Samsung and SK Hynix Target Mass Production of 8-Layer HBM4 for Nvidia by Q2 2027
nashnova research
Samsung and SK Hynix have pushed 8-layer HBM4 mass production back to Q2 2027 to align with Nvidia's Rubin Ultra launch — a delay that signals yield and thermal limits are resetting the pace of the HBM supply chain.
Why did the timeline slip?
Both companies originally targeted H2 2026 for 8-layer HBM4 volume ramp; the new target is Q2 2027.
This means → the earlier schedule was a best-case projection, not a confirmed production commitment.
The direct cause: higher-stack HBM faces thermal challenges and low yields, slowing supply planning across the board.
Why is Nvidia wavering between 8 and 12 layers?
Nvidia initially planned 12-layer HBM4E for Rubin Ultra but is now considering a flexible mix of 8-layer HBM4 and HBM4E, scaled to NVL rack size.
In plain terms = the 12-layer option delivers more performance, but yield and capacity cannot keep up; the 8-layer option is more mature and ships more accelerators on time.
Industry research points to tight DRAM wafer capacity and certification uncertainty around 12-layer parts as the core reasons for Nvidia's hesitation.
What does 8-layer gain — and give up?
An 8-layer stack uses one-third fewer DRAM dies than a 12-layer stack — at the same die capacity, per-package memory shrinks.
This means → the same DRAM output supports more accelerator units, trading per-unit memory for fleet breadth.
Industry research expects 8-layer HBM to become mainstream in H2 2027, while 12-layer HBM4 still dominates through H1 2027.
Could certification become a bottleneck?
Samsung and SK Hynix have reportedly passed Nvidia's qualification for 12-layer HBM4, but the 8-layer product requires a separate certification cycle.
In plain terms = passing the 12-layer exam does not grant an automatic pass on 8-layer — material prep and quality testing start from scratch.
Samsung began commercial HBM4 shipments in February 2026; SK Hynix confirmed mass shipments starting in Q2 this year.
Does per-GB cost actually go up?
Fewer DRAM dies per package means a lower package cost, but the base die's fixed cost is spread over less memory capacity.
This means → per-GB pricing rises — one research firm estimates 8-layer HBM could cost 10%–20% more per GB than 12-layer products in 2027.
This reflects a counterintuitive reality: fitting less memory per accelerator does not necessarily deliver a proportional drop in procurement cost.
Where is the final uncertainty?
Nvidia's Rubin Ultra is expected in H2 2027; the 8-layer HBM4 supply window sits just ahead of that launch.
This means → the final memory spec decision will determine how Samsung and SK Hynix allocate production lines — and how fast downstream AI compute capacity scales.
Until the spec is locked, supply-chain capacity planning remains in limbo.
市场有风险，内容仅供研究参考，不构成投资建议。