HBM Memory Pricing and Specifications (2026)
Last updated: August 2026 · · Methodology
As of August 2026, HBM3 costs approximately $200 per 24GB stack, HBM3E costs ~$300 per 36GB stack, and next-generation HBM4 is expected at ~$550 per 36GB stack (12-hi). SK Hynix leads with 50-55% market share.
HBM Specifications by Generation
| Generation | Capacity/Stack | Bandwidth/Stack | Cost/Stack | Cost/GB | Layers |
|---|---|---|---|---|---|
| HBM2e | 16 GB | 460 GB/s | $120 | $7.50/GB | 8-Hi |
| HBM3 | 24 GB | 819 GB/s | $200 | $8.33/GB | 12-Hi |
| HBM3E | 36 GB | 1.18 TB/s | $300 | $8.33/GB | 12-Hi |
| HBM4 (est.) | 36 GB | >2 TB/s | $550 | $15.28/GB | 12-Hi |
Factory-gate per-stack basis (~$7–14/GB): a component price excluding the capacity guarantees, yield underwriting and engineering support carried under long-term agreements. The per-accelerator table below is on the revenue-implied basis and runs ~2× higher; the two bases are not comparable.
Cost per stack and per GB, by generation
Bars: $/stack (left axis). Line: $/GB (right axis).
Recent Contract Pricing Prints
Newest dated points from our curated HBM contract-price series (full series with history). Contract $/GB — a different basis from the factory-gate specifications table above; the two are not comparable and are never mixed. See methodology.
| Series | Period | Contract Price | Source | Source Date |
|---|---|---|---|---|
| HBM3E (Contract) | 2026 (≈+20% renewals) | $13.00–$18.00/GB (mid $15.60) | Our model estimateSilicon Analysts model (hike: Chosun Biz via TrendForce Dec 2025; base: this series H2 2025) | Jul 2026 |
| HBM3E (Contract — per-stack prints) | 2026 (blended per-stack, post-renewal) | $11.10–$13.90/GB (mid $12.20) | Our model estimateSilicon Analysts model (blend of this series' own verified prints) | Jul 2026 |
| HBM4 (Contract — per-stack basis) | Expected level, Jan 2026 | $16.67/GB | TrendForce (citing Yonhap, Hankyung/KED, Dealsite, EBN) | Jan 2026 |
| HBM3E (Contract — per-stack prints) | 2026 (Samsung renewal, negotiated Dec 2025) | $12.50–$13.89/GB (mid $13.89) | Seoul Economic Daily via SammyGuru (Samsung HBM3E contract renewals) | Dec 26, 2025 |
Rows tagged “our model estimate” are Silicon Analysts derivations, not externally reported prints; every other row names the reporting outlet. Projections are excluded.
HBM3E 12-Hi spot price — daily ledger
Curated HBM spot values are published as ranges (e.g. “$2,600 – $2,800+”); the ledger tracks the low end of the published range. Daily ledger; history began 2026-06-06 (per-domain start may be later). Append-only — cannot be backfilled.
DDR5 16Gb spot price — daily ledger
Consensus median across independent public sources (TrendForce/DRAMeXchange marks); the shaded band is the cross-source min–max spread per run — consensus and disagreement, recorded from day one. Append-only — cannot be backfilled.
HBM Vendor Market Share
| Vendor | Market Share | Key Customers | Status |
|---|---|---|---|
| SK Hynix | 50-55% | NVIDIA, AMD | Leading supplier for H100/H200/B200 |
| Samsung | 35-40% | AMD, Google | Qualified for MI300X, TPU v5p |
| Micron | 5-10% | NVIDIA | Qualified for H200, expanding capacity |
HBM vendor market share
Midpoint of the estimated range; label shows the full range.
HBM per AI Accelerator
| Chip | HBM Type | Stacks | Total Capacity | Total BW | HBM Cost |
|---|---|---|---|---|---|
| H100 SXM5 | HBM3 | 5 | 80 GB | 3.35 TB/s | $1,350 |
| H200 SXM5 | HBM3e | 6 | 141 GB | 4.8 TB/s | $2,400 |
| B200 | HBM3e | 8 | 192 GB | 8 TB/s | $3,250 |
| MI300X | HBM3 | 8 | 192 GB | 5.3 TB/s | $2,900 |
| MI325X | HBM3e | 8 | 256 GB | 6 TB/s | $4,350 |
Per-device HBM procurement cost (revenue-implied basis), from the same canonical chip record behind /api/v1/accelerators. This basis runs ~2× the factory-gate per-stack basis used in the specifications table above, because it carries capacity guarantees, yield underwriting and engineering support. The two are not interchangeable — see methodology.
Key Trends
- HBM cost per GB is rising with each generation due to increased layer counts (8-Hi to 12-Hi), tighter TSV pitch requirements, and lower manufacturing yields on advanced stacking processes. HBM4 is expected to reach ~$15.28/GB (12-hi basis) compared to $8.33/GB for HBM3 — the figures in the specifications table above.
- HBM now represents 30-40% of total AI accelerator manufacturing cost, up from under 20% two generations ago. For the B200 with eight HBM3e stacks, memory alone costs $3,250 — several times the $850 logic die cost.
- Supply remains constrained through 2026. SK Hynix and Samsung are expanding capacity at their Icheon, Cheongju, and Pyeongtaek fabs, but demand from NVIDIA, AMD, and hyperscalers continues to outpace available production. Lead times for HBM3E stacks remain at 20-26 weeks.
Methodology
Pricing estimates are based on a composite of industry supply agreements, analyst teardown reports, and publicly disclosed contract pricing. Cost-per-stack figures represent typical volume pricing for Tier-1 customers and may vary by contract terms, volume commitments, and qualification status. Market share estimates are derived from quarterly revenue disclosures and shipment tracking data. All figures are approximate and intended for reference purposes. Data is reviewed and updated monthly.
Related Tools
HBM, wafer, and capacity movements each week — sourced, dated, human-reviewed.
One short email a week — only when the data moves. No marketing, ever. One-click unsubscribe.