AMD CEO Dr. Lisa Su personally traveled to South Korea and dropped two bombshell strategic alliances that shook the industry. The first was with Samsung.Signing of Memorandum of UnderstandingThis ensures the availability of HBM4 memory and advanced foundry capacity needed for next-generation AI chips; it was then announced that...NAVER Cloud, a South Korean cloud computing providerThe two companies will collaborate to build South Korea's "sovereign AI" infrastructure. In these two major announcements, AMD further stated that it will use the Instinct MI455X GPU and the sixth-generation EPYC processor, codenamed "Venice," as the foundation for computing power.
A powerful alliance with Samsung: ensuring HBM4 production capacity and aiming for one-stop OEM manufacturing.
At Samsung's advanced chip manufacturing campus in Pyeongtaek, South Korea, Lisa Su and Samsung Electronics Vice Chairman Jeon Young-hyun witnessed this crucial alliance. At its core, this memorandum of understanding paves the way for AMD's future AI infrastructure blueprint in terms of hardware and supply chain.
• Instinct MI455X GPU paired with HBM4:AMD has announced that its next-generation AI accelerator, the Instinct MI455X GPU, will be among the first to use Samsung's HBM4 memory. Samsung's HBM4 uses a 6th generation 10nm-class DRAM process (1c) and a 4nm logic substrate, achieving processing speeds of up to 13 Gbps and a maximum bandwidth of an astonishing 3.3 TB/s, bringing extreme energy efficiency to AI model training and inference.
• Optimization of the sixth-generation EPYC server processor "Venice" with DDR5:For the 6th generation EPYC server processor codenamed "Venice", the two parties will work together to develop a dedicated, optimized, high-performance DDR5 memory solution.
• Driving the AMD Helios rack-mount architecture:These advanced CPU, GPU, and memory technologies will all be integrated into AMD's next-generation rack-level system architecture, called "Helios," to meet the massive scalability needs of data center customers.
• Potential collaborations in wafer foundry:The statement specifically mentioned that the two parties will explore the opportunity for Samsung to provide "wafer foundry services" for AMD's next-generation products. This means that AMD will not be limited to TSMC in its choice of advanced manufacturing processes in the future, and will have the opportunity to diversify its foundry sources through Samsung.
Partnering with NAVER Cloud: Leveraging the ROCm software ecosystem to aggressively target "sovereign AI"
After securing hardware and supply chain support, AMD simultaneously announced a deep strategic partnership with NAVER Cloud, a subsidiary of NAVER, South Korea's largest portal and cloud service provider. The significance of this partnership lies in the "implementation of a hardware and software ecosystem."
• Early deployment of MI455X and EPYC:NAVER Cloud will significantly expand the deployment of AMD EPYC processors (including the future "Venice"), and AMD will make an exception by providing "early access" to the next-generation Instinct MI455X GPU, allowing NAVER to conduct testing and verification in the cloud production environment ahead of time.
• ROCm optimization that breaks the CUDA moat:The two parties will conduct in-depth optimization of NAVER Cloud's AI services and software stack on AMD's open ROCm software platform.
• Implementing Sovereign AI:By leveraging AMD's open architecture, South Korean companies can gain greater autonomy in data control, privacy protection, and computing performance, creating a national-level AI infrastructure specifically for South Korea.
Analysis of viewpoints
AMD's series of moves in South Korea this time is not only a highly targeted strategic move, but also a key step in countering NVIDIA's hegemony.
In the current AI chip market, NVIDIA enjoys the support of the golden triangle of TSMC (foundry and CoWoS packaging) and SK Hynix (HBM memory). For AMD to break through its production capacity ceiling, partnering with Samsung, which offers a complete "memory manufacturing + advanced packaging + wafer foundry" service, is undoubtedly the smartest choice. This not only ensures that future MI455X chips, and even the next generation, won't be hampered by the inability to purchase HBM4 memory or by being stuck in advanced packaging, but also increases its bargaining power with wafer foundries.
On the other hand, the partnership with NAVER Cloud represents a significant achievement for AMD within the ROCm software ecosystem. As more and more country-specific cloud providers (CSPs) like NAVER are willing to migrate their AI services and optimize them on AMD's architecture, the software moat that NVIDIA's CUDA has long built will begin to crumble. AMD is attempting to demonstrate to the global market that, in the wave of "sovereign AI," adopting the open and cost-effective AMD architecture is the best solution to avoid being held hostage by a single vendor's technology.
