AMD's Korea Play Reshapes the AI Chip Trilateral

The Visit That Changes Everything, Quietly

Lisa Su arrived in Seoul on October 7 and did what she always does: she worked the room with a precision that makes it look casual. A morning meeting with Korean NPU companies. An afternoon at AMD Gaming Day in Dragon City. A doorstop interview. No fanfare, just the steady accumulation of relationships that no one else in the chip world is building this aggressively.

Seven months separate this trip from her first visit to South Korea in March. That initial tour produced the headlines — the handshake with Samsung Electronics CEO Lee Jae-yong at the Blue House, the MOU with Naver, the talks with AI startup Upside. What didn’t make the press releases mattered just as much: the formalization of Samsung as the priority supplier of HBM4 memory for AMD’s next-generation Instinct AI accelerators.

That detail — HBM4 — is where the real story lives.

The Memory Moat

Nvidia’s dominance in AI training isn’t just about GPU architecture. It’s about memory bandwidth. Every generation of HBM has been a bottleneck, and every bottleneck has been a moat. HBM4 raises the bar again: higher density, lower power, faster signaling. Whoever controls HBM4 supply controls the pace at which their AI chips can scale.

By locking in Samsung as its primary HBM4 partner, AMD has done something structurally significant. For years, Nvidia has enjoyed de facto first refusal on Samsung’s most advanced memory — a reality no one stated publicly because it was understood. Samsung doesn’t need to announce it. The customers know. The competitors know.

Su’s March meeting with Lee and her subsequent formal agreement shifted that implicit hierarchy. Samsung’s HBM4 output for AMD is now a public commitment. Whether Samsung can simultaneously satisfy Nvidia’s demand and AMD’s growing appetite is an open question, and one that will define the next two years of the AI chip supply chain.

Intel, meanwhile, is still hunting for a memory partner at this tier. Its foundry ambitions are years away from producing the kind of HBM-integrated packages that Samsung and SK Hynix already ship. AMD’s Korea play widens that gap further.

The Heterogeneous Gamble

What Korean media haven’t fully captured is the architecture strategy underneath all these handshakes. The July meeting in San Francisco between AMD and Deputy Prime Minister Bak Hyun-hoon produced a framework that deserves more attention: heterogeneous AI computing infrastructure that pairs AMD CPUs and GPUs with domestic Korean NPUs.

This isn’t just about selling more chips. It’s about building a platform where Korean AI startups can plug in their own NPU designs alongside AMD’s silicon — something Nvidia has never offered and likely never will. Nvidia’s CUDA ecosystem is its entire business model. It runs best when every component in the stack is designed by Nvidia. Heterogeneous computing, by contrast, thrives on openness.

AMD has been building toward this since 2020. Its ROCm open-source software platform is the answer to CUDA’s lock-in. UALink, the ultra-high-speed interconnect standard AMD co-founded with Google, Apple, and Microsoft, is the plumbing that makes heterogeneous clusters viable. Both are bets that the future of AI computing isn’t a single vendor stacking silicon on silicon but a distributed ecosystem where different processors handle different workloads efficiently.

Korea’s AI research centers, promised under the MOU framework discussed in July, would become testbeds for exactly this kind of architecture. If AMD can prove that a cluster combining its own accelerators with Korean NPUs outperforms a homogenous Nvidia stack on specific workloads, the value proposition becomes undeniable for cost-conscious buyers.

Who Wins, Who Loses

Samsung wins immediately — guaranteed HBM4 volume, a anchor customer for its most advanced memory product, and leverage against SK Hynix in the bidding war for AI memory contracts.

Naver and Upside win secondarily — access to AMD’s GPU portfolio at a time when Nvidia lead times are measured in quarters, plus a seat at a table where they help shape infrastructure policy rather than simply buying capacity at market prices.

AMD wins strategically. Every Korean partnership is a foothold in a market where Nvidia has held everything by default. Su’s three-month cadence of engagement — March, July, October — signals that Korea isn’t a side project. It’s a deliberate expansion of AMD’s ecosystem beyond the American cloud giants.

Nvidia loses the most, though it won’t admit it. The company’s entire moat depends on keeping customers inside a closed stack. Every heterogeneous computing demonstration, every MOU signed between AMD and a Korean partner, every HBM4 agreement that diverts Samsung’s output weakens that premise. Nvidia can respond with better chips or lower prices. It cannot respond with openness.

What Comes Next

The immediate deliverable is HBM4 volume ramp. Samsung needs to prove it can ship enough high-grade memory to satisfy both AMD and Nvidia without compromising yield. That timeline — late 2026 into 2027 — will set the pace for AMD’s Instinct series competitiveness against Nvidia’s next-generation GPUs.

The medium-term bet is the Korean research center network. If the AI excellence centers materialize and demonstrate real heterogeneous performance gains, they become proof points that AMD can replicate in other markets. The ROCm ecosystem needs more than software; it needs hardware diversity and real workload data. Korea offers both.

The long-term question is whether the US-Japan-Korea trilateral on semiconductors, quietly strengthening since the 2023Chip 4 discussions, absorbs AMD’s Korean partnerships or treats them as competitive friction. Washington has every reason to welcome a second source for advanced memory and a second GPU vendor in the ecosystem. Tokyo has reason to watch carefully — Japan’s own AI chip ambitions could benefit from the same heterogeneous model AMD is promoting in Seoul.

Su doesn’t need to say any of this in a press interview. She just keeps showing up. The partnerships do the talking.