This guide used to rank four cards. In 2026 it ranks one, because the GDDR7 shortage lifted everything else above the line: the RTX 4060 went from $280 to $479, and the Intel Arc B580 from $250 to $310. Neither is a bad card. Neither is under $300 any more.
Quick answer: A used RTX 3060 12GB (~$250) is the only GPU still under $300 in 2026, and it is a genuinely good one for the money — 12GB of VRAM runs Stable Diffusion, Flux at Q8, and 13B LLMs at Q4. If you can stretch to $310 the Intel Arc B580 gives you the same 12GB on a new card with a warranty.
NVIDIA GeForce RTX 3060 12GB
12GB GDDR612GB VRAM at roughly $250 used — the only card left under $300, and more memory than the new cards that outgrew this budget.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
What can you actually run under $300?
Managing expectations matters here. Here is what each VRAM tier handles:
| VRAM | What runs | What does not |
|---|---|---|
| 8GB | SD 1.5, SDXL (optimized), 7B LLMs (Q4), Whisper | Flux at full quality, 13B+ LLMs comfortably |
| 12GB | SD 1.5, SDXL, Flux (Q8), 13B LLMs (Q4) | 32B+ models, video generation |
| 16GB | Everything up to 20B models, Flux full quality | Very large video models |
The used RTX 3060 12GB sits in the middle row, which is why it survives as a recommendation at all. The 4GB it holds over an 8GB card is the difference between running a 13B model and not.
The one card left: used RTX 3060 12GB
At roughly $250 on the used market you get 12GB of GDDR6 — more VRAM than the RTX 4060 that now costs nearly twice as much.
What it runs well:
- Stable Diffusion 1.5 and SDXL at comfortable speeds
- Flux.1 Schnell, and Flux.1 Dev at 8-bit
- 7B LLMs via Ollama (Llama 3.1 7B, Mistral 7B, Qwen2.5 7B)
- 13B LLMs at Q4, which an 8GB card cannot load at all
- Whisper transcription, embeddings, Real-ESRGAN upscaling
Why VRAM matters more than compute at this budget:
- Flux.1 models run at higher quality settings with 12GB vs 8GB
- 13B LLMs (Q4) fit in 12GB where they do not in 8GB
- Fewer out-of-memory crashes when running ComfyUI with multiple models loaded
The trade-offs are the usual ones for used Ampere: slower than Ada Lovelace on compute-heavy work, no Frame Generation, and no warranty. For inference, where memory capacity and bandwidth decide more than shader throughput, the generational gap is smaller than gaming benchmarks suggest. Run through the used GPU buying checklist before you pay, and note that NVIDIA’s RTX 3060 relaunch has put new units back on shelves if you would rather not buy secondhand.
NVIDIA GeForce RTX 3060 12GB
12GB GDDR612GB VRAM at roughly $250 used — the only GPU still under $300, and more memory than new cards costing twice as much.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
What left this price point
Two cards that anchored earlier versions of this guide are simply no longer here:
| GPU | Was | Now | What happened |
|---|---|---|---|
| RTX 4060 (8GB) | ~$280 | ~$479 | GDDR6 supply squeezed by AI accelerator demand |
| Intel Arc B580 (12GB) | ~$310 | ~$310 | Popular as a budget inference card; demand followed |
The RTX 4060 is the more awkward of the two. At $479 for 8GB it is now more expensive than the RTX 4060 Ti 16GB at ~$425, which has twice the VRAM. If you were planning to buy a 4060, buy the 4060 Ti 16GB instead — it is cheaper and it is the better AI card. That inversion is a shortage artifact rather than a considered product decision, and it may not last, but it is what the market looks like today.
See the full budget GPU guide for the wider picture across budget tiers.
If you can stretch to $310: Intel Arc B580
The Intel Arc B580 ($310 new) is the closest thing to a new card at this budget, and $60 over the used RTX 3060 buys you a warranty. It ships with 12GB GDDR6 — the same capacity as the 3060, on newer silicon with more bandwidth. Intel has invested significantly in oneAPI and IPEX-LLM for Arc GPU inference.
