Citrini Analyst: Rubin Ultra’s reduction in HBM stack layers is not necessarily bearish, and may expand total HBM demand.

Citrini analyst Jukan stated in a recent report that, per his sources, NVIDIA’s upcoming Rubin Ultra may have its HBM specification downgraded from 12 layers of HBM4E to 8 layers. Clients including OpenAI and Anthropic had even requested a 4-layer HBM product, but memory manufacturers rejected the request, with the current downgrade likely stopping at 8 layers. The analyst attributes the adjustment primarily to yield and cost pressures: if Rubin Ultra uses 12-layer HBM4E, combined with price hikes, memory costs could account for roughly 70% of the accelerator’s total bill of materials (BOM). Software optimizations such as model quantization, multi-head latent attention (MLA), and compute task offloading are shifting low-frequency KV caches and model states to LPDDR, CXL, and NAND, leaving HBM to only hold the working set required for real-time computation. As a result, after meeting minimum capacity requirements, customers now prioritize HBM bandwidth over capacity. Jukan added that reducing stack layers would improve packaging yield, lift HBM and AI accelerator shipments, and may even expand total HBM demand. Meanwhile, higher bandwidth requirements reduce the share of chips passing speed binning on wafers, further straining DRAM wafer capacity. Long-term, HBM will eventually be replaced by new architectures, with the ultimate direction being the integration of memory and logic chips. The next two years will be a critical period for memory manufacturers to expand into the logic sector.