The tiny memory glued to every AI chip.
AI chips need very fast memory stacked right next to them. Three companies sell almost all of it. Here is why, in plain English, with every source.
1. What HBM is
HBM (high-bandwidth memory) is thin memory chips stacked on top of each other and placed right next to an AI processor, so data moves very fast. In Q2 2026, SK hynix had about 50% of HBM sales, Samsung 33% and Micron 18% (rounded, so it adds to 101%). These are shares of money, not of chips. [1]
2. Who is cleared
In June 2026, Nvidia’s CEO said all three makers had passed its tests for HBM4, the memory on its next AI chip. Passing these tests and packaging the stacks well is hard, which keeps newcomers out. [2]
3. China is trying
China’s CXMT announced new ordinary memory chips in 2026. Reports say it is testing HBM in small amounts, but the company has not confirmed it. [3] [4]
4. Not every AI device needs HBM
Nvidia’s Jetson Thor, a computer for robots, uses 128 GB of LPDDR5X, the phone-style kind. The average new phone went from 7.4 GB of memory in December 2024 to 8.4 GB in December 2025. [5] [6]
5. The test
Shares are rounded and research firms revise them. Three makers today does not mean three forever.
Sources
- Counterpoint Research, Global DRAM and HBM market share, Sep 2026. counterpointresearch.com
- Bloomberg Línea, Nvidia approves three memory makers for HBM4, Jun 2026. bloomberglinea.com
- CXMT news release, 2026. cxmt.com
- Reuters via StreetInsider, citing The Information, Aug 2026. streetinsider.com
- Nvidia, Jetson Thor product page. nvidia.com
- Counterpoint Research, smartphone average DRAM 8.4 GB in 2025. counterpointresearch.com
Image credits
- Cover and carousel slide 1: illustration made for YosQuant.
- “Semiconductor clean room”, Aileen Devlin / Jefferson Lab from Newport News, USA, Public domain. Wikimedia Commons
- “Silicon wafer researcher”, U.S. Department of Energy, Public domain. Wikimedia Commons
