Lisa Su's Korea Visit Tests AMD's AI Gamble Against NVIDIA

A Second Look at Seoul

Lisa Su is expected to return to South Korea next month. The AMD CEO first visited in March, meeting with Samsung Electronics’ leadership and expanding partnerships with domestic AI firms. Now, roughly seven months later, she may sit down again — this time at an October 8 government-organized symposium on AI semiconductors in Seoul.

The question is not whether this visit will happen. Reports say it is being finalized. The question is what it signals about AMD’s strategy in a market dominated by one company: NVIDIA.

The Symposiun That Matters

The October 8 event is organized by the Ministry of Science and ICT, led by Vice Prime Minister Bae Kyung-hoon. Attendees are expected to include representatives from three Korean NPU companies: Revibel, DeepX, and Mobilint. These are not household names globally, but they represent South Korea’s attempt to build a domestic alternative to foreign AI chip supply chains.

If Su attends, it would be the first time AMD’s CEO has appeared at a Korean government-hosted semiconductor forum. That alone is notable. It signals that AMD takes the Korean NPU ecosystem seriously enough to put its leader in the room.

Why NPU Companies Matter

An NPU — a neural processing unit — is a chip designed specifically for AI inference, the task of running trained models rather than training them. NPUs are efficient at specific workloads. They are not general-purpose accelerators like GPUs. But for certain AI applications, efficiency matters more than raw flexibility.

Korea’s NPU firms are betting that there is room for specialized chips alongside AMD’s broader GPU and CPU offerings. That bet aligns with AMD’s own direction. In July, AMD and South Korea signed a memorandum of understanding to develop heterogeneous AI infrastructure — essentially combining AMD CPUs and GPUs with domestic NPUs in a single computing environment. The goal is to let workloads run on whichever chip is best suited for the task, rather than forcing everything through NVIDIA’s CUDA stack.

If that vision materializes, it would be a direct challenge to NVIDIA’s dominant position. NVIDIA sells complete systems — hardware, software, and ecosystem. AMD’s approach is different: it offers components that can be mixed with other suppliers’ products. That is either a strength or a liability, depending on who you ask.

The CUDA Problem

NVIDIA’s moat is not just its chips. It is CUDA, the software platform that developers have built their AI tools around over more than a decade. Switching away from CUDA is not simply a matter of buying different hardware. It requires rethinking entire development pipelines.

AMD has been working on ROCm, its open alternative to CUDA. Progress has been real but uneven. Some models run well on ROCm. Others do not. The ecosystem is younger and less mature than NVIDIA’s.

South Korea’s heterogeneous computing bet could be one way around the CUDA problem. If NPU vendors build their software stacks to interoperate with AMD’s hardware and software, developers might not need to choose between NVIDIA and AMD exclusively. They could run some workloads on AMD GPUs, some on Korean NPUs, and manage the split through a shared infrastructure layer.

That is ambitious. It is also potentially disruptive if it works.

Samsung Looms in the Background

Samsung Electronics remains the dominant force in South Korea’s semiconductor industry. Its memory business is global. Its foundry business is competitive but still catching up to TSMC. In AI chips, Samsung has made quieter moves, including its Xeon Phi collaboration with Intel and internal efforts in custom silicon.

Su met with Samsung’s leadership in March. If she returns in October, a second meeting is possible. That would be significant. Samsung and AMD have a long history in CPU partnerships — Samsung was one of AMD’s early manufacturing clients for Ryzen and EPYC processors. A renewed focus on AI infrastructure cooperation could deepen that relationship.

But Samsung is also a potential competitor in AI chips. The dynamics are complicated. Samsung’s Foundry Division wants to attract AI chip designers to its process technology. AMD already manufactures its CPUs and GPUs through external foundries, primarily TSMC. Samsung would prefer AMD to shift some of that volume in-house. Whether AMD sees that as advantageous depends on several factors, including TSMC’s capacity constraints and pricing.

The US-China Dimension

This visit cannot be understood without considering the broader geopolitical context. The United States has tightened export controls on advanced semiconductor technology to China. South Korea, home to Samsung and SK Hynix, is caught between complying with American restrictions and maintaining business relationships with Chinese customers.

AMD’s partnership with South Korea fits into a larger pattern of US-aligned semiconductor supply chain consolidation. The July MOU between AMD and the South Korean government includes provisions for establishing an AMD AI research presence in Korea. That is effectively an extension of AMD’s engineering footprint into a allied jurisdiction — a move that reinforces the technological alliance between the two countries while subtly narrowing the options available to competitors.

China is aware of this dynamic. Beijing has been accelerating its domestic AI chip development through firms like Huawei’s Ascend division. The South Korean NPU ecosystem, if it matures, could serve as a counterweight to both NVIDIA’s dominance and China’s self-reliance push — but only if the technology proves viable and commercially competitive.

What Happens Next

Several uncertainties remain. It is not yet confirmed that Su will attend the October 8 symposium. Industry sources told local media that details are still being coordinated. Even if she attends, the scope of announcements is unclear. The July MOU sets a framework, but frameworks do not ship products.

What is clear is that AMD is investing political and executive capital in this relationship. Su has now visited Korea twice in under a year. She has signed an MOU with the South Korean government. She has expanded partnerships with Korean AI companies. This is not casual diplomacy.

The litmus test will come later. If the heterogeneous computing platform described in the MOU reaches production readiness, AMD gains a differentiated offering in the AI chip market. If it stalls, the visits and agreements look like gestures without substance.

For South Korea, the stakes are equally real. The country wants to reduce its dependency on foreign AI infrastructure while strengthening its semiconductor ecosystem. A working partnership with AMD could deliver both. A failed one leaves Korea chasing another option — possibly one that is harder to find as NVIDIA tightens its grip and China closes its own circle.

The October visit is a moment to watch. Not because it will change everything, but because it reveals whether AMD is playing to win in AI or merely staying in the game.