← Back to feed

Nvidia has certified all three major HBM manufacturers — SK Hynix, Samsung, and Micron — to supply HBM4 for Vera Rubin, its next-generation AI accelerator. The qualification clears a supply chain constraint that analysts had flagged as a potential limiter on Vera Rubin ramp timing. SK Hynix is taking 60–70% of HBM4 allocation, Samsung 25–30%, with Micron holding the remainder.

Alongside the certification news, Nvidia announced a multiyear technology partnership with SK Hynix that goes beyond supply agreements. The two companies will co-develop memory across Vera Rubin AI supercomputers, Vera CPUs, RTX Spark-powered PCs, and NVIDIA Jetson Thor robotics platforms.

What the Partnership Actually Covers

The SK Hynix collaboration has three distinct layers:

Memory co-design. HBM can no longer be designed independently after the GPU architecture is finalised. Vera Rubin’s memory subsystem requires years of joint engineering, fabrication planning, and capital commitment. Both companies are now locked into that roadmap together.

AI tools in chipmaking. Nvidia is embedding its own software stack — CUDA-X, PhysicsNeMo, Omniverse, OpenUSD, cuOpt — into SK Hynix’s semiconductor design and simulation workflows. The goal is to accelerate physics modeling, lithography simulation, and internal engineering code generation.

Fab digital twins. SK Hynix plans to deploy digital replicas of its production lines using Nvidia’s simulation tools, enabling changes to be tested in software before hitting real wafers, robots, and logistics.

Supply Chain Implications

Qualifying all three major memory suppliers rather than relying on a single source significantly de-risks Vera Rubin’s production ramp. SK Hynix has been the dominant HBM supplier through the Hopper and Blackwell cycles. Samsung’s inclusion at 25–30% signals it has cleared Nvidia’s qualification bar after prior supply difficulties. Micron’s allocation, though smaller, marks its full entry into the Vera Rubin supply chain.

The timing matters. Vera Rubin was announced with substantial AI lab commitments, including NVIDIA’s 5-gigawatt DSX deal with IREN and server agreements tied to Anthropic and Meta capacity buildouts. Supply disruption at the memory layer would have cascaded across all of those deployments.

Key Numbers

• SK Hynix HBM4 allocation: 60–70% of Vera Rubin supply
• Samsung allocation: 25–30%
• Micron: remainder
• Platforms covered: Vera Rubin, Vera CPU, RTX Spark, Jetson Thor
• AI tools being co-deployed: CUDA-X, PhysicsNeMo, Omniverse, OpenUSD, cuOpt