High Bandwidth Memory (HBM) stacks DRAM dies vertically with through-silicon vias to break the memory-wall bottleneck, and its newest generation, HBM4, reaches 2 TB/s per stack. But supply sits with only three makers — SK Hynix at 58%, Samsung Electronics and Micron at 21% each — so HBM output and advanced packaging capacity, not GPU logic alone, now set the pace of AI chip shipments.
What is HBM, and why do AI chips need it?
HBM stacks DRAM dies vertically, linked by through-silicon vias (TSVs) next to the GPUCITE:E1. This design — multiple DRAM dies bonded on top of one another and connected through TSVs on a silicon interposer sitting beside the GPU or accelerator — trades a conventional wide-and-flat memory layout for an extremely wide, extremely short data channelCITE:E1. The reason AI accelerators need this architecture comes down to the "memory wall": compute units process data far faster than memory can supply it, so when bandwidth can't keep pace, even the most powerful processor sits idle waiting for dataCITE:E2. HBM's wide, chip-adjacent channel exists specifically to close that gapCITE:E2.
How has HBM bandwidth evolved across generations?
HBM bandwidth rose from 819 GB/s per stack in HBM3 to 2 TB/s in HBM4CITE:E3. The generation in between, HBM3E, reached roughly 1.2 TB/s per stack using a 12-layer, 36GB configurationCITE:E3. HBM4, standardized by JEDEC in 2025, pushes per-stack bandwidth to roughly 2 TB/sCITE:E3.
| Generation | Bandwidth per stack | Configuration |
|---|
| HBM3 | ~819 GB/s | — |
| HBM3E | ~1.2 TB/s | 12-layer, 36GB |
| HBM4 (2025 JEDEC standard) | ~2 TB/s | — |
Each generational jump exists to keep pace with newer AI chips that would otherwise run into the same memory-wall limitCITE:E2CITE:E3.
Why is the HBM market so concentrated, and who controls supply?
SK Hynix holds 58% of the HBM market, with Samsung Electronics and Micron at 21% eachCITE:E4. That breakdown, drawn from Q1 2026 supplier-share data, means only three companies worldwide currently produce HBM at allCITE:E4. This concentration is a direct factor behind AI chip shortages and pricing pressure, since the entire market's HBM output depends on just these three suppliers' capacityCITE:E4.
What does HBM4 NVIDIA certification mean for AI chip shipments?
SK Hynix, Samsung Electronics, and Micron have all secured NVIDIA certification for HBM4CITE:E5. The certification corresponds to NVIDIA's upcoming Vera Rubin platformCITE:E5. With all three suppliers now qualified, the determining factor for AI chip shipment timing shifts to which company can ramp HBM4 volume production first and sustain the highest yieldCITE:E5.
Why is HBM so hard to manufacture?
HBM manufacturing grows harder as stack height and TSV density increaseCITE:E6. Taller stacks and denser TSV connections raise both fabrication complexity and yield riskCITE:E6. HBM also cannot reach the GPU on its own — it requires advanced packaging, such as TSMC's CoWoS, to physically integrate with the processorCITE:E6. As a result, what looks like a "GPU shortage" is frequently, in practice, a shortage of HBM supply and advanced packaging capacityCITE:E6.
What this means
The same three companies that supply nearly all HBM — SK Hynix at 58%, Samsung Electronics and Micron at 21% eachCITE:E4 — have also all cleared NVIDIA's HBM4 certification for Vera RubinCITE:E5, meaning the next bottleneck is not qualification but each supplier's stacking yield and TSV manufacturing capacityCITE:E6. Because HBM must still pass through advanced packaging like CoWoS before it reaches a GPUCITE:E6, bandwidth gains from HBM3's 819 GB/s to HBM4's 2 TB/s per stackCITE:E3 only translate into AI chip shipments as fast as packaging capacity allows.