South Korea Just Locked In $950 Billion in AI Chip Deals. Companies Not at the Table Face Years Without Supply

There is a detail in the $950 billion announcement that landed in San Francisco on Friday that is easy to miss beneath the headline number: all three major suppliers of high-bandwidth memory — SK Hynix, Samsung, and Micron — had their entire 2026 production sold out before the summit even started. Meaningful new capacity is not expected until 2027 or 2028. What the agreements signed on Friday represent, therefore, is not just the largest single-day semiconductor supply commitment in history. They are the formalisation of a supply constraint that will define which AI companies can build what infrastructure for the next two years, and which cannot.

South Korean President Lee Jae-myung convened a high-stakes AI Summit at The Midway in San Francisco on July 25, 2026, bringing together the heads of NVIDIA, OpenAI, Anthropic, and Broadcom alongside the chairmen of Samsung, SK Group, Hyundai, and NAVER. When the summit concluded, Korea’s presidential chief of staff for policy, Kim Yong-beom, announced $950 billion in semiconductor and AI infrastructure commitments: $750 billion from SK Group and $200 billion from Samsung Electronics. The world’s AI hardware supply chain for the rest of the decade had just been substantially allocated.

The SK Group Agreements: $750 Billion

SK Hynix and NVIDIA: $500 Billion-Plus

The anchor agreement in the $950 billion package is a long-term partnership between SK Hynix and NVIDIA that industry sources and the Korean presidential office described as exceeding $500 billion in value. Under the agreement, SK Hynix will supply high-bandwidth memory to NVIDIA at scale and co-develop next-generation AI memory solutions specifically optimised for large language model training, AI agent workloads, and physical AI applications. The two companies will, in SK Hynix’s own statement, ‘co-develop and optimize next-generation AI memory solutions, including HBM, to meet evolving infrastructure demands ranging from large language model training to physical AI.’ That co-development clause is as significant as the supply commitment: it means SK Hynix’s HBM roadmap from HBM4 through the as-yet-unannounced HBM5 generation will be shaped, in part, by NVIDIA’s requirements rather than being developed in isolation. The partners go back through several prior HBM generations, as covered in our earlier reporting on NVIDIA’s South Korea AI infrastructure deals across SK Hynix, telecom, and NAVER. Friday’s agreement formalises and dramatically scales that relationship into the largest committed supply partnership either company has publicly announced.

SK Telecom’s 2-Gigawatt Data Center

A separate but connected announcement confirmed that SK Telecom will build a 2-gigawatt data center using NVIDIA’s Vera Rubin chips and SK Hynix’s HBM4 high-bandwidth memory. Two gigawatts is an extraordinary power budget for a single data center project — for comparison, NAVER’s GAK Sejong expansion announced the same week targets 200 megawatts, one-tenth of the SK Telecom commitment. The Vera Rubin GPU generation, NVIDIA’s planned successor to the Blackwell platform, is not yet in mass production, which means SK Telecom’s 2GW data center commits the company to a capital investment that will reach its full deployment in the timeframe when Vera Rubin silicon is in volume supply — likely 2027 and beyond. SK Group Chairman Chey Tae-won also disclosed at the summit that OpenAI had asked SK to supply chips directly to the ChatGPT maker, describing the scale of compute OpenAI requires as enormous. The exact terms of any SK-OpenAI supply arrangement were not disclosed.

Samsung Electronics: The $200 Billion Broadcom MOU

The Samsung component of the $950 billion announcement is the $200 billion memorandum of understanding with Broadcom covering HBM4, HBM4E memory supply, sub-2nm foundry manufacturing at Samsung’s Pyeongtaek campus, and 2.3D and 2.5D advanced packaging. This agreement was detailed in full earlier this week when Samsung and Broadcom confirmed it jointly. What the summit announcement adds is context: the Samsung-Broadcom MOU is one component of a coordinated Korean semiconductor diplomacy effort, not an isolated commercial transaction. Samsung Executive Chairman Jay Y. Lee attended the summit alongside President Lee, providing visible corporate endorsement of a government-level commitment that both the Korean administration and Korea’s private sector are aligned behind.

Anthropic’s Supply Deals: A Significant Confirmation

One of the summit’s underreported disclosures came from Anthropic CEO Dario Amodei, who confirmed that Anthropic had signed supply deals with both Samsung Electronics and SK Hynix without providing financial details. Amodei did not comment on a Reuters report suggesting Anthropic was exploring the possibility of designing its own chips, saying only that the company was ‘continually examining options and using a range of chips as part of a diversified computing strategy.’ The SK Group Chairman separately noted that Anthropic was building its own computing power without giving details. That Anthropic — which faces its own AI infrastructure expansion following the resolution of the June 2026 export control episode — has secured committed supply from both major Korean HBM producers before 2026 production is fully allocated confirms the company is among the beneficiaries of the supply relationships formalised at the summit.

