Nvidia Ramps Up Audits of Samsung's HBM4 Packaging to Power Next-Gen Rubin GPUs

Nvidia has initiated intensive audits of Samsung Electronics' HBM4 high-bandwidth memory packaging, focusing on mass production yields and delivery performance for its next-generation Rubin GPUs, also known as Vera Rubin. Samsung, a key player in the memory market, is in the final adjustment phase and plans to kick off HBM4 mass production as early as February 2026, following Nvidia's approval. This comes after Samsung qualified its HBM3E stacks in late 2025 and delivered HBM4 samples, leveraging its unique turnkey capabilities that integrate DRAM production, logic dies on 4nm processes, and advanced 3D hybrid bonding packaging.

The audits are critical as Nvidia seeks to secure a stable supply chain for Rubin GPUs, slated for release in the second half of 2026, which promise up to 900 times the performance of Hopper at a fraction of the cost. By qualifying Samsung alongside leader SK Hynix—which is expected to supply around 70% of HBM4—Nvidia reduces reliance on TSMC and mitigates risks from surging AI demand. Samsung's advancements, including energy-efficient 1c DRAM and 16-layer stacks enabling up to 512GB per GPU in 8-stack configs, position it to capture about 30% market share, narrowing the gap with rivals while Micron faces setbacks meeting stricter specs like over 11 Gbps per pin.

This development intensifies competition among Samsung, SK Hynix, and Micron in the HBM4 race, driven by Nvidia's leverage in setting exacting standards. Looking ahead, successful audits could accelerate Rubin's production ramp-up, enabling massive AI models on compact hardware and fostering 'Sovereign AI' data centers. Industry watchers anticipate 16-Hi HBM4 volume shipments through 2026-2027, paving the way for HBM5 and further semiconductor supply chain realignments.