According to media reports, Samsung Electronics has recently allocated approximately 75,000 wafers to HBM4 production, which is half of its monthly HBM DRAM wafer input volume of 150,000 pieces, while the remaining half is dedicated to fifth-generation HBM3E 12-Hi products. It is said that production of the relatively lower-demand HBM3E 8-Hi products has been temporarily suspended, whose capacity has been shifted to HBM4 manufacturing.
HBM3E is primarily used in NVIDIA's current-generation AI accelerator Blackwell, while HBM4 is destined for the next-generation AI accelerator Rubin. Beyond maintaining HBM3E production to meet existing demand, Samsung Electronics has concentrated its entire capacity on HBM4 production.
Samsung Electronics choosing to focus on HBM4 is driven by two main reasons.
On one hand, Samsung aims to revitalize its competitiveness in the HBM market by increasing HBM4 shipments. Earlier reports indicated that within just four months of the first mass production shipments of HBM4 in February, Samsung Electronics generated $1 billion in revenue. Total annual HBM4 revenue for the year is expected to reach $10 billion.
On the other hand, as major cloud providers such as Google, Amazon, and Microsoft ramp up development of their own AI accelerators to reduce reliance on NVIDIA GPUs, demand for HBM from these systems is growing rapidly. It is anticipated that starting with HBM4, the competitiveness of base chips and packaging customized to specific client design requirements will become critically important.