Samsung & NVIDIA deepen foundry partnership; LP40 Groq chip, HBM4E/HBM5 next-gen memory on roadmap

Samsung Electronics and NVIDIA held their most significant strategic discussion yet on June 8, 2026, when Vice Chairman Jun Young-hyun met with CEO Jensen Huang in Seoul to discuss deepening semiconductor collaboration across foundry services and high-bandwidth memory (HBM). The two companies discussed near-term priorities (HBM4 and SOCAMM2 memory supply for Vera Rubin platform this year) and long-term cooperation roadmaps extending into HBM4E, HBM5 supply starting next year, and Samsung Foundry's potential role in manufacturing next-generation Groq LP40 inference chips.

Samsung currently manufactures NVIDIA's Groq 3 LPU (LP30) on its 4nm process and Nvidia's autonomous driving chips on 4nm and 8nm nodes. The new discussion signals Samsung's candidacy to produce the next-generation Groq LP40 at advanced process nodes, potentially including Samsung's 3nm and 2nm processes. Samsung's HBM4E offers 14 Gbps pin speed (scalable to 16 Gbps) and 3.6 TB/s bandwidth, with 16% better energy efficiency and 14% lower thermal resistance than HBM4. The HBM5 architecture was revealed for the first time at GTC 2026.

For architects, this partnership expansion matters because Samsung is positioning itself as an alternative to TSMC for AI inference chip manufacturing while simultaneously increasing its share of NVIDIA's memory supply chain. Currently SK hynix holds 60-70% of HBM4 allocation for Vera Rubin, with Samsung at 25-30%. Samsung's aggressive HBM4E and HBM5 development, plus its proven track record manufacturing Groq inference chips, strengthen its competitive position. Long-term foundry cooperation signals Samsung is attempting to capture advanced-node AI ASIC orders that might otherwise default to TSMC.