Nvidia has pulled SK Hynix into the design process for its next-generation AI chips — not just the parts bin.
The two companies announced a multi-year agreement covering both the design and manufacture of high-bandwidth memory, or HBM, during Nvidia CEO Jensen Huang's visit to South Korea. SK Hynix moves from supplier to co-developer, gaining formal input into the memory architecture that will power Nvidia's Vera-class accelerators. HBM is the specialized memory stacked directly on AI chips — the component that determines how fast a GPU can move data during training or inference. The bottleneck framing is accurate: compute performance has scaled faster than the memory bandwidth that feeds it, and that gap has become an engineering constraint the whole industry is racing to close.
Elevating a supplier to co-development partner is a structural shift, not a procurement decision. Nvidia isn't just securing supply — it's locking in architectural influence before rival customers can shape it. That's a quiet disadvantage for Samsung and Micron, both of which manufacture HBM but now negotiate from a more arm's-length position with the company that sets the pace of AI hardware demand.
Jensen Huang doesn't fly to Seoul to sign a routine supplier contract. The public announcement, timed to a CEO visit, signals that this relationship has moved well past purchase orders.