If DDR4 recycling is a volume business, then HBM (High Bandwidth Memory) recycling is the crown jewel of the memory recycling industry. HBM is the core memory of AI compute chips, with per-unit value far exceeding that of regular DRAM — but it is also the hardest to recycle. In 2026, AI servers deployed in 2022–2023 are hitting their first refresh peak: A100/H100 accelerator cards carrying HBM2e/HBM3 are retiring in bulk, and HBM recycling has moved from theory to real-world demand.
HBM configurations vary enormously across accelerator generations, directly determining residual value:
| Accelerator Model | HBM Generation | HBM Capacity | Stack Height | Whole-Card Recycling Price (Reference Price (CNY)) |
|---|---|---|---|---|
| NVIDIA A100 80GB | HBM2e | 80GB (4 × 20GB) | 4-Hi/8-Hi mixed | 30,000–50,000 |
| NVIDIA H100 SXM5 | HBM3 | 80GB (5 × 16GB) | 8-Hi | 80,000–120,000 |
| NVIDIA H200 | HBM3e | 141GB (6 × 24GB) | 8-Hi | 150,000–220,000 |
| NVIDIA B200 | HBM3e | 192GB (8 × 24GB) | 12-Hi | 250,000–350,000 |
| AMD MI300X | HBM3 | 192GB (8 × 24GB) | 8-Hi | 180,000–280,000 |
As the table shows, in the HBM3e era, per-card residual value has broken through CNY 300,000 — 5–7 times the A100. The core logic behind this: per-stack capacity jumped from 16GB on HBM2e to 24GB on HBM3e, and 12-Hi stacking raised density by 50%.
HBM recycling currently faces a key decision: whole-card recycling or teardown to die level? Each route has its trade-offs:
If teardown to die level succeeds, current market reference prices are:
| HBM Model | Capacity per Stack | Price per Stack (Reference Price (CNY)) | Supplier |
|---|---|---|---|
| HBM2e 4-Hi | 16GB | 800–1,500 | Samsung/SK Hynix |
| HBM3 8-Hi | 16GB | 2,000–3,500 | Samsung/SK Hynix |
| HBM3e 8-Hi | 24GB | 3,500–5,500 | SK Hynix/Samsung |
| HBM3e 12-Hi | 24GB (36GB stack) | 5,000–8,000 | SK Hynix only |
Note: die-level HBM recycling prices are highly volatile. The ranges above are Q3 2026 references; actual quotes depend on die health, test reports, and batch size.
Die-level HBM recycling cannot be done with ordinary teardown operations — the technical bar is far above DDR/NAND:
The HBM market is highly concentrated, split among three manufacturers:
On the supply side, AI server retirement is the largest source of HBM recycling volume. A100 server clusters deployed in 2022 are entering retirement as their 3–4-year warranty periods expire — a single cluster can hold several thousand cards, creating one-off supply windows for HBM recyclers.
| Category | Residual Value per Unit/Stack | Technical Barrier | 2026 Supply Volume |
|---|---|---|---|
| HBM3e | CNY 3,500–8,000/stack | Very high (CoWoS teardown) | Low (just starting to retire) |
| HBM3 | CNY 2,000–3,500/stack | Very high | Medium |
| LPDDR5X | CNY 80–120/unit | Medium (BGA rework) | High |
| DDR4 8Gb | CNY 5–15/unit | Low (standard teardown) | Very high |
| NAND 256Gb TLC | CNY 8–20/unit | Low | High |
HBM's per-unit value is hundreds of times that of DDR4, but volumes are small and barriers high — it suits a premium, high-precision model rather than a volume play.
MemoryRecycle continuously tracks HBM technology evolution and the retirement pace of AI accelerator cards. We have established whole-card recycling capabilities for HBM2e/HBM3 and are building out a CoWoS teardown line. For data centers and cloud providers holding retired AI accelerator cards, we provide free evaluations and professional recycling plans to maximize the residual value of high-value HBM chips.