NVIDIA and SK hynix Expand HBM Pact for AI Factories
A multiyear co‑development deal aims to ease high‑bandwidth memory bottlenecks for large AI systems
NVIDIA and SK hynix on June 7–8, 2026 announced a multiyear technology partnership to co‑develop next‑generation memory aimed at supporting global AI factory buildouts and accelerating semiconductor design and manufacturing.
The agreement formalizes joint engineering work on advanced memory tailored to NVIDIA’s infrastructure roadmap, and says SK hynix will align new DRAM and HBM lines to Nvidia’s product set across data center, personal AI and robotics markets.
The move targets one of AI’s persistent bottlenecks: demand for high‑bandwidth memory (HBM) in modern accelerators, which can outstrip the pace of memory development and factory capacity expansions.
Under the pact, SK hynix will codevelop memory for several upcoming NVIDIA platforms named in the announcement, including the Vera Rubin AI supercomputer family, Vera CPUs, RTX Spark PCs and Jetson Thor robotic modules.
The partnership comes as SK hynix and peers invest heavily to add capacity for HBM and advanced DRAM, including large orders of extreme‑ultraviolet lithography tools to support next‑generation fabs and packaging lines. Those investments are meant to close a gap between rapid AI demand and long semiconductor lead times.
SK hynix also told customers and markets this month that it has begun shipping samples of next‑generation HBM chips, a sign its development pipeline is moving toward production even as qualification and full ramp remain years‑long processes.
Supply dynamics are shifting from single‑vendor dependencies to multi‑supplier ecosystems; NVIDIA has said it certified several memory makers to supply HBM4 for upcoming accelerators, while continuing deep co‑engineering with SK hynix under the new pact.
Industry observers caution that multiyear codevelopment agreements help smooth future supply but do not instantly expand wafer or packaging capacity, because advanced HBM stacks require complex tools, test flows and time‑consuming qualification.
Beyond chips themselves, the deal emphasizes software‑driven fabs: SK hynix will use NVIDIA CUDA‑X libraries, PhysicsNeMo and Omniverse tools to accelerate semiconductor simulation, design automation and digital‑twin efforts for fab automation. The companies say these steps can shorten development cycles and improve yield.
For cloud providers and hyperscalers racing to build AI factories, a coordinated supplier roadmap matters because HBM shortages can bottleneck whole systems, from large model training clusters to inference and edge AI deployments. The partnership signals a shared roadmap between a dominant accelerator designer and a leading memory maker.
The announcement also includes the usual forward‑looking caveats: NVIDIA and SK hynix note that expectations about performance, availability and market trends are subject to risks from development cycles, capital investments and changing demand patterns. That underscores how memory supply depends on multi‑year capital plans rather than single contracts.