Samsung’s HBM ambitions just hit a yield-shaped wall.
Korean outlet IT Chosun claims Samsung is struggling to achieve decent yields with its 1D DRAM, the seventh-generation 10nm-class node it needs for future HBM stacks. The report suggests the outfit may be ready to shelve plans to mass-produce next-generation HBM parts if the numbers do not improve.
The technology has already secured pre-production approval, but Samsung is worried that a trial run would be a money pit, with yields below target. That turns the usual “just ramp it” talk into an uncomfortable spreadsheet exercise.
A Samsung source said: “Samsung Electronics plans to postpone mass production indefinitely until the D1d yield reaches the target level, and as of now, the resumption date has not been determined. They are focusing on further increasing the yield by completely re-examining the process roadmap.”
The D1d node is meant to be a key plank in Samsung’s HBM roadmap, with Chosun linking it to HBM5E, the firm’s ninth-generation HBM product. If that slips, the knock-on is not just a delayed SKU, it is a delayed seat at the table.
Meanwhile, the current 1c DRAM is doing the rounds across HBM4, HBM4E and HBM5, which is fine until customers start demanding the next step. HBM4 is expected later in 2026 and is tipped for Nvidia’s Vera Rubin and AMD’s MI400 platforms, with HBM4E tied to Rubin Ultra and MI500 accelerators.
HBM5 and custom variants are expected to land in Nvidia’s Feynman lineups and rival designs, assuming suppliers can actually ship in volume. Samsung has reportedly been trying to cut its HBM development cycle, but development speed is useless if production keeps tripping over yields.
It looks like production is the real bottleneck. Samsung is not sitting still on facilities, at least on paper. Samsung Electronics is pouring resources into a large chip fabrication plant in Onyang, Korea, described as spanning four soccer fields and aimed at next-generation DRAM, including HBM.
The Onyang site is set up for packaging, testing, logistics and quality control, which are the unglamorous steps that decide whether “next gen” becomes “next quarter”. Big buildings do not magically fix a flaky process, though.
SK Hynix is not waiting around, with the report saying it has already developed and secured yields for its own D1d DRAM. Both firms are chasing major AI customers, and the one with stable yields and believable ROI gets to do the victory lap.
