Q4 NAND Prices Seen Up 30%, DRAM 20% as Memory Squeeze Widens

Research houses forecast NAND prices up 30% and DRAM contract prices up about 20% in the fourth quarter. The memory squeeze is now visible everywhere: PC cafes are deferring upgrades, Microsoft and Meta are rethinking memory efficiency, Amazon’s new GPUs sit idle for lack of power, and Micron and AMD are designing new memory for robots and AI servers.

Key takeaways

• Research houses forecast Q4 NAND prices up 30% and DRAM contract prices up about 20% quarter on quarter, with gains moderating from earlier in 2026.
• Nvidia is reportedly re-evaluating 12-layer HBM4e for Rubin Ultra, also weighing 8-layer HBM4e and HBM4, a supply-chain report says.
• Microsoft is exploring composable memory and Meta is reportedly reusing DDR4 via CXL as high memory prices force new efficiency approaches.
• Amazon is taking delivery of roughly $5 billion in Nvidia GB300 racks that it cannot yet power, Business Insider reports.
• Micron is developing edge inference memory for robots, targeting mass production in the second half of 2028; AMD’s 2027 EPYC Verano will use LPDDR5X.

Q4 NAND seen up 30%, DRAM contract prices up 20%

Memory prices keep climbing despite rising market doubts. According to Taiwan’s Commercial Times, research houses forecast fourth-quarter NAND prices up 30% and DRAM contract prices up about 20% quarter on quarter, supported by AI inference demand and limited new supply at mature process nodes.

The pace is moderating from roughly 80% in Q2 and 40% in Q3, the institutions say, as prior increases, weakening end-market tolerance and price caps in long-term supply agreements (LTAs) take effect.

Nvidia is reportedly re-evaluating Rubin Ultra’s memory configuration: beyond the planned 12-layer HBM4e, it is weighing 8-layer HBM4e and HBM4. Dropping from 12 to 8 layers would cut HBM capacity per GPU by about 33%, but the same DRAM supply could support roughly 50% more GPU shipments, the report notes.

As AI shifts from training to large-scale inference, growing context windows and KV cache needs keep lifting server DRAM and enterprise SSD demand. With the big three makers tilting capital spending toward high-end memory, mature-node supply additions remain limited — a structure research houses expect to prove more resilient than past cycles.

Rising PC hardware costs put internet-cafe upgrades on hold

Consumer hardware is feeling it too. A Nandu N-video survey of Guangzhou’s computer markets found a RMB 12,000 budget no longer comfortably buys a mid-to-high-end PC with an RTX 5070 or better; merchants say prices accelerated after August, some GPUs rising nearly RMB 1,000 within weeks.

Internet cafes are deferring large-scale refreshes, cutting mid-range machine memory from 48GB to 32GB or 24GB, or reusing retired graphics cards in general seating. Newly opened venues face the heaviest pressure, and the hoped-for traffic boom has not materialized as competition and the post-summer lull keep customer traffic declining at some venues.

High prices push Microsoft and Meta toward memory-efficiency ideas

Tight supply is forcing new efficiency approaches. Microsoft chief architect Samir Rajadnya said composable memory — pooling memory across servers and allocating it dynamically by workload rather than binding it to one CPU or GPU — will be a key direction. Its effectiveness depends on high-speed CPU-to-XPU interconnect, where optical links could help; dynamic allocation across workloads remains to be worked out.

Meta is reportedly exploring CXL to reuse DDR4 memory slated for retirement, adding cheap capacity for less demanding workloads.

Such techniques could improve utilization, but their effect on supply, demand and prices depends on deployment scale and remains uncertain.

Amazon stockpiles $5 billion in GB300 racks it cannot yet power

The AI buildout is hitting a power wall. Per Business Insider, Amazon took delivery of roughly $5 billion of Nvidia GB300 racks in August and expects about $3 billion more each month under a new Early Rack Receipt program. Some AWS sites have rack space but not enough electrical capacity, leaving finished equipment staged in warehouses as dark GPUs.

Trailing twelve-month free cash flow fell from about $50 billion in Q1 2024 to negative $7.6 billion in Q2 2026, while 2026 capital expenditure plans still run near $200 billion; AWS CFO John Felton reportedly expects the pressure to persist for years.

Demand is not the constraint — the source report cites AWS second-quarter revenue up 37% year on year with a $496 billion backlog. The challenge is synchronizing chip deliveries, data-center construction and power availability.

