Affiliate note: as an Amazon Associate I earn from qualifying purchases, at no extra cost to you. Prices change constantly, so we do not list them. Click through for the current price.
Before you buy RAM
Will this RAM work in my PC?
RAM is the cheapest local AI upgrade, but only if you buy the right type. Five checks:
- DDR4 vs DDR5 are not interchangeable. The notch is in a different place, so a DDR5 stick will not fit a DDR4 slot. Your CPU platform and motherboard decide which one you need. AMD AM4 (Ryzen 1000 to 5000) is DDR4. AMD AM5 (Ryzen 7000 and newer) is DDR5 only. Intel 12th to 14th gen boards are either DDR4 or DDR5 depending on the board model, never both, so check yours.
- Desktop vs laptop size. Desktops use full size DIMMs. Laptops and most mini PCs use the short SO-DIMM. Same DDR type, different size, not swappable.
- Soldered memory cannot be upgraded. Many thin laptops and some mini PCs (and every Mac mini) use LPDDR5 soldered to the board. If the spec sheet says LPDDR or "onboard", there are no slots, so buy enough memory up front: 32GB minimum for local AI, 64GB or more for 30B+ models.
- Speed runs at the slowest link. RAM runs at the lowest speed supported by the CPU, the motherboard and the kit. A faster kit just downclocks, which is safe. Out of the box most kits run at a slower JEDEC default speed (for example DDR4 2133 or DDR5 4800). To get the rated speed, turn on XMP (Intel) or EXPO (AMD) in the BIOS.
- Capacity limits. Every board has a max per slot and a max total (common examples: 32GB per slot and 64GB total on older DDR4 boards, 48GB per slot and 192GB total on newer DDR5 boards; many mini PCs top out at 64GB or 96GB). The board maker's spec page lists it.
How to check what you have (2 minutes)
- Task Manager > Performance > Memory shows your total RAM, speed, slots used (for example "2 of 4") and form factor (DIMM or SO-DIMM).
- CPU-Z (free) > Memory and SPD tabs show DDR4 or DDR5 and the exact sticks installed.
- msinfo32 (press Windows key, type msinfo32) shows your motherboard maker and model ("BaseBoard Product"). Search that model plus "specifications" or "QVL" on the maker's site to see the supported RAM type, speed and max capacity.
Buy a matched kit. Mixing different kits, even the same size and speed, can cause crashes, failed boots or forced slow speeds. Replace with one matched 2 stick kit instead of adding a single stick to what you have.
You do not need a $5,000 “AI workstation” to run helpful local chat and light coding help on Windows. You need enough RAM, decent storage, and realistic model choices (see 16GB model guide). This page is framed as good enough for local chat/code help, not training foundation models.
Want specific models? See best mini PC for local AI, best budget GPU for Ollama, and the RAM and VRAM chart.
Build vs buy mini-PC
Buy a mini-PC (fastest path)
- Pros: Quiet, small, low power, fewer compatibility headaches.
- Cons: RAM sometimes soldered; GPU upgrades limited or impossible.
- Target: 32GB RAM configs when the budget allows; 16GB works if you stick to 7B–8B models.
DIY / tower build
- Pros: Swap RAM/GPU later; used GPU market can be a steal.
- Cons: Time, cable management, PSU sizing, more ways to mis-order parts.
Rule of thumb: If you’ve never built a PC, buy a mini-PC or a simple prebuilt with upgradeable RAM. Spend brainpower on Ollama + Open WebUI, not debugging boot loops.
Parts list bands under $1,000
Prices move, treat these as budget bands, not a static cart.
Path 1: Mini-PC ready (~$500–$900)
- Mini-PC with modern AMD/Intel CPU
- 32GB RAM preferred (16GB minimum)
- 512GB–1TB NVMe
- Optional: USB SSD for model overflow
Outcome: solid Tier A/B local chat; quiet enough for a desk or shelf.
Path 2: Upgrade the PC you own (~$80–$400)
- 32GB RAM kit matched to your laptop/desktop
- 1TB NVMe if storage is full/slow
- Skip GPU until RAM and disk are healthy
Often the best ROI, don’t buy a new box if RAM was the bottleneck.
Path 3: Tower + used GPU (~$700–$1,000 all-in)
- Modest used/refurb CPU + B-series / mid chipset board
- 32GB DDR4/DDR5
- 1TB NVMe
- Used GPU with 8GB+ VRAM (watch PSU watts and case clearance)
- Decent 550W+ known-brand PSU if adding GPU
Outcome: faster tokens/sec on 7B–14B models; still not a 70B playground.
What “good enough” means
- Private drafts, brainstorming, rewrite help, basic coding Q&A.
- Offline-capable once models are downloaded.
- Not realtime 4K video generation or training giant models.
Where to shop
Pair with software setup
Hardware alone isn’t a product. Install stack: Ollama + Open WebUI on Windows, then stay in the model sizes that fit: best models for 16GB RAM.