AMD explores Samsung foundry partnership amid shortages

Lisa Su reached the entrance of Samsung Electronics’ Seocho office building in Seocho-gu, Seoul, at about 4:44 p.m. on October 7, seven months after her last visit. Kim Jae-joon, head of strategic marketing at the Samsung memory business division, and Cho Sang-yeon, head of Samsung Electronics DSA Americas, were already at the doors to greet her. Jack Huynh, AMD senior vice president and general manager of the Computing and Graphics Group, and Joseph Macri, an AMD vice president, came in with her for the session with Jun Young-hyun, the vice chairman heading Samsung’s Device Solutions institutional sector.

Asked what the afternoon would cover, she said, “we will talk about our broader partnership.” On whom she would see inside, she answered, “we have many partners at Samsung.” One more question landed: whether the visit would also take her to Samsung Electronics Chair Lee Jae-yong.

“It’s hard to go that far.”

That morning at the Four Seasons Hotel, Su had greeted Ryu Je-myung, second vice minister of Science and ICT. Ahead of the Korea AI Semiconductor & Startup Roundtable, Jack Huynh confirmed he would meet Samsung representatives during the visit. She met SK hynix CEO Kwak Noh-jung on the trip as well. After AMD Gamer Day at Seoul Dragon City in Yongsan-gu, she held a press Q&A and placed the Korean suppliers at the center of AMD’s AI supply. “Samsung Electronics is one of the important strategic partners,” she said. “Samsung Electronics and SK hynix are AMD’s key supply chain partners.” The scale of the need was already fixed in the planning language: “AI demand is very high, so we are planning long-term supply chains with partners.”

“We have started shipping the next-generation MI455 and the Helios system. They are all systems that use HBM4.” High-bandwidth memory stacks dies into dense packages so an accelerator can move large volumes of data at speed. “We are receiving strong support from both Samsung Electronics and SK hynix.”

Demand was surging. Memory was the bottleneck.

In March she had met Jun Young-hyun and the two companies signed a memorandum of understanding to expand collaboration in next-generation AI memory and computing technology. The MOU covered supply of HBM4 and next-generation DDR5. The sides also agreed to discuss designating Samsung as a major supplier of HBM4 for the Instinct MI455X and producing AMD’s next-generation products at Samsung’s foundry. AMD designated Samsung the priority HBM4 supplier for that accelerator. The visit closed with a dinner between Su and Chairman Lee Jae-yong at the Seungjiwon guesthouse in Yongsan-gu, where they discussed cooperation in the AI semiconductor sector.

Samsung aimed to tie scarce high-bandwidth memory to foundry orders and move from memory supply into the overall production of AI semiconductors. AMD needed the volume. The broader partnership was that unfinished March bargain: memory plus contract manufacturing.

No manufacturing agreement has been announced. Seven months later the foundry half still waited for a signature.

AMD still relies on TSMC for the advanced production of its central processing units and graphics processing units. The MI455X is made on that process. A foundry is the contract manufacturer that turns a finished chip design into working silicon at scale. Adding Samsung would mean re-validating performance, yield, and delivery on a line already tuned to one maker. Yield is the share of chips that leave the fab working as designed. Every new process step has to match what the existing supplier already delivers before volume can move.

In its annual report the previous year, AMD warned that shifting some products to a new manufacturer could delay launches or hurt yield and performance. The caution landed harder while AI shipments had to ramp fast. A second source promised more wafers. It also promised the parts would behave the same way and arrive on the same calendar. At a moment when every working accelerator counted, that verification risk was the real price of spreading the work.

The MI455X itself showed how little room the schedule left. The Instinct MI455X holds 432 gigabytes of HBM4. That is 50 percent more than the 288 gigabytes of HBM3E in the previous MI355X. Eighteen units of eight-stack HBM4 supply that capacity on every accelerator. Nvidia’s planned Rubin GPU is set for 288 gigabytes of HBM4.

Even an unchanged sales count would lift total memory demand by half. Each accelerator sold multiplies the stacked memory the market has to produce. On September 29, TrendForce said HBM and general DRAM compete for limited advanced-process capacity. Improving yield and output on new products takes time, so memory supply is expected to remain tight through the next year. Accelerator shipments and memory capacity per product were rising together.

Su has described memory supply as very tight. AMD is working with customers to optimize how memory is used so more compute demand can be met, and it is asking partners to ramp production as quickly as possible. That same tightness keeps the manufacturing discussion on the table alongside the memory orders.

The foundry question—whether Samsung would produce AMD’s next-generation products—remained open under the same roof where the greeters had waited at the door.