Where it works well:
- Ollama with Intel extension (IPEX-LLM) runs 7B-13B models
- Stable Diffusion via DirectML or OpenVINO
- Image inference workloads
Where it falls short:
- CUDA does not run on Arc — some AI tools are Nvidia-only
- Performance in PyTorch training is weaker than equivalent Nvidia cards
- Community support and tutorials assume Nvidia by default
If your workflows are inference-only (running models, not training them) and you are comfortable troubleshooting compatibility, the Arc B580 is genuinely worth considering. For a detailed breakdown of which AI workloads it handles and which it struggles with, see our Intel Arc B580 AI performance guide. For anyone who wants the path of least resistance, stick with Nvidia.
Check Intel Arc B580 on Amazon→Buy on Shopee SG→Comparing the options, in and just above budget
| GPU | VRAM | Bandwidth | Price | Best for |
|---|---|---|---|---|
| RTX 3060 12GB (used) | 12GB | 360 GB/s | ~$250 | The only card under $300 |
| Intel Arc B580 | 12GB | 456 GB/s | ~$310 | Same VRAM, new, with a warranty |
| RTX 4060 Ti 16GB | 16GB | 288 GB/s | ~$425 | Most VRAM in the budget bracket |
| RX 7800 XT | 16GB | 624 GB/s | ~$450 | 16GB with real bandwidth, if ROCm suits you |
| RTX 4060 | 8GB | 272 GB/s | ~$479 | Nothing — the 4060 Ti 16GB is cheaper |
Only the first row is actually under $300. The rest are here so you can see what each additional step buys, and how little sense the RTX 4060 makes at its current price. For the full picture one tier up, see the best GPU for AI under $500 guide.
Which GPU should YOU buy?
- Your ceiling really is $300: RTX 3060 12GB used. It is the only option, and it is a good one — 12GB runs more than most people expect.
- You can go to $310 and want a warranty: Intel Arc B580. Same 12GB, new card, but check your tools support Arc before buying.
- You can go to $425: RTX 4060 Ti 16GB. Sixteen gigabytes changes what you can load, and it costs less than the 8GB RTX 4060.
- You want to run Flux.1 Dev at full quality: 12GB is the floor, so the 3060 is the minimum entry — but 16GB makes it comfortable.
- You were about to buy an RTX 4060: Do not. At ~$479 it is outpriced by the 4060 Ti 16GB with twice the VRAM.
Common mistakes to avoid
- Buying an 8GB card thinking VRAM does not matter for “small” models. A 7B model at Q4 uses about 4-5GB of VRAM — but ComfyUI, your OS, and browser are competing for the rest. 8GB gets tight fast.
- Refusing to buy used. At this budget that decision costs you the entire bracket — there is no new card under $300 in 2026. If used is genuinely off the table, your real budget is $310 and the Arc B580 is the answer.
- Working from a 2025 price list. The RTX 4060 at $280 and the Arc B580 at $250 were both true a year ago and are both wrong now. Check current listings before committing to any budget guide, this one included.
- Expecting 13B model quality from a $300 budget. It is achievable on 12GB, but only with quantization. Do not expect GPT-4-level outputs from a budget GPU — the model matters more than the card at inference time.
- Buying AMD RX 6600/6700 for AI. These cards have good gaming performance but weaker AI inference support compared to Nvidia CUDA. Unless you know your tools support ROCm, stick to Nvidia at this price point.
Final verdict
| Priority | Pick |
|---|---|
| Under $300, full stop | RTX 3060 12GB used (~$250) |
| New card with a warranty | Intel Arc B580 (~$310) |
| Most VRAM if you can reach it | RTX 4060 Ti 16GB (~$425) |
Under $300 you can still run a useful local AI setup, but the bracket has narrowed to a single card. A used RTX 3060 12GB handles Stable Diffusion, Flux at Q8, and 13B LLMs at Q4 — which is more than the price suggests. If you can find another $60, the Arc B580 does the same on new silicon with a warranty behind it.