Updated August 12, 2026 · reviewed against primary sources
Buyer's guide
Best DDR5 for local AI
Disclosure: product links on this page are Amazon affiliate
links — as an Amazon Associate this site earns from qualifying purchases. Links go to
live searches, not sponsored picks; prices move too fast to quote.
Two numbers govern a local-AI memory buy: capacity (what fits) and bandwidth (how fast it generates — see the bandwidth explainer). Everything below serves those two, in order of the decisions you'll actually make.
Pick your capacity tier
| Tier | What it runs | Configuration |
|---|---|---|
| 64GB | 8–14B models generously; 32B quantized with room for context and life | 2 × 32GB — the default answer |
| 96GB | 32B comfortably, 70B quantized (slowly — bandwidth, not capacity, is the complaint) | 2 × 48GB — the quiet sweet spot |
| 128GB | 70B + big context, multiple models resident | 2 × 64GB where supported — keeps 2-DIMM speed |
| 192GB | Ceiling hosting for MoE/huge-context, speed-tolerant work | 4 × 48GB — accept the 4-DIMM speed penalty knowingly |
Speed: the 6000–6400 CL30 sweet spot
On current mainstream platforms, DDR5-6000 to 6400 at CL30–32 is where price, stability, and the memory controller's happy zone intersect (AMD's controllers run 1:1 in this window; pushing past it buys latency compromises and training drama). Faster kits (7200–8000+) are real token-speed gains if your platform validates them — check the board's QVL before paying the premium.
Three rules that save regret
- Two DIMMs, not four. Populating all four slots typically forces speeds down — the opposite of what AI workloads want. Hit capacity with bigger modules, not more of them (128 vs 192 walks the tradeoff).
- Dual-rank helps. 32GB/48GB modules are typically dual-rank, which usually benchmarks a few percent ahead in bandwidth-bound work — a free win that comes with the capacity you wanted anyway.
- EXPO/XMP is not optional. Out of the box, kits run JEDEC-slow; enable the profile or you paid for speed you're not getting.
When DDR6 changes this page
Consumer DDR6 (2027+, timeline) roughly doubles the
bandwidth column and this guide gets rewritten. Nothing you buy above is wasted in the
meantime — mature-generation DRAM is the value point of the whole cycle
(the price curve).