$Micron Technology(MU)$ $SK hynix(SKHY)$ $Direxion Daily MSCI South Korea Bull 3x Shares(KORU)$ $Direxion Daily Semiconductors Bull 3x Shares(SOXL)$ HBM is an ultra-fast 3D stacked memory that's physically packaged right next to the GPU chip. The key point is that HBM and the GPU can't be bought and installed separately in an AI server; they come as one combined physical unit.

Because of this, the production of $NVIDIA(NVDA)$ 's high-end data center, AI, and supercomputing GPUs is directly tied to the availability of HBM memory. The supply of these GPUs is essentially capped by the supply of the HBM that goes into them.

This applies across the product line:

- **Blackwell & Blackwell Ultra Series:** Flagship AI chips like the B200 (with 180 GB HBM3e) and the B300 / GB200 (with 288 GB HBM3e).

- **Hopper Series:** The industry-standard H100 and H200, which require HBM3.

- **A100 Series:** The prior-generation data center workhorse, which requires HBM2e.

- **Earlier Enterprise Lines:** Older specialized compute GPUs like the Tesla V100 (HBM2) and Tesla P100 (HBM1) also used this memory architecture.

- **Next-Generation Rubin:** Future architectures like the Rubin series are designed specifically around upcoming HBM4 memory.