HBM2e/HBM3 valued within GPU sale — why GPUs price by memory. Updated July 26, 2026 — DDR5 shortage pricing is live; memory is the hottest lot category of 2026.
Indicative per-GB trend, trailing 12 months. Memory reprices with the DRAM spot market — faster than any other category we buy.
| Per-GB market rangeworking pulls, labeled | valued within GPU sale |
| Typical per-stickbrand/condition dependent | n/a |
| Sealed / OEM-boxedpremium grade | +10–20% |
| Untested / unlabeledbulk grade | −35–45% |
| Minimum lotwe buy volume | no minimum |
The 2025–2026 DRAM supply squeeze — driven by HBM fab allocation eating conventional DRAM capacity — pushed server memory prices to their strongest levels in a decade. HBM GPU Modules (on-package) (HBM2e/HBM3 valued within GPU sale — why GPUs price by memory) sits squarely in that demand: data centers refreshing to AI-capable hosts need dense, fast memory, and the secondary market is the only place to get it without OEM lead times. That makes pulled-working sticks from decommissions genuinely liquid assets rather than drawer stock.
Pricing is set per GB and adjusted for speed grade, rank/density, brand (Samsung, SK hynix, Micron price ahead of third-party), and condition. Labels matter: sticks with intact part-number labels grade as working pulls, while unlabeled or bent-pin sticks fall to bulk grade. If you're clearing hosts, pull the memory before the chassis ships — memory sold separately consistently outperforms memory left in low-value servers.
We buy by the stick, the tray, or the pallet, with counts verified on receipt and settlement per category. NIST 800-88 documentation covers the host systems; memory itself is non-data-bearing, which keeps compliance simple. Wire goes out on verification.