NVIDIA VeraRubin AI Accelerator Samsung and SK Hynix Get Exclusive HBM4 Supply
NVIDIA has picked its main suppliers for the VeraRubin AI accelerator which is set to launch in late 2026. Samsung Electronics and SK Hynix will supply the 6th generation High Bandwidth Memory HBM4 for this top tier product. Micron didnt make the cut for the VeraRubin HBM4 supply meaning the two South Korean tech giants are in control of high performance memory production.
Samsung and SK Hynix were chosen after tough quality tests to meet NVIDIAs high performance goals. While the Joint Electron Device Engineering Council JEDEC standard for HBM4 is 8 Gbs NVIDIA needed speeds over 10 Gbs for its best hardware.
- Data Transfer Speeds Samsung aced tests for 10 Gbs and 11 Gbs versions. SK Hynix is putting the finishing touches on its 11 Gbs modules.
- Chassis Capacity The VeraRubin has 16 HBM4 chips giving it a total capacity of 576GB. Thats way more than the 432GB in competitors like the AMD MI450.
- Manufacturing The HBM4 stacks use 8 to 16 advanced DRAM layers made with 11-13nm processes placed right on a base die to keep the time between the GPU and memory as low as possible.
Even though Micron used to supply HBM3E for NVIDIAs older stuff it wont be part of the VeraRubin HBM4 launch. Industry watchers think this is because of the tough output and speed needs for VeraRubin. Instead Micron will likely focus its HBM4 production on mid range AI accelerators like the Rubin CPX which dont need the crazy speeds of VeraRubin.
NVIDIA and Samsung are working together even more. Besides getting the biggest slice of HBM4 supply for VeraRubin Samsungs Foundry division is back to making the RTX 3060 GPU. These chips will be made using Samsungs 8nm process giving NVIDIA a solid source for its more budget friendly graphics while it keeps its focus on high end AI.
Even with the HBM4 plans set the DRAM market is still up and down. The price of regular server DRAM has shot up matching HBM3E prices in some areas. This means Samsung and SK Hynix have a decision to make balance the expensive HBM4 stacking with the growing profits from standard server memory. Still Samsung is very hopeful about its HBM division predicting that HBM sales will triple in the 2026 fiscal year.
As GTC 2026 gets closer all eyes are on the production ramp up. Samsung and SK Hynix are expected to start cranking out HBM4 wafers this month to make sure theres enough supply for the VeraRubin systems coming in the second half of the year.