Micron develops edge inference memory as robots emerge as a memory market

Physical AI is becoming a new memory battleground. Micron VP Lee Cheol said at SEDEX 2026 on October 16 that the company is developing edge inference memory (EIM) — low-power DRAM stacked on a logic die to cut latency and raise bandwidth — and is exploring it for LPDDR6. Industry observers expect mass production in the second half of 2028.

Micron expects future high-performance humanoids to need about 200GB of DRAM and 4TB of storage, versus 12GB/256GB in today’s service robots, and says advanced robots may eventually adopt HBM. Counterpoint Research estimates humanoid-robot DRAM demand growing more than twentyfold between 2026 and 2030.

Dell plans 3–5% monitor price increase from October

Dell has notified partners of a shipment-price increase of roughly 3% to 5% from October 2026, mainly for mid-to-high-end monitors, according to ChannelGate; entry-level models are expected to be excluded. Dell had absorbed a July cost increase without raising prices.

Shipment-price increases do not guarantee immediate retail increases: distributors and retailers may absorb part of the cost in their margins, so the timing and extent of any consumer-level price rise remain uncertain.

AMD’s next-generation EPYC “Verano” targets LPDDR5X server memory

Low-power DRAM is moving into AI servers. AMD unveiled its next-generation EPYC “Verano” server CPU at the OCP Global Summit on October 12 — a Zen 6-based processor planned for 2027 and aimed at AI accelerator host nodes. Verano is the first EPYC to support the SOCAMM2 memory module, which pairs LPDDR5X’s low power and high bandwidth with server-grade replaceability and maintenance.

The design mirrors NVIDIA, which also plans to adopt SOCAMM2 in its next-generation AI platform. AMD has not confirmed final specifications — reported core counts vary between 72 and 192.

Sources

• Commercial Times, October 20, 2026 — research-house forecast of Q4 NAND +30%, DRAM contract +20% — https://www.ctee.com.tw/news/20261020700183-439901
• Business Insider, October 20, 2026 — Amazon’s Early Rack Receipt program and GB300 racks awaiting power — https://www.businessinsider.com/amazon-nvidia-ai-gpu-data-centers-power-shortages-2026-10
• The Elec, October 18, 2026 — Micron developing EIM for physical AI and humanoids — https://www.thelec.net/news/articleView.html?idxno=14669
• TrendForce, October 19, 2026 — Micron EIM overview and 2H28 production outlook — https://www.trendforce.com/news/2026/10/19/news-micron-develops-edge-inference-memory-2h28-production-seen-as-hbm-potential-grows-in-humanoids/
• TechPowerUp, October 13, 2026 — AMD EPYC Verano with LPDDR5X SOCAMM2 at OCP 2026 — https://www.techpowerup.com/353662/amd-showcases-its-next-gen-epyc-verano-cpu-up-to-192-cores-and-lpddr5x-memory
• ITHome, October 13, 2026 — Verano 2027 launch, Zen 6, LPDDR5X — https://www.ithome.com/1/012/535.htm

FAQ

How much will memory prices rise in Q4 2026?

Research houses cited by Taiwan’s Commercial Times forecast NAND prices up 30% and DRAM contract prices up about 20% quarter on quarter — a slower pace than the roughly 80% and 40% DRAM gains of Q2 and Q3, as prior increases, end-market resistance and price caps in long-term agreements moderate the rally.

Why are Amazon’s new Nvidia GPUs sitting in warehouses?

According to Business Insider, AWS is taking delivery of GB300 racks before its data centers have enough power, under an “Early Rack Receipt” program worth about $5 billion in August and roughly $3 billion a month thereafter. Racks wait in intermediary warehouses as “dark GPUs,” tying up cash: trailing twelve-month free cash flow swung to negative $7.6 billion in Q2 2026.

What are composable memory and EIM, and will they ease the shortage?

Composable memory pools server memory across workloads instead of binding it to one CPU or GPU — an approach Microsoft is exploring — while Meta is reportedly reusing DDR4 via CXL and Micron’s EIM stacks LPDDR on a logic die to cut latency for robot AI. All could improve utilization, but their impact on supply, demand and prices depends on deployment scale and remains uncertain.

Will LPDDR5X come to AI servers?

Yes. AMD’s EPYC Verano, unveiled at the OCP Global Summit on October 12 and planned for 2027, is the first EPYC processor to support LPDDR5X through serviceable SOCAMM2 modules, and NVIDIA is also adopting SOCAMM2 for its next-generation platform. Final specifications, including core counts, remain unconfirmed.