Mobile DRAM Jumped 89% in One Quarter — Blame AI Servers
AI servers are not a single-memory story. Nvidia's Vera Rubin platform pairs HBM4 with a second, cheaper memory type — and in the second quarter of 2026, that second memory moved faster than HBM did. A 12GB LPDDR5X module that cost $77.10 a quarter earlier jumped to $145.90: an 89% increase in three months.
KEY TAKEAWAYS
1. Nvidia's Vera Rubin platform pairs HBM4 with SOCAMM2, a new server module built on LPDDR5X. SK hynix started mass-producing a 192GB version on April 19, 2026.
2. 12GB LPDDR5X pricing rose 89% quarter-over-quarter in Q2 2026 — from $77.10 to $145.90 — per SIGMAINTELL data reported by Seoul Economic Daily.
3. TrendForce estimates the three memory makers' combined 2027 LPDRAM output will cover only about 60% of Nvidia's requirement.
What SOCAMM2 actually is
Server memory has run on RDIMM for two decades — the long, upright modules that sit next to the CPU. LPDDR, by contrast, was built for phones: low power, soldered close to the processor, never meant to be swapped out.
SOCAMM2 turns that phone-grade LPDDR5X into a server-pluggable module. It carries the JEDEC standard number JESD328, still in its finalization stage as of August 2026, and supports up to 9.6Gb/s per pin.
SK hynix is furthest along: its 192GB SOCAMM2, built on 10nm-class 1c LPDDR5X, went into mass production on April 19, 2026 for Nvidia's Vera Rubin platform. The company claims more than 2x the bandwidth and 75% better power efficiency versus standard RDIMM. Samsung says its version offers more than 2x bandwidth and 55% lower power draw, and is currently sampling to customers. Micron is reportedly sampling a 192GB module as well. All three figures are company-disclosed and use different baselines, so they should not be read as a direct one-to-one comparison.
Company-reported figures: SK hynix 192GB module vs. standard RDIMM. Not independently benchmarked.
Why servers started using mobile-grade memory
An AI rack has a fixed power budget. Every watt a memory chip doesn't use is a watt the GPU can use instead. LPDDR was designed from the start to run on a battery for days, so it wins on performance-per-watt almost by construction.
Workload matters too. Much of inference work is more sensitive to throughput and efficiency than to raw latency. Giving up some of RDIMM's latency advantage in exchange for bandwidth and power headroom is a trade that pencils out for that workload.
HBM can't simply take this role instead — it's expensive, and package capacity is capped. The split that has emerged is straightforward: HBM sits next to the GPU, LPDDR sits next to the CPU.
The price move came first
SIGMAINTELL's Q2 2026 memory pricing report puts 12GB LPDDR5X at $145.90, up 89% from $77.10 the prior quarter. 4GB LPDDR4X rose 75%, and 512GB SSDs rose 54% over the same period.
The driver is allocation, not scarcity of silicon in the abstract. As the three memory makers route limited wafer capacity toward HBM and premium server DRAM first, smartphone and PC makers are absorbing price increases just to secure volume. This looks like the first case of AI demand spilling directly into mobile memory pricing — and it flows straight into smartphone bills of materials, where memory is not a trivial line item.
Samsung + SK hynix + Micron combined, indexed to Nvidia's requirement = 100. Source: TrendForce (Jun 2026).
Supply is not catching up
TrendForce's June 2026 outlook puts the three makers' combined 2027 production at roughly 60% of Nvidia's LPDRAM requirement. Nvidia's own response is telling: Vera Rubin's SOCAMM module capacity per unit was cut in half, with more modules used to compensate. TrendForce reads this as a reallocation of supply, not a cut in total demand.
Nvidia's response is one data point on how tight this looks from the demand side. TrendForce also expects AI servers to overtake smartphones as the largest LPDRAM demand source sometime between 2028 and 2030.
Stage of SOCAMM2 rollout, as reported by each company (as of Aug 2026).
What I actually watch
| Checkpoint | Why it matters |
|---|---|
| How LPDDR capacity gets added | Repurposed mobile lines vs. new fab capacity changes the mobile supply impact entirely. |
| JESD328 finalization | A finalized standard speeds up entry for later movers. |
| Samsung and Micron mass production dates | Only SK hynix has announced volume production so far; a second entrant is the pricing inflection point. |
| Cost pass-through by device makers | Whether phone and PC makers pass the increase on to consumers signals if this cycle curbs its own demand. |
Value chain read-through
| Segment | What's sold | Indicator to track |
|---|---|---|
| Memory | LPDDR5X, server-grade low-power DRAM | Mass-production timing and capacity allocation across the three makers |
| Module | SOCAMM2 modules | Capacity per module, platform adoption |
| Substrate | Module-level PCB | Customer qualification pass rates |
| Mobile devices | Smartphones, PCs | Cost burden and pricing pass-through |
| Legacy servers | DDR5 RDIMM | Replacement speed, price resilience |
Risks to this view
• A price spike cuts both ways: it's earnings for memory makers, but cost for phone and PC makers. If end demand breaks, volume reverts in the next cycle.
• The 89% figure comes from a single research firm. It's worth cross-checking against other trackers.
• SOCAMM2 is not yet a finalized standard. A late spec change could force costly rework on equipment already purchased.
• TrendForce's 60% coverage estimate is a 2027 forecast and will move as capacity announcements come in.
Next up: why Nvidia's inference-only Rubin CPX GPU skips HBM4 entirely and uses GDDR7 instead.
Sources: SK hynix 192GB SOCAMM2 press release (Apr 19, 2026); Samsung Semiconductor technical blog on SOCAMM2; JEDEC JESD328 (SOCAMM2) notice; SIGMAINTELL "Q2 2026 Global Memory Price Trends," via Seoul Economic Daily (Jun 23, 2026); TrendForce (Jun 2026); Tom's Hardware (Jun 2026).
Disclaimer: This post is for informational and educational purposes only. It does not constitute investment advice or a recommendation to buy or sell any security. All investment decisions are your own responsibility.
Comments
Post a Comment