The $700B US-Korea AI deal pipeline runs through Korean high-bandwidth memory

Samsung, SK, and Nvidia announced a $700 billion deal pipeline at the San Francisco AI Summit on July 24. Most of the line items are memoranda, not firm orders. The mechanism that decides whether the announced number materializes sits in Korean high-bandwidth memory fabs, the same chokepoint that throttles every AI training run.

South Korean President Lee Jae Myung hosted the one-day investment summit, drawing Nvidia CEO Jensen Huang, Samsung Chairman Lee Jae-yong, and other U.S. and Korean tech leaders to the Bay Area. Nikkei put the announced value at $700 billion across semiconductors and data centers. Korea JoongAng Daily reported the same summit produced partnership commitments worth more than $950 billion. The gap is not a typo: the two outlets appear to be measuring different scopes, one counting firm contracts and another counting memoranda plus pipeline. President Lee framed the gathering as positioning South Korea as a "global hub" of the AI market; the Asiae outlet published what it calls the full text of the "San Francisco AI Declaration." That is the host's pitch and the host's document, not an independent verdict.

High-bandwidth memory, or HBM, is the stacked DRAM that sits next to a GPU on the same package and feeds it data fast enough to keep training or inference running. AI accelerators from Nvidia, AMD, and Google cannot operate at full throughput without it. SK Hynix (the memory arm of SK Group) and Samsung Electronics are the two firms in the world with mature, high-volume HBM production. Both are Korean, and both sit inside chaebol groups whose other subsidiaries can absorb the long-dated capital a memory build-out requires. That is why a deal pipeline nominally centered on Nvidia's compute and Samsung's foundry capacity runs through Seoul's memory fabs. The bottleneck is specifically HBM, where capacity additions are tied to multi-year fab builds, advanced packaging yields, and qualification cycles with chip designers rather than quick equipment moves.

Nikkei's phrasing is "deals worth" a combined $700 billion. Few of the line items appear to be firm purchase orders with committed financing, delivery dates, and penalties for non-performance. A memorandum of understanding is a statement of intent; the financing, offtake, and export-control approvals still need to be arranged. KEIA's analysis of the summit frames the meeting as a coordination step between Washington and Seoul on AI supply chains, not a binding treaty or a capital commitment.

Three external dependencies will compress or expand the announced trajectory. First, U.S. export controls on advanced AI chips to China: the deal flow is implicitly routed through non-China Asia, and a tightening of those rules would redirect Korean memory and foundry output to U.S. customers, which the announced pipeline seems to anticipate. Second, financing: most of the announced data-center capacity requires long-dated capital that has not been committed in dollar terms, and the cost of debt has moved up since the post-2022 AI buildout began. Third, AI demand itself: the dollar scale of the pipeline assumes the 2025-2027 training and inference run rate continues, which is a forecast, not a contract.

The pipeline is large and the timing is short. The watch items: which of the announced line items convert to firm orders in the next two quarters, how SK Hynix and Samsung allocate HBM3E and HBM4 output between Nvidia and other customers, and whether U.S. export-control posture toward Korea stays permissive through 2027. The Korea-U.S. AI deal flow is real; the headline dollar total is not yet. The mechanism that decides which one holds is in Korean memory fabs.