AINEWS
Behind a San Francisco summit's $950B in pledges, Samsung and SK Hynix become the two suppliers every AI accelerator now depends on.
AI's next growth ceiling sits in high-bandwidth memory and advanced chip packaging, where two Korean companies hold the chokepoint. A round of long-term agreements announced this week, headlined as roughly $950 billion in Korea-US commitments, is the moment Nvidia and Broadcom formally conceded that point, while OpenAI has asked SK to supply chips directly.
The combined figure mixes at least three different kinds of commitment, and treating them as equivalent obscures what actually changed.
The largest single bucket is SK Group's announced $750 billion in deals, where SK Group is the parent of memory maker SK Hynix and data-center builder SK Telecom. The headline is anchored by a partnership with Nvidia valued at more than $500 billion. That partnership covers large-scale AI data centers and next-generation memory, including a long-term SK Hynix supply deal and joint development of HBM (stacked memory chips that sit next to an AI accelerator and feed it data) for AI training, agents, and physical AI. A SK Telecom facility targeted to come online in 2027 will draw two gigawatts of continuous power, roughly the demand of a mid-sized US city, and run Nvidia's Vera Rubin platform, the company's next-generation AI accelerator succeeding the current Blackwell line, paired with SK Hynix's HBM4 memory.
Samsung's contribution is a separate MoU with Broadcom worth up to $200 billion, covering memory chips, foundry services, and advanced packaging. The two companies will co-develop Broadcom's next AI accelerator using Samsung HBM, a sub-2-nanometer foundry process (the most advanced manufacturing node Samsung can run commercially today), and 2.5D/3D packaging technology that lets multiple silicon dies behave like one larger chip. The "up to" language in the announcement, and the MoU structure rather than a signed purchase order, is why the Korean side's combined $950 billion number should be read as an envelope, not a check.
AI training runs are throttled by memory bandwidth per accelerator, and GPU count has stopped being the binding constraint. HBM4 is the next step in stacking more memory closer to the compute die. Nvidia's Vera Rubin platform and Broadcom's custom AI accelerators are designed against that memory roadmap. When a single accelerator needs multiple gigabytes of HBM4 stacked and packaged onto the same substrate, the constraint moves from TSMC's fabs to Korean memory and packaging lines.
On the same stage, SK Group Chairman Chey Tae-won said AI customers have wanted more memory than suppliers had projected, and that Nvidia's five-year memory requirement could be revised higher. Broadcom's Hock Tan said memory demand is likely to exceed supply. Anthropic, whose May 2025 funding round valued the company at $965 billion with Samsung, SK Hynix, and Micron named as critical partners, told the wire it regularly checks future chip availability. OpenAI, according to Chey, has asked SK to supply chips directly, bypassing the usual Nvidia-bundled path.
Korean President Lee Jae Myung used the summit to unveil a "San Francisco AI Declaration" on US-Korea AI cooperation and met Jensen Huang, Sam Altman, Dario Amodei, Hock Tan, and the chiefs of Samsung, SK Group, Hyundai, and Naver. A follow-on Saturday event put South Korea's National Pension Service, the world's third-largest pension fund, in the same room as US venture investors, an explicit bid to pull Korean sovereign capital into US AI infrastructure deals.
The strategic backdrop includes a Korean industrial strategy Lee announced last month, anchored by state-driven projects spanning semiconductors, AI infrastructure, and data centers. The summit deals are the foreign-customer side of that strategy, locking in demand for the same memory and packaging capacity the state is underwriting at home.
Two qualifications matter. First, summit deal totals in this category have historically overstated committed capital, especially when the same executives share a stage. The Korean $950 billion headline mixes firm purchase orders (a smaller share), MoUs (Samsung-Broadcom), and multi-year "up to" envelopes that depend on memory yields, customer uptake, and packaging capacity all hitting plan. Second, the bottleneck can shift: if HBM4 yields climb faster than expected or if a third supplier enters at scale, the pricing power Samsung and SK Hynix currently enjoy will compress.
The watch item through 2027 is the SK Telecom two-gigawatt facility, which is the first concrete capacity anchor in the new agreements and the cleanest test of whether the memory-and-packaging chokepoint is binding or breakable. If the site comes online on schedule, the chokepoint is real and the deals harden. If it slips, the headline numbers will be the first thing to get revised.