AMD, which leads the artificial intelligence (AI) infrastructure market alongside Nvidia, said a final agreement with Samsung Electronics for the mass supply of HBM4, the sixth-generation high-bandwidth memory, is imminent. At the event where AMD declared mass production of its "Helios" AI server rack, Samsung Electronics dispatched the head of its foundry (contract chip manufacturing) division and its chief of Device Solutions Americas (DSA) to engage in final-stage discussions to sign the HBM4 supply contract.
Lisa Su, AMD's chief executive officer (CEO), met with reporters during the "AMD Advancing AI 2026" event held in San Francisco on the 23rd (local time). Asked whether AMD was discussing a substantive final contract to receive a significant volume of supply after signing a memorandum of understanding (MOU) with Samsung Electronics for priority HBM4 supply, she answered, "almost close."
CEO Su visited Samsung Electronics' Pyeongtaek campus in March to sign the MOU. It covers receiving priority supply of Samsung Electronics' HBM4 for AMD's new AI accelerators and cooperation with Samsung's foundry. Following the priority supplier selection, she signaled on this day that a final contract for additional HBM4 supply is imminent. In her keynote address, Su said, "Helios is currently in full-scale mass production and is scheduled to ship starting at the end of the third quarter."
Han Jin-man, president of Samsung Electronics' foundry, and Cho Sang-yeon, executive vice president heading DSA, personally attended Su's keynote stage that day to engage in discussions for full-scale HBM4 supply. They stayed for the entire two-hour keynote and, immediately after the speech ended, continued discussions with key executives including Joseph Macri, AMD's vice president and chief technology officer (CTO). Asked about follow-up HBM4 contract discussions, President Han and Vice President Cho avoided answering, saying, "There is an important meeting soon." When asked about follow-up HBM supply and the possibility of producing AMD chips at Samsung's foundry, Vice President Macri said it was a "collaborative relationship."
AMD's 'Helios' carries 50% more memory per chip… Samsung HBM demand set to rise
If Samsung Electronics moves into full-scale mass supply following its selection as AMD's priority supplier of sixth-generation high-bandwidth memory (HBM4), demand for HBM4 is expected to surge. As AI infrastructure demand has been confirmed in the earnings reports of major big-tech firms such as Google and Intel, once AMD begins supplying AI infrastructure to major big-tech companies, the importance of HBM4—which determines server performance—will also grow. Bae Kyung-hoon, deputy prime minister and minister of science and ICT, held a surprise meeting with AMD CEO Lisa Su, throwing his support behind cooperation between AMD and Korean companies including Samsung Electronics.
CEO Su unveiled the AI server rack system "Helios" through her keynote address at "AMD Advancing AI 2026," held in San Francisco on the 23rd (local time).
Helios is a system that integrates 72 "Instinct MI455X" graphics processing units (GPUs) dedicated to AI computation and 18 "EPYC (code name Venice)" central processing units (CPUs) that control the entire system into a steel rack structure. Helios contains not only AMD's GPUs and CPUs but also all networking and software. AMD introduced Helios at "CES 2026," held in Las Vegas in January this year, and disclosed its detailed performance on this day.
The AI rack is core infrastructure for developing and running AI models such as large language models (LLMs). Until now, Lenovo, Dell, and HPE have built and sold AI servers based on AMD CPUs, but Helios packages multiple AMD AI chips together and supplies them by the rack. With the arrival of the era of agentic AI—in which AI reasons and executes on its own—CPUs have become important, and AMD has decided to sell AI infrastructure centered on its own CPUs. In effect, AMD has declared a head-to-head fight against Nvidia's rack system "Vera Rubin NVL72/NVL144," which combines the GPU "Rubin" and the CPU "Vera."
Helios to ship at end of Q3… secures customers including OpenAI and Anthropic
AMD has already secured major hyperscalers (operators of large-scale computing infrastructure) and ultra-large AI developers—including OpenAI, Anthropic, Microsoft (MS), Meta, and Oracle—as Helios customers. Helios will begin shipping at the end of the third quarter and enter full-scale supply from the fourth quarter. Sachin Katti, head of infrastructure at OpenAI, who took the keynote stage that day, said, "We plan to begin deploying Helios at scale from the end of this year and expand that throughout next year." Anthropic has formed a partnership to adopt Helios systems of up to 2 gigawatts (GW) in scale.
Especially notable is that Samsung Electronics' HBM4 will be mounted in Helios, which enters mass production at the end of the third quarter. Samsung Electronics shipped HBM4 first in the industry in February this year and, a month later, signed a priority supply agreement with AMD on the occasion of CEO Su's visit to Korea. Bart Walker, senior vice president of the memory business division at Samsung Electronics' DS division, held a briefing at the exhibition hall that day, explaining that Samsung has maintained a 20-year cooperative relationship since supplying GDDR4 for AMD's Radeon HD 2000 series in 2007, and that this led to the March priority HBM4 supply. Following its supply of fifth-generation HBM3E to AMD last year, Samsung has secured major partner status in HBM4 as well.
If Samsung begins full-scale HBM4 supply to AMD, Samsung Electronics and SK Hynix—the two leaders in HBM—could further solidify their standing as memory companies in the AI rack competition as well. Both companies' HBM4 has been supplied for Nvidia's AI rack "Vera Rubin," and if AMD joins in, HBM shortages will deepen, which could lead to long-term contracts. From the perspective of domestic memory companies, this can also be expected to reduce dependence on Nvidia. While one Nvidia GPU "Rubin" is connected to 288 gigabytes (GB) of HBM4 memory, AMD's GPU "MI455X" carries 50% more at 432GB, increasing the supply volume as well.
Deputy Prime Minister for Science Bae, the top official overseeing sovereign AI (AI sovereignty) policy including independent foundation models, also made a surprise appearance at the AMD event that day. CEO Su personally introduced AMD's new products, including Helios, and Deputy Prime Minister Bae toured physical AI (technology combining AI with the physical world) using AMD technology, cloud servers, and more. The originally scheduled time was 30 minutes, but he left the venue after touring booths of domestic companies such as MangoBoost and Moreh for about an hour. He said, "It is meaningful for Korean companies to participate in this open ecosystem as AMD expands its share in an Nvidia-centered ecosystem," adding, "Since Korea has strengths in semiconductors, it should work to cooperate with AMD while creating Korea's own independent ecosystem and drawing in participation from global companies."
