This year, Samsung started mass production of HBM4 chips for NVIDIA's flagship Vera Rubin AI platform. This deal will generate billions of dollars in revenue. Now, the South Korean firm appears to have an advantage over its rivals in supplying custom HBM chips for NVIDIA's next flagship AI accelerator architecture.NVIDIA recently announced NVHBM, a custom high-bandwidth memory (HBM) chip designed for its next-generation AI platforms, Rubin Ultra, which is scheduled to debut next year. Samsung is in a uniquely strong position to supply NVHBM to Nvidia because of its turnkey capability to design, fabricate, and package both DRAM base dies and logic chips completely in-house.In comparison, rivals firms Micron and SK Hynix depend on third-party foundries like TSMC to fabricate advanced logic base dies and handle packaging. Since custom HBM requires tight integration of custom logic chips and memory base dies, Samsung's all-in-one manufacturing ecosystem offers a distinct technological advantage.Modern HBM designs rely on 12-layer or 16-layer stacks, which introduces design complexity and yield issues. To counter these issues, NVIDIA has requested for an 8-layer HBM4E design for NVHBM. This approach reduces stack height by roughly 33% and substantially improves production yields, which in turn helps ease memory shortages.While an 8-layer configuration reduces raw memory capacity per stack, it also reduces stack capacity. NVIDIA aims to offset this disadvantage by driving data transfer speeds up to 18Gbps, which is around 20% faster than initial 14.4Gbps HBM4E samples.When paired with individual GPUs and linked across hundreds of compute units via NVIDIA's high-speed NVLink interconnect, this faster custom memory configuration delivers massive aggregated throughput across the entire cluster. If Samsung succeeds in getting a deal from Nvidia for NVHBM supply, production could start sometime next year.