you don't need a $2,000 card to run AI at home. here's the honest budget GPU ladder, from $250 up

The most useful thing I can tell you about budget local-AI GPUs is that the benchmark scores barely matter. What matters is one number — VRAM — because it decides which models you can load at all, and a slower card with more memory beats a faster card with less every single time. So this isn't a spec dump ranked by tokens per second. It's a ladder: three cards under $500, what each actually runs, and how to pick your rung without needing to read a chart. Spoiler on the whole genre: for local AI, buy the most VRAM your budget allows, and stop optimising anything else.

The one rule that replaces every benchmark

VRAM sets your ceiling; everything else sets your speed. A model has to fit in your GPU's memory to run on the GPU at all — spill over and it crawls on system RAM at a few tokens per second, which nobody tolerates. So the first question isn't 'how fast is this card,' it's 'what's the biggest model it can hold.' A rough map at Q4 quantization: 8GB handles a 7–8B model tightly, 12GB comfortably fits 7–14B, and 16GB gives 13–14B real breathing room plus longer context. Check any specific model in the VRAM calculator before buying — it turns this from guesswork into a yes/no.

The rungs

The RTX 3060 12GB (~$250 used) is where I'd point most people starting out. It's the cheapest way onto the CUDA ecosystem, where every local-AI tool works with zero fuss, and its 12GB runs every 7–8B model and most 13–14B at Q4. It's not fast, but 'not fast' still means a perfectly usable conversation. For a first serious local-AI card, it's hard to beat on value.

The Intel Arc B580 (~$249 new) is the value wildcard: the same 12GB, newer silicon, a warranty, for used-3060 money. The catch is software maturity — Intel's stack has improved a lot but still trails CUDA, so expect more setup and the occasional tool that assumes NVIDIA. If you're comfortable troubleshooting and want new hardware for cheap, it's compelling. If you want it to just work, the 3060's boring CUDA reliability is worth the used-market gamble.

The RTX 5060 Ti 16GB (~$549) sits at the ceiling of 'budget,' and its extra 4GB matters more than the price bump suggests — hardware-corner measured it running 14B models with more context headroom than the 12GB cards, and it draws a modest ~180W with a full warranty. If you can stretch, the 16GB buys real future-proofing. Just make sure it's the 16GB variant, not the 8GB one — we compared it against a used 3090 if you're weighing new-16GB against used-24GB.

When to stop saving and just buy up

Honest advice a budget guide should give: sometimes the budget card is the wrong buy. If what you actually want is to run 27–32B models — the ones that feel genuinely capable — no sub-$500 GPU gets you there, because that needs 24GB and these cards top out at 16. Spending $549 on a 5060 Ti when your real goal needs a used 3090 means buying twice. Be honest about your target model first: if it's 14B-and-under, this ladder is perfect; if it's bigger, save the extra few hundred and skip a tier.

The whole budget game comes down to one honest question: what's the biggest model you actually want to run? Answer that in the VRAM calculator, buy the cheapest card that clears it, and ignore everything else the spec sheets shout about. If your answer turns out to be 32B, skip this ladder and read our used-3090 comparison instead — the right budget card is the one that fits your model, not the one with the biggest number in the review.