Glass is the HBM4 bottleneck: Samsung boosts substrate purchases 2.5x

Glass volume shows how many high-layer-count HBM4 stacks Samsung expects to assemble for NVIDIA accelerators. In 2025, the company bought 10 thousand sheets a month; in 2027, it will buy 50 thousand.

Why glass. The carrier temporarily holds the ultrathin DRAM wafer during processing and prevents it from warping and breaking. Without it, stacks of 12 layers or more cannot be assembled, and HBM4 uses exactly those.

In 2026, HBM4 accounts for about 40% of Samsung's HBM shipments; in 2027, its share will reach 80%. Actual output of HBM4 and HBM4E, by industry estimates, will at least double. Total wafer starts for HBM will increase from 180 thousand a month to 250 thousand, nearly 40%.

The bottleneck is not just glass. Samsung makes HBM4 base dies on 4-nm lines in Pyeongtaek: they produce about 30 thousand wafers a month, and as many as 15 thousand already go to HBM4.

Samsung launched mass production of HBM4 in February 2026 and was first in the industry to ship commercial samples. NVIDIA is taking the volumes: HBM4 is for its Vera Rubin-series GPUs, while Samsung sent HBM4E samples to customers in May.

Micron is expected to reach about 100 thousand HBM wafers per month by the end of 2026 after adding 60 thousand. Samsung and SK hynix are estimated at 150–200 thousand.

Samsung promises custom HBM packages from 2027 and is considering a 2-nm process for its next base dies.