Why AMD and Samsung's Reset Changes the HBM Supply Chain
AMD and Samsung are restarting talks on HBM and foundry cooperation, a move that could reshape the AI chip supply chain and signal deeper hedging against Taiwan risk.
A Second Meeting That Isn’t Routine
Jack Huynh didn’t need to confirm much when asked if AMD would meet Samsung during his visit. He said the company already had plans. The details matter more than the handshake.
Seven months ago, in March, Lisa Su dined with Samsung chairman Lee Jae-yong at the presidential guesthouse Seungjijwon in Yongsan. That visit produced an MOU covering next-generation AI memory and computing. AMD named Samsung the priority supplier for HBM4 to be used in its Instinct MI455X accelerator. The numbers were striking: 432 gigabytes of HBM4 per chip, eight stacks, 18 chips total per accelerator—50 percent more memory than Nvidia’s upcoming Rubin GPU will carry.
This visit is different. It is about what comes after memory.
Su has publicly stated that AMD will significantly increase its AI accelerator output and that the company is now planning supply chains out three to five years instead of one to two. That extension alone changes the conversation. It moves the relationship from a procurement arrangement into something closer to a structural partnership.
Huynh declined to say what topics the Samsung meeting would cover. But Su was more direct when asked about foundry: she said she would have time to discuss it this afternoon. The afternoon agenda includes an AMD Gaming Day event in Dragon City. Foundry may not be on that public bill, but the subtext is clear.
The Capacity Constraint No One Ignores
TSMC is at full cap on its 3-nanometer, 5-nanometer and 4-nanometer nodes. That is not speculation. It is the current state of the business Su acknowledged when she told reporters, ahead of this trip, that she had met with TSMC and Foxconn in Taiwan and that AMD needs more supply. The phrase she used was blunt: more volume is required.
Samsung foundry is the only realistic alternative at this point. Not the same process nodes. Not yet the same yield curve. But a second source for leading-edge logic production where none existed before, and a diversification layer that reduces the tail risk of a single-fab dependency.
The MI455X will use TSMC’s most advanced process—that is the current plan. But the next generation, and the volume ramp that Su described, may not fit on one fab line. Samsung is the place where that overflow goes.
HBM4 Is a Bottleneck With a Price
HBM4 is not yet a mass-produced commodity. Samsung holds the lead on volume and quality, a fact codified when AMD selected it as the priority supplier for MI455X. SK Hynix, Nvidia’s preferred HBM4 source, is running a tighter ship but cannot fill the entire market. Micron is still catching up.
The capacity constraint Su flagged applies to memory as well as logic. The supply side of HBM is constrained across the board. AMD’s decision to lock in Samsung for HBM4 priority supply is both a hedge and a dependency. It buys allocation security now. It also cedes bargaining power later, because Samsung controls the throughput.
That trade-off is worth naming. AMD gains volume. Samsung gains a strategic anchor client for its memory business, something it has chased since losing ground to SK Hynix in the high-performance memory race. The MI455X deal gives Samsung a revenue floor and a technical reference design that reinforces its positioning.
What This Signals About Taiwan Risk
The geopolitical dimension is understated in most coverage of this reset, but it is the reason the reset exists at all.
Taiwan concentrates the majority of the world’s leading-edge logic fabrication. A conflict, blockade or escalation would sever that line. AMD and every other fabless company that relies on TSMC knows this. The question is not whether to hedge. The question is how aggressively.
Samsung’s foundry push in South Korea is the most serious alternative currently available. The government has backed it with subsidies and tax incentives. The technology gap remains real—Samsung’s yield rates and process maturity lag TSMC by roughly one to two nodes. But the gap matters less when the alternative is total supply disruption.
AMD’s extension of planning horizons to three to five years is a signal that the company is treating this as a permanent structural shift, not a temporary workaround.
Who Wins, Who Loses, What Happens Next
Samsung wins first. Priority HBM4 supply agreements are revenue. Foundry design wins are a longer game, but they start here. Every AMD chip that flows through Samsung’s fabs in the next round adds to the foundry pipeline and strengthens the company’s case to other AI accelerator designers.
AMD wins second, but the win is conditional. More supply means more revenue. But Samsung will price that access. The dependency cut both ways: AMD needs Samsung’s HBM4 volume; Samsung needs AMD’s brand and design volume. Neither side can walk away cleanly without pain.
Nvidia loses margin leverage. Its HBM4 dominance is under pressure. If AMD secures enough Samsung allocation, the gap between Nvidia’s accelerator memory supply and its competitors narrows. TSMC loses margin upside on AMD logic, at least partially, because some volume shifts south.
The next moves are predictable: a formal foundry agreement if talks this week go beyond conversation, a second round of HBM4 design wins for Samsung, and a public acknowledgment of diversified supply chains from AMD that has been deliberately silent on the details so far.
Su will not explain the specifics today. She rarely does. But the arc is visible. The reset is not about nostalgia. It is about capacity, risk and the reality that no single fab can serve the AI accelerator market alone.