Guide · Local AI
Best GPU for local AI under $500
By Tiny · how I pick
As an Amazon Associate I earn from qualifying purchases. Buying through these links supports the site at no extra cost to you. Full disclosure
For running language models at home with Ollama or LM Studio, one spec decides almost everything: VRAM. If the model fits in the card's memory, it runs fast. If it spills into system RAM, it slows down by 5 to 20 times. CUDA core counts barely matter for this, and memory bandwidth comes a distant second. So under $500, the job is simple: get the most usable VRAM you can.
Prices moved a lot this year. Two cards I used to recommend here, the RTX 5060 Ti and a used RTX 3090, now cost well over $500, so they've moved to their own section below. If you already own a PC, run it through the Server Calculator first. You might not need a new card at all.
Words you'll see in this guide
- VRAM: the memory on your graphics card. AI models load into it, so more VRAM means you can run bigger models.
- Parameters (7B, 14B, 30B): a rough measure of a model's size. "8B" means 8 billion. Bigger models are usually smarter and need more memory.
- Quantization: shrinking a model by storing its numbers less precisely, so it fits in less memory with a small drop in quality. Think of saving a photo as a smaller JPEG. "4-bit" (Q4) is the common setting.
- Tokens per second: how fast a model writes. A token is roughly three-quarters of a word.
- CUDA: Nvidia's software for running AI on its graphics cards. Most local AI tools support it first and best.
How much VRAM you need
At the usual 4-bit setting, plan on about 0.6 GB of VRAM per billion parameters, plus a couple of GB for overhead. That works out to:
- 8 GB: 7B and 8B models, comfortably
- 12 GB: most 13B and 14B models
- 16 GB: 14B models with room for longer conversations
- 24 GB: 30B-class models on a single card
Pick the card for the model size you'll actually use day to day.
The two picks that are actually under $500
RTX 3060 12 GB (used)
Still the cheapest way into 12 GB of Nvidia VRAM. A used RTX 3060 12 GB fits every 7B and 8B model and most 13B and 14B models at 4-bit. Community benchmarks put it at roughly 40 to 55 tokens per second on an 8B model, which is faster than you can read. CUDA support is mature, so Ollama and LM Studio work without fiddling. It's also the card I picked for my own AI rig.
Not for you if: you want 30B models. 12 GB isn't enough, and no amount of tuning changes that.
Check price on AmazonIntel Arc A770 16 GB
The Intel Arc A770 16 GB is now the only new 16 GB card I can find under $500. Ollama and LM Studio both run on it. Intel's software is less polished than Nvidia's CUDA, so expect the occasional setup problem and fewer guides when something breaks.
Check before you buy: the A770 also comes in an 8 GB version, and listings don't always make that obvious. Make sure the one you order says 16 GB.
Check price on AmazonWorth it if your budget stretches
These two used to sit in the top picks. They're still good cards for local AI, but neither is anywhere near $500 right now.
-
RTX 5060 Ti 16 GBThe newest card with 16 GB. It launched at $429, but new cards were going for roughly $620 to $790 at retailers as of October 2026. You get Nvidia's newer tensor cores with FP4 support and much more memory bandwidth than the old 4060 Ti, which means faster replies. Worth it if you want a warranty and the price drops back toward launch.Check price on Amazon
-
RTX 3090 24 GB (used)The cheapest route to 24 GB, which runs the 30B-class models a 16 GB card can't. Two of them pool to 48 GB for a 70B model. Used prices climbed all year, from about $700 last winter to roughly $1,200 on eBay as of October 2026. It also draws a lot of power and needs a big power supply.Check price on Amazon
Side by side
| Card | VRAM | Runs comfortably | Under $500? |
|---|---|---|---|
| RTX 3060 (used) | 12 GB | 7B to 14B models | Yes |
| Intel Arc A770 | 16 GB | 14B models | Yes |
| RTX 5060 Ti | 16 GB | 14B with headroom | Not right now |
| RTX 3090 (used) | 24 GB | 30B-class, or 70B with two | No |
Price ranges above come from retailer and eBay trackers in October 2026, not from Amazon. They change weekly, so check the current price before you buy.
What I'd buy
- Just starting out: a used RTX 3060 12 GB. Nothing else gets you this much model for the money.
- Want 16 GB and can live with some troubleshooting: the Intel Arc A770 16 GB.
- Want 16 GB on Nvidia with a warranty: wait for the RTX 5060 Ti to come back down toward $429.
- Set on 30B-class models: a used RTX 3090, and budget well past $500 for it. Before you do, read whether a used RTX 3090 is still worth it at today's prices.
Whatever you pick, buy for VRAM first. Once you have the card, the next step is building the rest of the home AI server around it.
Free email updates
GPU prices moved a lot this year. Subscribe and I'll tell you when a card on this list drops back under $500.
No spam. Unsubscribe anytime. Privacy policy.