HWNEWS
Seoul is funding AI independence, training and running its own most advanced AI models on home turf rather than depending on US labs. The chip money still flows to Nvidia. SK Hynix, Korea's memory chip champion, ends up on the wrong side of the bill.
Frontier-model access is now at the mercy of a handful of US labs and Washington. Fable 5 was temporarily banned by the USG this summer. GPT 5.6 and Astra were similarly delayed. Sovereign AI is the answer: a country training and running its own frontier models on domestic hardware rather than renting from labs whose access can be throttled from the other side of the Pacific, per the SemiAnalysis newsletter.
Korea has the most concrete public answer. In mid-2026 the Ministry of Science and ICT selected five teams (SKT, Upstage, LG AI Research, Naver, and NC AI) to pretrain Korean foundation models, per News1. Each team is getting roughly 1,000 GPUs in 2026, and the program narrows to two teams by 2027 for the final sovereign model build-out. SKT ranked first in the ministry's sovereign AI scoring, with the Korea Development Institute evaluation as the public record for the team-selection decision.
Decoupling from US frontier-model providers does not mean decoupling from US chip suppliers. Korea is publicly committing to the first, and the trillion-dollar capex still routes to Nvidia's GPUs, per the SemiAnalysis analysis. That leaves Korea's own memory-chip champion, SK Hynix, on the wrong side of the capex split.
Nvidia is the compute partner of record. On July 24, 2026, SK Group and Nvidia expanded a strategic partnership across AI factories and next-generation memory. Naver, Nvidia, and Brookfield have announced a national AI factory buildout in Korea. On August 10, 2026, Nvidia, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR formalized an AI compute infrastructure financing platform to mobilize over $500 billion of third-party capital. That platform is the financial backstop for sovereign-AI builds, and Korea is funding its own AI stack on credit supplied by the same vendor that is selling the GPUs.
HBM, or high-bandwidth memory, is the type of DRAM (dynamic random-access memory) that stacks memory chips vertically to feed data to accelerators at high speed. It is the most valuable slice of the memory market, and SK Hynix is the volume leader. The SemiAnalysis analysis draws the structural point: training-compute demand concentrates at the GPU layer, where the capex lives. SK Hynix's high-bandwidth memory sits inside those GPUs, but the value capture is upstream, and Korea's own memory-chip champion ends up supplying a US vendor's accelerator rather than a domestic one. Sovereign AI can be a national project and a foreign-supplier project at the GPU layer.
HBM is a growing market, and a HBM-led recovery in inference or a different training build could rebalance the split. The Moonshot license shows the open-weight side of the trend tightening in parallel: any model-as-a-service business making over $20 million a year has to negotiate a separate agreement to serve Kimi K3, per the newsletter. Open-weights is the fallback sovereign AI is meant to make redundant, and the fallback is narrowing.
Korea's program is a trillion-dollar proof that a country can fund its own AI stack and still send the silicon bill to a US supplier. The next sovereign-AI announcement will answer the same question at a smaller scale: where the capex actually flows, and which economy is the host of record at the chip layer. The interesting question is not whether the next country tries. It is whether any of them changes the routing.