The Geopolitical Dimension: Korea’s Semiconductor Diplomacy

President Lee held separate meetings before the summit with NVIDIA CEO Jensen Huang, OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, and Broadcom CEO Hock Tan. The summits were attended by the executive chairmen of Samsung and SK Group, the executive chairman of Hyundai Motor Group, and NAVER founder Haejin Lee — effectively the heads of Korea’s four largest technology and industrial conglomerates, arriving in San Francisco together under a presidential visit that was simultaneously a trade mission, a diplomatic event, and a semiconductor supply-chain negotiation. South Korea’s government described the country as ‘an irreplaceable, trustworthy production base’ for AI semiconductors. The framing reflects a deliberate positioning: in a US-China AI competition where semiconductor supply chains have become instruments of geopolitical leverage, Korea is asserting that its position — home to the world’s two largest HBM producers and essential to both American and allied AI infrastructure buildouts — gives it a role as a trusted production partner that no other country can currently replicate.

What the Supply Shortage Means for Everyone Else

The summit’s most consequential strategic fact was stated plainly in post-event analysis: with all three major HBM suppliers — SK Hynix, Samsung, and Micron — running at full capacity through 2026, the commitments signed in San Francisco have allocated a substantial portion of global HBM supply through the rest of the decade to a specific set of companies. NVIDIA, Broadcom, and their hyperscaler customers — Google, Meta, Amazon, Microsoft — are now in long-term committed supply relationships with the producers of the chips their AI infrastructure depends on. Companies that did not have a seat at the summit, or whose supply relationships are less formally committed, face a multi-year hardware constraint that capital alone cannot resolve in the near term. Meaningful new HBM production capacity from new fabs currently under construction is not expected until 2027 or 2028. The connection to broader AI infrastructure energy and capacity pressures is direct: the AI data center energy and resource demand surge that has been building throughout 2025 and 2026 is now also a memory supply constraint story, with the same few chipmakers controlling a bottleneck that every major AI infrastructure programme in the world depends on.

What Happened to the $576 Billion in June Announcements

Friday’s $950 billion announcement does not replace Korea’s June 2026 semiconductor cluster announcement — it layers on top of it. In June, the Korean government announced over $576 billion in state-backed projects covering semiconductor clusters, physical AI, and data centres. The San Francisco summit announcements are private-sector supply commitments between Korean chipmakers and US technology companies, representing a distinct capital flow from the government-backed cluster investments. Together, they describe a Korean AI semiconductor strategy operating at two levels simultaneously: building domestic manufacturing infrastructure with government capital, and securing long-term supply relationships with the US companies whose AI products drive the demand that justifies that infrastructure.

What Happens Next

The supply commitments formalised in San Francisco will take years to fully execute. HBM4 mass production is underway at Samsung and progressing at SK Hynix; HBM4E production is in the sample phase at Samsung and approaching mass production at SK Hynix. The Vera Rubin GPU generation that SK Telecom’s 2-gigawatt data center will use is in development for volume availability in 2027. The most immediate commercial effects will be felt in the pricing and availability of HBM allocation for companies outside the committed supply agreements: with production sold out through 2026 and the largest buyers locked into long-term partnerships, the secondary market for HBM capacity is effectively closed until new fabs come online. The San Francisco AI Declaration, which President Lee introduced at the summit, is expected to be followed by formal diplomatic and trade framework documents that convert the commercial commitments into a multilateral AI infrastructure partnership structure.

Why It Matters

The $950 billion in commitments signed in San Francisco on July 25, 2026 marks the point at which the AI semiconductor supply chain became formally committed rather than transactionally managed. For the companies in the room, the agreements provide multi-year supply certainty for the chips their AI infrastructure requires. For the companies not in the room, the agreements create a structural disadvantage that will take years and billions in new manufacturing capacity to close. And for Korea, the summit is the clearest statement yet that the country’s semiconductor leadership is not merely a commercial asset but a geopolitical one — a position of structural indispensability in the global AI infrastructure buildout that Korea’s government is now actively managing as a strategic resource.

Sources

AFP via The Jakarta Post, July 26, 2026. Reuters wire, July 25, 2026. Gulf News, July 25, 2026. TechTimes analysis, July 25, 2026. Today (Reuters/AFP), July 25, 2026. GV Wire (Reuters), July 25, 2026. Malay Mail (AFP), July 25, 2026.