Lisa Su meets 14 Korean AI chip firms — AMD bets on mixing GPUs with Korean NPUs
On October 7, 2026, AMD chair and CEO Lisa Su met Ryu Je-myung, Korea's second vice minister of science and ICT, in Seoul and held a roundtable with 14 Korean AI chip and software companies, including Rebellions, FuriosaAI and DeepX. The core agreement is to test, in Korea, 'heterogeneous' AI infrastructure in which AMD CPUs and GPUs run alongside Korean-designed NPUs in one system, and to turn it into a reference design Korean firms can sell abroad. AMD will open an AI Center of Excellence in Korea, its first in Asia, by year-end and said it will hire hundreds of researchers and engineers there over several years. It also signaled it will join the government's three AI megaprojects in chips, physical AI and AI data centers
The three lines
- Meeting — Lisa Su met Korea's science ministry and 14 local firms, incl. Rebellions and FuriosaAI, Oct 7
- Deal — test AMD CPUs/GPUs + Korean NPUs together; ROCm participation; export support
- Pledges — first Asian AMD AI research center by year-end; hundreds of hires over several years
Key questions
- Which Korean companies met Lisa Su
- **Fourteen firms spanning chips to software.** | Area | Companies | |---|---| | AI chips (NPUs) | Rebellions, FuriosaAI, HyperAccel, DeepX, Mobilint | | Data-center chips, interconnect, storage | MangoBoost, Panmnesia, FADU, Dnotitia | | AI software and platforms | Moreh, Nota AI, Lablup, Elice Group, Upstage |
- AMD AI research center in Korea
- **AMD's first Asian AI Center of Excellence, due by year-end.** | Item | Detail | |---|---| | Partners | Korean universities and research institutes | | Fields | Drug discovery, climate, energy, manufacturing and robotics | | Core task | Heterogeneous computing with AMD chips + Korean NPUs | | Location | To be disclosed at a November steering committee | | Hiring | Hundreds over several years |
- What is heterogeneous AI infrastructure
- **Different chip types in one system, each doing what it is best at.** | Chip | Role | |---|---| | CPU | Scheduling, data preparation, general work | | GPU | Training, large parallel jobs | | NPU | Running trained models (inference) cheaply |
Nvidia's biggest rival gathered Korea's AI chip startups in one room. The goal is not to replace Nvidia but to mix. On October 7, 2026, AMD chair and CEO Lisa Su met Korea's second vice minister of science and ICT, Ryu Je-myung, at the Four Seasons Hotel in Seoul, then held a roundtable with 14 Korean AI chip and software firms. It was her first visit in seven months. The centerpiece: heterogeneous AI infrastructure that runs AMD CPUs and GPUs together with Korean-designed NPUs, proven at home first and then offered abroad.
Context for global readers: Korea is the world's memory powerhouse through Samsung and SK Hynix, but in AI processors it has relied on startups such as Rebellions and FuriosaAI, which design NPUs — chips specialized for running trained AI models efficiently. Their weak spot has been software and customer references, not silicon.
1. Who was in the room
| Area | Companies |
|---|---|
| AI chips (NPUs) | Rebellions, FuriosaAI, HyperAccel, DeepX, Mobilint |
| Data-center chips, memory interconnect, storage | MangoBoost, Panmnesia, FADU, Dnotitia |
| AI software and platforms | Moreh, Nota AI, Lablup, Elice Group, Upstage |
The guest list matters. A mixed-chip system needs more than chips: software to split work across processors, fast links such as CXL to move data between them, and optimization tools. These 14 firms cover most of those layers.
2. What they agreed
| Item | Detail |
|---|---|
| Heterogeneous testbed | AMD CPUs/GPUs + Korean AI chips run together in Korea, developed into a global reference model |
| Software | Korean participation in ROCm, AMD's open-source GPU software platform |
| Exports | Linking Korean firms' products to AMD's global AI ecosystem |
| AMD AI Center of Excellence | First in Asia, by year-end. Drug discovery, climate, energy, manufacturing and robotics; location announced at a November steering committee |
| Talent | Graduate programs in AI; AMD to hire hundreds of researchers and engineers over several years |
| Megaprojects | AMD signaled participation in Korea's chips, physical-AI and AI-data-center projects |
Su said AMD will keep hiring "hundreds of specialized researchers and engineers in Korea over several years" to strengthen the country's AI R&D. Ryu said growth in AI inference is moving competition from individual chip performance to whole systems and platforms.
From 2027 the government plans a national "full-stack" AI infrastructure program combining Korean NPUs, foreign GPUs and CPUs, CXL and domestic software, run by consortia. At the roundtable, firms also proposed using next year's roughly 4.7 trillion won frontier-AI model program as a testbed for Korean NPUs.
3. Why AMD, why now
| Party | What it gains |
|---|---|
| AMD | More developers on ROCm to challenge Nvidia's CUDA; an "open" reference that Nvidia lacks |
| Korean NPU firms | A system validated alongside a global GPU vendor — a credential for overseas data-center buyers |
| Government | An "open ecosystem" not locked to one foreign Big Tech firm; a concrete sovereign-AI blueprint |
AMD's market value passed $1 trillion this year, driven heavily by data-center CPUs. As inference demand grows, splitting work across CPUs, GPUs and NPUs becomes the economical design — the same logic behind soaring valuations for inference-chip startups such as Etched.
4. What remains unconfirmed
- Reports differ on whether the ministry–AMD memorandum was signed in July or September.
- The center's site, size and exact hiring figures await the November steering committee.
- Using the frontier-AI program as an NPU testbed is an industry proposal; the government has committed only to building large Korean-NPU references next year.
- Budgets and company roles for the heterogeneous testbed were not announced.