This article was originally published on runaihome.com
TL;DR: DDR5 RAM hit a $375 floor for a 32GB kit in June 2026 — roughly 4× what it cost 18 months ago. Phison's CEO confirmed all of 2026's NAND production is already sold out, and TrendForce projects another 58–63% QoQ DRAM increase for Q2. The cause is structural: HBM for AI accelerators consumes 23% of all global DRAM wafer capacity at a 3:1 wafer ratio vs DDR5. Relief isn't coming before late 2027.
| Mid-2025 | June 2026 | Delta | |
|---|---|---|---|
| 32GB DDR5 kit | ~$80–$90 | $375 floor | +317–369% |
| 64GB DDR5 kit | ~$150–$170 | $630–$679 | +299–352% |
| 1TB Gen4 NVMe | ~$70–$80 | ~$180–$250 | +125–257% |
| 2TB Gen4 NVMe | ~$120–$150 | ~$280–$380 | +87–217% |
Honest take: Buy the RAM and storage you need now — you'll spend less today than in six months. But if budget is the constraint, a used Ryzen 5000 / AM4 platform with DDR4 still runs local LLMs fine. GPU VRAM is still the actual bottleneck for inference; don't sacrifice your GPU budget to pay a 4× premium for DDR5.
Why HBM Is Eating Your DDR5 Budget
The mechanism behind the shortage explains why this won't self-correct quickly.
High Bandwidth Memory — the stacked DRAM inside every H100, MI300X, and B200 — requires roughly three times the silicon wafer area per gigabyte compared to standard DDR5. Samsung, SK Hynix, and Micron have been systematically converting production lines from consumer DRAM to HBM. The economics are not subtle: HBM earns 3–5× more revenue per wafer than DDR5, and every hyperscaler running an AI buildout is prepaying for it.
The result: HBM now consumes 23% of all global DRAM wafer capacity, up from roughly 10% in 2023, according to TrendForce. That's nearly a quarter of the world's DRAM manufacturing output going to a product that doesn't fit your motherboard.
GDDR7 adds further pressure. The RTX 50-series launched with GDDR7, pulling additional wafer allocation away from DDR5. AI data centers alone are expected to consume 20% of total DRAM wafer output in 2026, with HBM and GDDR7 together accounting for most of that. Meanwhile, overall DRAM supply growth is running at just 16% year-over-year — well below historical norms — because manufacturers are optimizing for HBM revenue, not volume.
New DRAM fabrication lines take 2–3 years to build and qualify. The financial incentive to build consumer DDR5 capacity instead of HBM capacity doesn't exist at current price differentials. This shortage is structural, not cyclical.
The Numbers: How Bad Is It?
DDR5: TrendForce recorded conventional DRAM contract prices up 90–95% quarter-over-quarter in Q1 2026, the largest single-quarter jump on record. Q2 came in at another 58–63% QoQ increase. A 32GB DDR5 kit that bottomed out below $80 in mid-2025 now has a confirmed floor of $374.97 on PCPartPicker as of early June 2026 — that's the cheapest available option (a Silicon Power Zenith RGB DDR5-6000 kit). Named-brand kits from Corsair and Crucial at 64GB are running $630–$679.
HP's CFO disclosed in a recent earnings call that memory and storage jumped from roughly 15% to 35% of its PC bill of materials in a single quarter. HP stated directly that memory costs doubled within that timeframe. Dell, HP, and Lenovo have all raised retail PC prices by 15–20% in response.
NAND flash: Phison CEO Khein-Seng Pua confirmed in late 2025 that NAND flash prices had more than doubled in six months — a single TLC 1-terabit NAND chip went from $4.80 in July 2025 to $10.70 by November 2025. The harder fact he disclosed: every NAND manufacturer told Phison that all of 2026's production capacity is already sold out. TrendForce projects a further 70–75% QoQ increase for Q2 2026 NAND contracts.
