. SK Hynix is the lead supplier of HBM to Nvidia; Samsung is pushing to close the performance gap as HBM3E ramps and HBM4 looms. Local press stressed that “HBM 증설이 시급” (HBM expansion is urgent). The 900,000 DRAM wafers per month cited by SK Group for 2029 implies years of staggered capex, substrate and TSV constraints, and yield learning curves. Packaging is the shadow bottleneck: even if DRAM capacity arrives, CoWoS and HBM interposers remain tight, spreading gains to foundry and OSAT ecosystems in Taiwan and Japan. Another underplayed angle is custom silicon. Both Korean memory giants supply Broadcom, which is building custom AI chips for OpenAI. That diversifies demand away from Nvidia-only cycles, but it also adds complexity to qualification timelines. LOIs are not purchase orders; watch the cadence of binding orders and capex commitments.
Samsung Heavy Industries and Samsung C&T are now part of the story. Samsung Group said the companies will explore floating data centers and floating power. In Korean coverage: “부유식 데이터센터와 전력 제어 인프라 협력 검토” (reviewing cooperation on floating data centers and power control infrastructure). That is not a gimmick. Korea’s land constraints, grid congestion near Seoul, and cooling requirements push infrastructure to coastal or offshore solutions. Floating platforms can tap seawater cooling and modular power, but they demand marine engineering, mooring, and cybersecurity investments that only a handful of EPCs and yards can execute. The phrase you see in Korean boardrooms is “전력원 확보가 관건” (securing power sources is the key). Expect LNG-to-power, grid reinforcement, HVDC, and potentially SMR pilots to surface in policy and tenders if Stargate-scale campuses are to land outside Seoul.
Politics are fused to this buildout. The Basic Law on Artificial Intelligence is slated to take effect in January, with the government setting a target for Korea to be a top-three AI nation by 2027. Local media shorthand reads: “AI 기본법 내년 시행, 3대 AI 강국 목표” (AI Basic Law takes effect next year, goal is top three AI nation). That unlocks budget, permitting streamlining, and tax incentives for strategic assets including data centers and chip capacity. SK Telecom will co-develop an AI data center in the southwest as part of Stargate Korea, accelerating growth outside the capital region. Domestic demand is not theoretical. Kakao’s roughly 600 billion won data center plan and its OpenAI tie-up show local hyperscalers want in. Enterprise adoption should lift Samsung SDS and SK affiliates on ChatGPT enterprise deployments. The risk is overheating valuations if unit economics of AI services lag infrastructure build.
Korea’s memory lead intersects with others’ strengths. Taiwan still owns the packaging bottleneck and AI server integration stack; TSMC’s CoWoS and InFO capacity expansions are the pacing item for much of the AI accelerator pipeline. Japan will capture incremental tool sales via Tokyo Electron, Nikon, and Disco as HBM capex matures, and could benefit if Kioxia or Micron leverage local fabs for HBM4E-related flows. The US remains the nexus for accelerators and custom silicon. Export controls remain a ceiling: China demand is constrained for leading-edge AI, pushing more capacity toward US and allied buyers. For Korea, the asset is coordination. Conglomerate structures can align memory, EPC, shipbuilding, and telco demand in a single programmatic relationship. That is why a letter of intent can move multiple sub-sectors at once in the KOSPI, while similar news elsewhere lifts only one ticker.
Investors should separate spectacle from sequencing. Letters of intent are not revenue; they are an option on execution. The meaningful milestones are binding wafer supply contracts with price and node detail, confirmed HBM qualification wins at Nvidia and custom chip partners, grid interconnection approvals for specific sites, and EPC awards for power and cooling systems. In Korean coverage, one line stood out: “발주와 인허가가 병목” (orders and permits are the bottleneck). That is the real timeline risk. The 2029 wafer number signals ambition, but the calendar for each 12 to 18 month capacity tranche matters more for earnings models. On the infra side, watch for first steel on a southwest-region campus and for any pilot of floating platforms. The faster those happen, the more credible the non-memory leg becomes.
English-language coverage has focused on Samsung and SK Hynix. Local markets are already leaning into second-derivative plays: Samsung C&T for EPC, Samsung Heavy for floating platforms, SK Telecom for data center JV economics and enterprise AI upsell, and power equipment names tied to HVDC, substations, and cooling. Expect interest in LNG import terminals and grid storage to rise if policymakers bless offshore or coastal buildouts. ESG scrutiny is coming. Water use, power sourcing, and coastal environmental impact will become gating issues in permitting. Japanese press have used the phrase “電力と環境の両立” (balancing power and environment) when covering hyperscale expansion, and it applies here. The companies that can show lower PUE, recycled water systems, and credible green power sourcing will win tenders faster.
Two things are being missed. First, this is not just a memory upcycle; it is a coordinated industrial build where Korea’s conglomerates can monetize design, construction, power, and operations around AI campuses. The non-memory margin pool is under-modeled outside Korea. Second, the critical path is not HBM wafer starts alone; it is power and packaging. Price in the grid and interposer bottlenecks, not just the DRAM ASP curve. The Korean press is clear: “전력과 패키징이 병목” (power and packaging are the bottlenecks). If you only own the memory pair, you are long the right theme but short the enablers. Add exposure to Korean EPC and marine engineering where order books will fill first, and watch for binding supply contracts and interconnection permits as the true catalysts from here.
