Why Lisa Su's Late-Night Samsung Dinner Matters for AI Chips

A 3-hour dinner says more than a press release

Lisa Su didn’t come to Seoul on October 7 for a photo opportunity. She arrived at Samsung’s Seocho office at 4:44 p.m. and didn’t leave until nearly 8 p.m., having shared a meal and a strategy session with Jeon Young-hyeon, Samsung’s vice chairman and the executive who runs the company’s most critical division — its semiconductor business. The hands-on-heart answer when asked whether she likes Samsung wasn’t corporate politeness. It was an admission of something harder to say publicly: AMD needs Samsung now more than it let on.

The substance of what they discussed is only visible between the lines of Su’s carefully non-committal remarks. When pressed on foundry cooperation, Su said only that AMD and Samsung have “a very broad partnership overall.” On the future — particularly the question of whether they’d collaborate on physical AI, the robotics and edge-computing frontier — she said they would work together across “many different fields.” She said nothing about terms, volumes, timelines, or pricing.

That silence is the story.

What happened in March sets the stage

The foundation for this meeting was laid in March, when Su visited Korea and AMD signed a memorandum of understanding with Samsung covering HBM4 — the sixth generation of high-bandwidth memory — and DDR5 server DRAM supply. They also discussed the possibility of manufacturing AMD’s next-generation chips at Samsung’s foundry operations. That MOU was notable because it was the first clear signal that AMD was willing to diversify its foundry supply beyond its longstanding reliance on TSMC.

Nvidia has built its empire partly by controlling the most advanced nodes at TSMC. AMD’s strategy, articulated repeatedly by Su, has been to avoid putting all its chips in one foundry basket. But that strategy hit a wall in 2024 and 2025 when HBM capacity became the bottleneck that determined who could ship AI accelerators and in what volume.

HBM isn’t just memory. It’s the bandwidth bridge between compute and training data, and it’s the difference between an AI chip that works and one that doesn’t. Samsung and SK Hynix are the two companies that can produce HBM at scale. Micron is closing the gap but isn’t there yet. Whoever controls HBM supply controls the timeline for AI chip shipments worldwide.

The tension beneath the hospitality

Industry sources close to the negotiations told Chosunbiz that this meeting was about resolving gaps in position between AMD and Samsung. AMD needs more HBM to scale its AI processor shipments. Samsung wants more foundry revenue to compete with TSMC, which dominates the advanced-node market. Both sides want the deal. But both sides also face risk.

AMD’s risk is dependency. If it shifts a meaningful portion of its advanced chip production to Samsung, it becomes vulnerable to Samsung’s yield curves, capacity allocation decisions, and the political calculations of a government that sometimes treats semiconductors as leverage. Samsung’s risk is that it invests in capacity for a customer that might return to TSMC if the economics don’t work out — exactly what happened with other fabless designers who tried Samsung’s foundry and walked away.

The 3-hour meeting suggests they found enough common ground to move forward, even if the details remain private. Su’s public comments were calibrated to reinforce the partnership without revealing leverage points.

Physical AI changes the question

When reporters asked about collaboration in physical AI — a term that covers robotics, autonomous systems, and edge devices that interact with the real world — Su didn’t deflect. She said they would cooperate across “very diverse fields.” This matters more than it sounds.

Nvidia has staked its future not just on data-center GPUs but on the physical AI layer — robots, factories, self-driving systems, and the digital-twin infrastructure that connects them. AMD has been quieter on this front, but Su’s answer suggests the company is preparing to compete there too. Physical AI requires a different kind of silicon mix: not just raw GPU throughput, but efficient inference at the edge, memory that can keep up with sensor data streams, and foundry partners who can produce mixed-signal and analog-digital hybrid chips alongside the digital logic. Samsung can do all of that. TSMC can too, but at a cost that squeezes margins for everyone except the biggest buyers.

The Korea-U.S. dimension

This meeting shouldn’t be read solely through a commercial lens. It’s also a geopolitical signal. The United States has been pushing Korean semiconductor companies to deepen their integration with American chip designers and to reduce dependency on Chinese supply chains. Samsung and SK Hynix are American allies in this framework — they build critical infrastructure for the global AI stack, and they do it under U.S. export controls and investment incentives.

AMD’s decision to work more closely with Samsung, particularly on HBM and foundry production, aligns with Washington’s interest in building a supply chain that doesn’t run through Taiwan alone. Every billion dollars of advanced chip design that flows from AMD to Samsung’s foundry is a dollar that strengthens the Korean-American semiconductor alliance and weakens the argument that the world’s most important chips can only be made on one island.

Who wins, who loses

AMD wins if it secures reliable HBM4 supply and finds a viable secondary foundry option. That gives it pricing power with TSMC and shipping speed with its MI-series AI accelerators. It also positions the company to capture demand from customers who want a second source for advanced chips — a growing requirement as geopolitical risk makes single-source dependency uncomfortable for enterprise buyers.

Samsung wins if it converts this meeting into actual tape-outs and volume production. The company has been trying for years to win fabless design wins away from TSMC. AMD is the biggest prize in that effort. A successful AMD-Samsung foundry relationship would validate Samsung’s process technology and attract other customers who were watching to see whether the company could deliver at the cutting edge.

Nvidia loses if AMD gains traction. Not immediately — Nvidia’s software moat (CUDA) remains formidable — but over the next two to three years, every AI training and inference contract that AMD wins with Samsung-sourced chips is one less contract for Nvidia. And unlike AMD, Nvidia has no foundry alternative to leverage. It is entirely dependent on TSMC.

TSMC loses margin and narrative control. The company’s dominance is based on the assumption that no fabless designer can realistically qualify a second foundry for its most advanced products. If AMD proves otherwise — and if Samsung’s yield data supports the claim — the assumption cracks. Other customers will test the water. TSMC will respond with pricing pressure, but it can’t offer what Samsung can: a relationship with a U.S. government that views Korean foundry capacity as strategically preferable to Taiwanese concentration.

What comes next

The details will emerge slowly. Samsung’s next earnings call will mention foundry order trends. AMD’s investor materials will reference supply diversification. Neither will quantify the AMD-specific figures. Expect announcements about specific chip models — likely the MI400-series or whatever succeeds it — using Samsung’s process node within the next six months. Expect public statements about physical AI collaboration to become product launches, possibly at AMD’s next major conference.

The meeting itself, taking place at nearly 8 p.m. after hours with no formal agenda released, was designed to avoid the noise of a staged announcement. That’s the right approach for a deal that’s still being finalized. The signal was clear: AMD and Samsung are serious. The market now waits to see what they actually build together.