AMD Secures Samsung HBM4 for Instinct MI455X, but the Real Surprise Is the Foundry Angle
AMD has officially expanded its next generation memory partnership with Samsung, securing HBM4 supply for the upcoming Instinct MI455X while also aligning on DDR5 solutions for 6th Gen EPYC Venice platforms. The agreement was announced directly by Samsung after Lisa Su’s visit to Korea, and it confirms that Samsung will serve as the primary HBM4 supplier for AMD’s next wave of AI accelerators. That is the headline, but the more interesting part is the caveat attached to the deal: this is not just a memory agreement. Samsung says the 2 companies will also discuss opportunities for a foundry partnership for future AMD products.
Under the memorandum of understanding, Samsung and AMD will align on primary HBM4 supply for the AMD Instinct MI455X GPU, as well as advanced DRAM for EPYC Venice, with both feeding into AMD’s broader Helios rack scale AI platform. Samsung also used the announcement to stress that it is “uniquely positioned” to offer a turnkey combination of advanced memory, foundry, and packaging capabilities. That language matters because it shows Samsung is trying to sell AMD a broader platform level relationship, not just high bandwidth memory wafers.
From AMD’s side, this is a strategically important win, but it is also a reminder that Samsung’s leverage in the HBM4 race has grown fast. Samsung says its HBM4 is built on its 6th generation 10 nanometer class DRAM process with a 4 nanometer logic base die, delivering up to 13 Gbps processing speeds and up to 3.3 TB/s of bandwidth. That makes it one of the most attractive HBM4 options now entering the AI infrastructure market, especially as NVIDIA has already moved early to deepen its own Samsung relationship for the Rubin era.
The real twist is the foundry piece. Samsung’s official wording is careful, saying the 2 companies “will also discuss opportunities for foundry partnership,” which stops well short of confirming that AMD has already committed advanced AI chips or CPUs to Samsung fabrication. Reuters matched that interpretation, describing the deal as an expansion of AI memory cooperation that also includes discussions about Samsung providing foundry services for next generation AMD products. That means the foundry side is real, but still exploratory in official terms.
This is where the Korean reporting becomes more interesting. A Chosun Biz report has suggested the understanding around HBM4 may come with stronger expectations around Samsung Foundry participating in AMD’s future advanced chip manufacturing. I have not found a public official statement from AMD or Samsung confirming that specific condition in those exact terms, so it is best treated as reported local industry detail rather than a fully confirmed contractual requirement. Still, even Samsung’s own announcement makes clear that foundry cooperation was part of the conversation, which is enough to make this more than a straightforward supply deal.
The bigger business picture is easy to read. Samsung wants to turn its recent HBM momentum into broader semiconductor wins, and AMD wants reliable access to next generation memory as AI demand keeps tightening supply chains. If Samsung can convert memory wins into foundry and packaging business, it strengthens its position against TSMC at exactly the right time. For AMD, getting primary HBM4 capacity is the immediate prize, but the price of that access may be a deeper relationship with Samsung’s broader semiconductor stack.
In that sense, AMD may have achieved what it needed, but Samsung may have structured the deal so that it wins on multiple fronts. HBM4 supply for MI455X is the visible result. The foundry conversation is the part that could matter even more over the longer term.
Do you think AMD should deepen its manufacturing ties with Samsung beyond memory, or keep Samsung focused mainly on HBM supply?
