DRAM spot prices have staged a sharper-than-expected recovery in Q3, driven primarily by AI server build-outs and a surprise inventory drawdown at SK Hynix. The question heading into Q4 is whether demand is broad-based or confined to HBM pull-through — and how long suppliers can hold the line on contract prices.
Samsung's Pyeongtaek P4 ramp remains behind schedule, with yield issues on the 1b-nm node extending into late Q4. This constrains leading-edge DDR5 and LPDDR5X output at the worst possible time. Micron, by contrast, has quietly improved its 1-beta ramp and may exit Q4 with meaningfully better cost structure than its peers.
SK Hynix has effectively bet the farm on HBM3E for the next two quarters. Roughly 60% of its advanced DRAM capacity is allocated to HBM, leaving standard server DRAM undersupplied. If hyperscaler HBM pull slows — say, because of a GPU shipment delay — Hynix faces a painful rebalancing act.
"We see risk of a two-speed market: HBM tight throughout 2025, commodity DRAM potentially softening in H1 if PC and smartphone demand disappoints again."
— Bernstein semiconductor desk, Oct 2024
AI server DRAM content per box continues to climb. Current-generation H100 nodes consume ~8× the DRAM of a typical CPU server, and H200/GB200 configs push that higher still. A single 72-GPU NVL72 rack requires over 17 TB of HBM — numbers that were science fiction three years ago.
| Supplier | Q3 Rev (est.) | DRAM bit growth Q/Q | HBM share of output | Outlook |
|---|---|---|---|---|
| Samsung | $18.2B | +11% | ~25% | Cautious — yield headwinds |
| SK Hynix | $12.7B | +9% | ~58% | Positive — HBM ASP tailwind |
| Micron | $8.1B | +15% | ~20% | Positive — cost curve improving |
Our base case is a modest tightening in commodity DRAM through Q4 with spot prices up 8–12% from current levels. The bear case — another inventory build triggered by weak end-demand — looks less likely than it did 90 days ago, but shouldn't be dismissed entirely.