This was confirmed in ZDNet Korea reporting on 26 August 2026. Samsung Electronics and SK Hynix will raise the share of 8-high products in their sixth-generation high bandwidth memory (HBM4) shipments to Nvidia in the second half of this year. Both companies had been supplying 12-high HBM4 for Nvidia’s next-generation AI accelerator family, “Vera Rubin.”
In the memory industry, stack count has long been shorthand for technical capability. So the news sounds like a step backwards at first. It is not.
Why lower the stack count
The reason comes in three layers.
1. Heat
HBM4 doubled the I/O pin count to 2,048 compared with the previous generation. The price of that bandwidth is higher power and higher heat. The crux is that this does not end as a single chip’s problem.
One AI accelerator carries several HBM stacks, one server holds several accelerators, and one rack holds several servers. An increase in heat at the chip level multiplies at the rack level. That is the background to an industry source saying thermal management is what they are watching most closely.
2. Yield
Yield on the DRAM core itself climbs relatively quickly. The problem is what comes after. At the stacking and packaging steps, 12-high loses far more than 8-high. If one die goes wrong partway through stacking twelve, the eleven beneath it die with it.
Low yield does not only mean a higher unit price. It means you cannot deliver the promised volume on time. In a phase like this one, where accelerator shipments are themselves the bottleneck, that hurts more.
3. Keeping options
Nvidia wants to hold several memory options at once. Concentrate on one and the entire schedule stops when that line wobbles. Running 8-high and 12-high in parallel creates freedom in how configurations are combined.
Why this should be read alongside Nvidia’s results
Nvidia reported earnings the same day. CFO Colette Kress disclosed that gross margin would fall to 74% in Q3 and 71–72% in Q4, citing rising component costs including memory.
One sentence she added summarises the nature of this situation precisely.
Today’s memory shortage was in significant part brought about by the build-out of AI infrastructure itself.
Overlay the two pieces of news and it looks like this.
| Viewpoint | What it reads as |
|---|---|
| Nvidia | Memory prices push up cost. Margin gets given up for several quarters |
| The three memory makers | Pricing power has swung to the supply side |
| The supply chain as a whole | Volume is the bottleneck. Whoever yields reliably wins, not whoever stacks highest |
The Korea Economic Daily reported that Nvidia’s purchase commitments for memory and other components reached $279 billion, up 134% in three months. That is the signal that if you do not lock volume up in advance, you cannot buy it.
Where does HBM4E go
There is speculation that 8-high will be the mainstay for HBM4E as well, the next generation. Which memory is finally adopted for the Rubin Ultra family has not been settled.
One thing to remember here. The stack-count race has not ended; another axis has been added. The variables now dividing the field line up like this.
- How much less heat you can produce at the same stack count
- How reliably you can yield through stacking and packaging
- How fast you can switch to the configuration the customer asks for
The third matters especially. When a customer asks you to change the mainstay product, the gap between a company that can turn its lines within a few months and one that has bet everything on a single type opens up sharply in the following generation.
What is left for consumers
HBM is not something an ordinary user can buy. The effects arrive anyway.
- PC DRAM and NAND prices. When production capacity tilts toward HBM, supply of ordinary memory falls. It is one of the reasons behind the memory price rises felt through 2026
- Finished product prices. Memory is not a small share of the bill of materials for laptops, smartphones and graphics cards
- Cloud pricing. GPU instance rates ultimately follow accelerator costs
If you are looking to buy a device with generous memory right now, the strategy of waiting for prices to come down may not work for a while. The cause of the shortage is structural demand, not a business cycle.
In short
- Samsung Electronics and SK Hynix will raise the 8-high share in HBM4 for Nvidia in the second half. It is reported to be at Nvidia’s request
- The reason is not a performance retreat but heat (2,048 I/O pins) and 12-high stacking yield
- The same day Nvidia said margins would fall to 71–72% in Q4 because of memory prices. The cost pressure has been confirmed in the numbers
- HBM4E may also be mainly 8-high. Final adoption for Rubin Ultra is undecided
- For ordinary consumers it comes back as DRAM and NAND prices, and finished product prices
Frequently asked questions
How much performance difference is there between 8-high and 12-high?
Stack count mainly determines capacity. Within the same generation, 12-high holds 1.5 times the capacity of 8-high, while bandwidth itself is set by the interface specification and so does not scale proportionally with stacks. That is why 8-high becomes the more sensible choice on thermals and yield in configurations that need less capacity.
Which is better positioned, Samsung Electronics or SK Hynix?
This change alone does not separate them. The shift to 8-high is a requirement placed on both, and the contest is decided by who can reliably produce the requested volume and by when. With no public yield figures, there is no basis right now for declaring either ahead.
Is it true that HBM is pushing up ordinary RAM prices?
There is an effect, because the same production resources are shared. That said, ordinary DRAM prices are driven by PC and mobile demand, inventory levels and exchange rates together, so HBM alone does not explain everything.
Should I buy RAM or an SSD now?
If you need it now, buying is the better move. The supply structure described here moves on a scale of several quarters, so it is not the sort of thing where waiting a few weeks brings the price down. Conversely, if you have room, the usual principle of buying when you need it is always safest.
Leave a Reply