Samsung Electronics said Thursday morning in Seoul that its 12-hi HBM4 stack has passed Nvidia’s qualification process, and initial shipments will begin in September for use in the Blackwell Ultra 300 series that Nvidia is currently rolling out to hyperscaler customers. This is the news Samsung has been trying to write for the better part of two years, and the fact that it took this long is the actual story.
The HBM3E cycle in 2024 broke against Samsung in a way that turned into a market structure event. SK Hynix passed Nvidia’s HBM3E qualification clean and on schedule. Samsung’s samples ran hot and missed thermal spec on the top of the 12-hi stack. By the time Samsung’s HBM3E finally cleared, the H100 and H200 allocation windows were closed and hyperscaler orders had committed elsewhere. HBM3E turned into an SK Hynix monopoly by accident of manufacturing timing, and SK Hynix’s market cap ran up accordingly through 2025.
HBM4 was supposed to be the reset. It did not reset in Q1 or Q2. Samsung’s first HBM4 samples missed the Blackwell Ultra pre-qual thermal envelope in February. The April round still had signal-integrity problems at the top layer. What Samsung fixed between April and this week is not fully disclosed, and the press release language is “manufacturing process refinements,” but two supply-chain sources describe it as a combination of a redesigned TSV pattern and a new mold-compound supplier from a Japanese vendor Samsung is politely declining to name.
The market read is that HBM is officially back to being a duopoly. SK Hynix has enjoyed effective monopoly pricing on the highest-margin memory in the industry for eight straight quarters. That premium starts compressing in September when Samsung volume flows into the Blackwell Ultra 300 supply pool. Micron is also shipping HBM4 in volume this quarter, but the Nvidia qualification is the one that sets the floor price, and Samsung entering the Nvidia allocation is a bigger event than the Micron ramp because Samsung has the capacity to move share meaningfully.
The follow-on read is that everyone paying for cloud GPU capacity should expect the memory portion of their bill to drift down modestly starting in Q4, and expect the delta between “cloud GPU allocation” and “actually available Blackwell Ultra 300 nodes” to narrow, because the binding constraint through most of 2026 has been HBM supply, not GPU-die yield. The financial read is that SK Hynix will trade sideways for a while, and every analyst who has been calling it a memory-cycle secular winner will have to argue over whether that thesis was really about SK Hynix or really about HBM as a category. The answer is the boring one. It was really about HBM. There is now more than one company selling it at scale to Nvidia.