Global Memory: Memory becoming a burden of AI too?

Research evidence excerpt

Global Memory: Memory becoming a burden of AI too?

22 June 2026

Price Target Change

Global Memory

We update our memory & industry models (models for HBM, DRAM, NAND, Samsung, SK Mark Li

+852 2123 2645 hynix, Micron & KIOXIA) & maintain Outperform ratings on Samsung, SK hynix & Micron &

mark.li@bernsteinsg.com Underperform on KIOXIA.

Edward Hou, CFA HBM price needs to go higher to narrow the profitability gap vs. conventional DRAM.

+852 2123 2623 From 3QCY25 to 2QCY26, conventional DRAM price has risen c. 4.5x, but HBM price is edward.hou@bernsteinsg.com

locked by annual contracts and hasn’t moved. As the result, we estimate in CY26 deploying

Yipin Cai, CFA capacity to conventional DRAM will generate over 2x revenue & nearly 3x gross profit dollar

+852 2123 2669 per wafer capacity vs. HBM. This is why memory suppliers and GPU/XPU companies are

yipin.cai@bernsteinsg.com negotiating CY27 HBM price now to narrow the gap.

The possible “markup” of GPU/XPU suppliers will amplify the cost burden on

hyperscalers. Unlike conventional DRAM & NAND that hyperscalers can source directly

from memory suppliers, HBM is packaged in GPUs/XPUs and is part of the COGS of, for

example, NVIDIA. NVIDIA needs to mark up the HBM price hike 4x if NVIDIA has 75% gross

margin & wants to keep 75% unchanged, despite the HBM price hike. This will make the

HBM price hike even a heavier burden for hyperscalers.

With memory making hyperscalers’ capex 30% higher, a “re-calibration” is

inevitable and may squeeze weaker suppliers. Using Vera Rubin (VR, NVL72) rack as

an example, hyperscalers will find their capex higher by 30% if they deploy a data center

with this rack and with the same memory capacity installed in the rack, because of higher