AMD Chief Executive Lisa Su landed in Seoul on July 7 for a single-day sweep of meetings with Samsung Electronics, SK hynix, and 13 Korean AI chip vendors, all centered on expanding HBM4 supply for the company's next-generation Instinct MI455X accelerator. The visit came seven months after Su's previous Seoul trip.

What did AMD actually agree to?

Samsung Electronics confirmed in March that it would supply HBM4 for the MI455X. On July 7, Su met Samsung Device Solutions Vice Chairman Jun Young-hyun at the company's Seocho office in Seoul to widen that commitment into foundry, or contract chip manufacturing. The same day, Su sat down with SK hynix President Kwak Noh-jung to expand HBM4 allocation for the same accelerator.

What is Su's partnership model?

Speaking to reporters after AMD's Gamer Day event at Dragon City in Seoul's Yongsan district, Su called the chip supply chain "highly complex." She said AMD favors three-to-five-year-plus relationships over one-to-three-year deals. "Chipmaking is a team sport," Su added. "Meeting partners frequently and regularly matters."

Memory remains the binding constraint in AI infrastructure, officials at the meetings said. With inference computing demand rising on the back of agentic AI, the industry's center of gravity is shifting from GPU-only stacks toward heterogeneous CPU, GPU, and NPU combinations tuned for cost and power efficiency.

Which Korean firms joined the roundtable?

Su hosted chief executives at the Four Seasons Hotel in Seoul's Jongno district. The 13 attendees included:

• NPU developer FuriosaAI
• NPU developer DeepX
• AI infrastructure provider MangoBoost
• Other domestic AI chip and software vendors (full list undisclosed)

Topics covered linking AMD CPUs and GPUs with Korean-made AI accelerators, participation in open-source AI, and integration with the ROCm high-performance computing software stack. Su framed AMD's position this way: "AMD's principle has always been collaboration with an open ecosystem."

What is Korea contributing?

Second Vice Minister of Science and ICT Ryu Je-myung proposed alignment across the government's three mega-projects: semiconductors, AI data centers, and physical AI. Su called the plan "exactly the right direction and an ambitious vision" and said AMD will participate.

Lee Do-gyu, head of the information and communications policy office at the Ministry of Science and ICT, said the global question now is "how to optimize CPUs, GPUs and NPUs to deliver the best performance at the lowest cost."

The ministry will run a demonstration project for heterogeneous computing infrastructure built jointly by AMD and Korean AI vendors, with the aim of turning the pilot into a reference design exportable to overseas markets.

What is the hiring plan?

AMD will open an AI center of excellence in Korea this year. Hiring targets several hundred researchers. Su proposed "world models" used in physical AI as a joint research focus at the new site. AMD recently acquired World Labs, co-founded by Stanford professor Fei-Fei Li, to expand into physical AI; the Korean research footprint will feed into that workstream.

What changes for the MI455X supply chain?

Three concrete shifts emerged from the Seoul meetings:

• Samsung Electronics: HBM4 supply for MI455X confirmed in March; July talks extended into foundry cooperation
• SK hynix: expanded HBM4 allocation discussion for the MI455X
• Korean NPUs: pilot integration with AMD's CPU/GPU stack and ROCm software platform

Su's seven-month revisit cadence is unusual for a U.S. chip CEO in a single foreign market. It tracks with AMD's effort to lock multi-year memory commitments before inference-driven demand absorbs the global HBM4 output.

Korea's manufacturing density — heavy industry, automotive, electronics — positions the country as a test bed for AI applied to factories and physical products, ministry officials said.