Samsung and AMD have formalized their collaboration through a memorandum of understanding (MoU) that will focus on HBM4 memory and next-generation DRAM solutions for AI infrastructure. Samsung will supply HBM4 memory for the AMD Instinct MI455X accelerator and will develop optimized DDR5 modules for the upcoming sixth-generation EPYC processors, integral to the AMD Helios rack architecture.
The signing ceremony, attended by AMD CEO Lisa Su and Samsung Electronics Executive Vice President Young Hyun Jun, marks a deepening of a partnership rooted in nearly twenty years of shared history in the semiconductor industry. This agreement comes at a pivotal moment for the high-bandwidth memory (HBM) market. Samsung, which holds approximately 22% of the HBM market compared to SK Hynix’s 57%, is looking to solidify its position in this strategic segment of the AI value chain.
HBM4 represents Samsung’s sixth generation of stacked DRAM memory, built on a 10nm class 1c node with a 4nm base logic chip. It achieves transfer speeds of 13 gigabits per second and a maximum bandwidth of 3.3 terabytes per second, surpassing current industry standards. Samsung is the first manufacturer to begin mass production of this generation.
This announcement coincides with Nvidia’s GTC 2026 conference, where CEO Jensen Huang praised Samsung’s HBM4 capabilities and confirmed a foundry partnership with the Korean firm. The timing underscores Samsung’s emerging role as a central supplier for two of Nvidia’s main challengers in AI infrastructure: AMD and possibly other companies seeking to diversify their high-performance memory sources.
Co-Optimizing Memory Architecture and GPU
The AMD Instinct MI455X acts as the centerpiece of this collaboration on AMD’s side. Announced earlier this year and expected to enter production in the second half of 2026, this accelerator competes directly with Nvidia’s Vera Rubin platform. Samsung’s HBM4 will serve as the primary memory provider, a designation that extends beyond a simple supply contract to include a co-optimization of the memory architecture and GPU to maximize performance in training and inference for large AI models.
The AMD Helios rack architecture forms the second layer of this partnership. Designed for large-scale deployments in data centers, it combines Instinct GPUs, EPYC CPUs, and integrated memory solutions. The DDR5 modules that Samsung will develop for the sixth-generation EPYC processors, codenamed Venice, will be tailored for GPU-intensive nodes where processors orchestrate clusters of accelerators. Achieving system optimization between the HBM4 memory of the GPUs and the DDR5 DIMMs of the CPUs is a challenge both companies are aiming to tackle jointly.
Foundry Partnership Discussions Underway
Beyond memory, the agreement specifically mentions open discussions regarding a foundry partnership, in which Samsung could manufacture future AMD products. This prospective collaboration is strategically significant for Samsung Foundry, whose growth has historically been constrained by lower yields compared to TSMC on advanced nodes. Securing AMD as a foundry client would represent a major validation in a market where the credibility of customer references is crucial for winning substantial contracts.
Additionally, AMD recently secured a multi-year agreement for AI chip supply with Meta Platforms, valued at up to $60 billion, along with a similar deal with OpenAI. These foundational contracts naturally increase the demand for HBM memory from AMD, giving Samsung a vested interest in AMD’s commercial trajectory. Young Hyun Jun indicated that Samsung is equipped to offer comprehensive capabilities encompassing HBM4 memory, next-generation memory architectures, advanced foundry services, and packaging, positioning the company as a complete solution partner rather than merely a component supplier.
Memory Bandwidth: The New Bottleneck for AI
For companies and infrastructure architects tasked with scaling training or inference environments, this agreement highlights a profound dynamic within the semiconductor market. Memory bandwidth has emerged as the primary limiting factor for modern AI systems. The most powerful GPUs often remain underutilized when memory cannot supply computational matrices at the required speed. With its 3.3 terabytes per second, HBM4 represents a significant leap over the HBM3 currently deployed in AMD’s MI350X and MI355X accelerators, which are already supplied by Samsung.
The consolidation of the ecosystem around a few major memory-computing alliances—Samsung-AMD on one side and SK Hynix-Nvidia on the other—foreshadows a transformation in the semiconductor supply chain for AI. For data center operators and companies building their GPU infrastructure, this concentration will heighten interdependencies and reduce negotiating power, all while accelerating innovation cycles on the most critical components of the stack.



