Highly anticipated: Samsung Electronics is the world's largest memory manufacturer and one of the few companies with full chipmaking capabilities through its foundry facilities. But until recently, the Korean giant had fallen short of supplying the high-bandwidth memory Nvidia requires for its most powerful AI servers. That has now changed.
Samsung Electronics has resolved the technical hurdles it faced in producing 12-layer HBM3E memory chips, successfully passing Nvidia's strict qualification tests. According to people familiar with the matter cited by KED Global, the Korean chipmaker will soon begin supplying the high-bandwidth memory required for Nvidia's AI servers.
Sources said Samsung completed development of its HBM3E product about 18 months ago but initially failed to meet Nvidia's performance standards. Earlier this year, Jun Young-hyun, head of Samsung's semiconductor division, ordered a redesign of the DRAM core, a change that reportedly resolved the thermal issues seen in earlier versions.
With those problems addressed, Samsung is now cleared to provide HBM3E chips to Nvidia, though volumes will remain limited for the time being. The company becomes the third supplier of such memory to the GPU maker, joining rivals SK hynix and Micron, and is expected to play only a secondary role in the near term.
Samsung's ability to meet Nvidia's strict power and performance requirements is less about revenue and more about prestige, according to insiders. The achievement signals that the Korean manufacturer is back on track with its technology advancements.
HBM3E chips are highly sought-after components, and AMD is already using Samsung-made memory in its MI350 accelerators. Nvidia is expected to deploy the memory in additional B300 servers, designed to significantly accelerate AI inference and training workloads.
The HBM3E technology was first introduced by SK hynix in 2023 and is currently used by Nvidia in data center accelerators based on the Blackwell GPU microarchitecture. Meanwhile, the memory industry is shifting its focus to HBM4, the next-generation high-bandwidth memory standardized by JEDEC in April 2025.
All major memory manufacturers are racing to bring HBM4 products to market, with SK hynix preparing mass production for Nvidia and other major customers. Industry insiders say Samsung's early performance indicators for HBM4 are promising, and the company plans to deliver its first HBM4 samples to Nvidia for validation later this month.