Samsung Electronics said Friday that its 12-high HBM4 stack passed Nvidia’s qualification audit for Vera Rubin Ultra, the part of the Vera Rubin generation that ships into the largest training racks. Samsung will supply Q4 production volume alongside SK hynix and Micron, the two suppliers that have been sole-sourcing HBM4 for Vera Rubin since the qualification cycles started in late 2024. Samsung shares closed up 6.8 percent on the KOSPI. SK hynix shares closed down 4.2 percent. Micron held flat in pre-market US trading.
This has been a slow-motion narrative inside the Korean semiconductor press for about a year and a half. Samsung’s HBM3E qualification for the H200 took multiple revisions. The HBM4 cycles for Vera Rubin Ultra reportedly hit yield and thermal-cycling issues on three separate attempts. Each miss handed SK hynix more of the share of the most expensive memory component on the most expensive accelerator on the planet, and SK hynix happily accepted. The current consensus estimate is that SK hynix supplied roughly 58 percent of Vera Rubin Ultra HBM4 in the first half of 2026, with Micron at 32 percent and the rest covered by tactical SK hynix overflow. Samsung’s number was effectively zero.
What changes now is the supply-side bargaining. With three qualified suppliers, Nvidia can run procurement against all three on price for the back half of the year, and the memory pricing power that SK hynix has been enjoying since 2024 starts compressing. The other read, which the desks in Seoul are already trading on, is that Samsung’s foundry roadmap may have caught up in ways that have not yet been priced in. Vera Rubin Ultra is the bleeding-edge qualification gate in AI memory. Clearing it is meaningful evidence that the broader Samsung memory operation is no longer behind. The fact that this took eighteen months is also evidence that it really, really was behind. Both things can be true.