Consumer SSDs absorbed this upstream cost pressure with a 6–9 month lag. 1TB M.2 Gen4 drives that cost $70–$80 a year ago are now $180–$250 depending on model and retailer. 2TB options have followed proportionally.
What This Does to Local AI Build Costs
This is where the pain concentrates for home lab builders.
When our $500 local AI inference build guide was written, the RAM+storage slice of that build ran roughly $90–$120 total — enough budget headroom to also buy a GPU, CPU, and a cheap case. That same component combination now costs $450–$600. The $500 budget build is no longer a $500 build.
The $2,000 workstation guide is in a similar position. It assumed 64GB DDR5 + 2TB NVMe at around $280 combined. Those same components now cost $850–$1,050.
| Build tier (GPU budget unchanged) | 2025 RAM+storage | 2026 RAM+storage | Effective shortfall |
|---|---|---|---|
| $500 budget build | ~$110 | ~$540 | −$430 |
| $1,000 mid-range | ~$200 | ~$760 | −$560 |
| $2,000 workstation | ~$280 | ~$900–$1,050 | −$620–$770 |
The GPU side of the budget hasn't inflated at the same rate. RTX 40-series prices are relatively stable; used RTX 3090s remain in the $450–$550 range. Every dollar that migrates to DDR5 and NVMe is a dollar not going to GPU VRAM, which is the actual bottleneck for running larger local models. See the system RAM guide for minimum RAM requirements per model size, and the NVMe SSD loading guide for how storage speed actually affects inference.
Does DDR5 vs DDR4 Actually Matter for Local AI?
Worth asking directly before spending $679 on a 64GB DDR5 kit.
For LLM inference, the answer splits on workload type:
GPU-resident inference (the model fits in VRAM): system RAM speed is nearly irrelevant. The GPU's onboard memory handles all token generation; system RAM only moves model weights at initial load time, which happens once per session. DDR4 vs DDR5 makes no measurable difference here.
CPU-offload inference (llama.cpp with mixed GPU+CPU layers): RAM bandwidth matters. DDR5-5600 dual-channel delivers roughly 89 GB/s. DDR4-3200 dual-channel delivers roughly 51 GB/s. That bandwidth difference translates to approximately 2–4 tok/s more on a 70B Q4_K_M model being partially offloaded — real, but not decisive for most use cases.
The practical implication: if your build has a GPU with enough VRAM for your target model — an RTX 3090 for models up to 24GB, for instance — the DDR4 vs DDR5 difference is below the threshold of usefulness. GPU VRAM handles the inference; RAM is overhead.
A used Ryzen 9 5950X on an AM4 board with 64GB DDR4-3200 runs approximately $400–$450 total for CPU plus RAM. The AM5 + DDR5 equivalent costs $800+ for RAM alone, plus $300–$400 for a Ryzen 9000-series CPU and AM5 board. The AM4 path leaves $700–$900 more for a better GPU.
What to Actually Buy Right Now
RAM
On a DDR5 platform (AM5 or Intel 13th/14th/Core Ultra):
Buy now, not later. Q3 2026 prices will be higher. The 64GB (2×32GB) DDR5-6000 configuration is the target for local AI builds that do any CPU offload. A Corsair Vengeance 64GB DDR5-6000 kit is running $630–$679. DDR5-4800 entry kits are slightly cheaper but offer limited bandwidth headroom; the price difference to DDR5-6000 doesn't justify the step-down.
If you're building new and budget is tight:
AM4 / Ryzen 5000 with DDR4 is a legitimate choice. A used Ryzen 9 5900X paired with 64GB DDR4-3600 runs $380–$420 combined — $250–$300 less than the DDR5 kit alone. AM4 CPUs run llama.cpp without issue. The throughput trade-off vs AM5 for pure inference is small; the budget savings are large.
Storage
For model storage, Gen4 NVMe is sufficient. There is no meaningful benefit to PCIe Gen5 for loading LLM weights — VRAM saturation, not disk bandwidth, determines load time at any sequential read speed above ~3,500 MB/s. A Gen5 SSD at 12
Top comments (0)