What Is HBM?

High-bandwidth memory, in plain English — why AI can't run without it, how the generations differ, and who actually makes it.

The one-sentence version

HBM (high-bandwidth memory) is DRAM with the memory dies stacked vertically and wired to the processor through a very wide channel, so it moves data far faster than ordinary DDR memory. That speed is exactly what AI accelerators need.

Why AI needs it

Large AI models are bandwidth-bound: the expensive compute cores in an NVIDIA GPU sit idle if memory can't feed them data fast enough. HBM solves that. It is why essentially every leading AI accelerator ships with HBM instead of standard DRAM — and why HBM demand has tracked the AI buildout.

• HBM demand grew an estimated +130% year-over-year in 2025, with a further +70% projected for 2026.
• One HBM bit requires roughly 3× more wafer capacity than a DDR5 bit, so HBM absorbs a disproportionate share of fab output — about 19% of total DRAM wafer output in 2025, rising toward ~23% in 2026.

Source: TrendForce free press releases, Nov 13 2025. Demand figures are vendor/tracker estimates and projections — directional.

Who makes it

Only three companies make HBM at scale:

Maker	Position
SK hynix	The HBM leader across recent generations; reported to supply more than half of NVIDIA's 2026 HBM.
Samsung	Pushing hard on HBM4; began HBM4 mass production in Feb 2026.
Micron	Passed Samsung for #2 in a 2Q25 snapshot; HBM4 ramp targeted for 2Q26.

HBM share differs sharply by methodology (bit-output vs revenue) and by snapshot vs full-year; treat any single share number as directional, not gospel. See the live HBM section on DRAMWatch for the dated, sourced data.

HBM3E vs HBM4

• HBM3E — the generation shipping in volume into 2026. Samsung has shipped HBM3E to NVIDIA since Q3 2025.
• HBM4 — the next generation, with a wider interface and higher bandwidth. Samsung began HBM4 mass production in Feb 2026; Micron and others ramp through 2026; SK hynix targets HBM4E mass production in 2027. HBM4 is the memory behind NVIDIA's Vera Rubin platform.

Why it matters for the market

Because HBM eats so much wafer capacity and sells at a premium, it has pulled the entire memory market tight — conventional DRAM contract prices have spiked alongside the AI buildout. If you want to track that, the live DRAMWatch dashboard keeps the dated, sourced numbers on HBM, DRAM pricing, vendor share, and the HBM4 roadmap.