Every guide below enforces our 3-Zone Street-Price Stoplight: buy verified Prime hardware at fair MSRP, inspect secondhand 24GB/48GB workhorses safely, or rent hourly cloud GPUs whenever retail markups destroy hardware payback math.
Buy the NVIDIA RTX 5060 Ti 16GB at $429 ($26.81/GB) and strictly avoid the $379 8GB SKU for local AI. Both share the GB206 die and 448 GB/s GDDR7 memory bus, but 16GB holds Qwen 2.5 14B Q4_K_M, OpenAI gpt-oss 20B in native MXFP4, and ComfyUI FLUX.1-dev weights entirely in VRAM without triggering 8x-slower PCIe x8 system RAM spillover.
A verified secondhand NVIDIA RTX 3090 24GB between $650 and $720 ($28.33/GB) remains the undisputed price-to-performance champion for 30B-class local LLMs. Its 384-bit GDDR6X interface delivers 936 GB/s bandwidth, sustaining 38–44 tok/s on Qwen 2.5 Coder 32B Q4_K_M and Gemma 4 31B, provided buyers replace dried backside thermal pads and pair an ATX 3.1 PSU against 550W transient spikes.
At its $599 MSRP ($37.44/GB), the AMD Radeon RX 9070 XT 16GB (RDNA 4 Navi 48, formerly rumored as RX 8800 XT) pushes 640 GB/s over a 256-bit GDDR6 bus—43% faster raw bandwidth than NVIDIA's $429 RTX 5060 Ti 16GB. It excels in Ollama and Vulkan llama.cpp (~49 tok/s on Qwen 2.5 14B), though Windows ComfyUI and FlashAttention-2 users still face ROCm friction.
Three graphics cards dominate sub-$500 local AI in 2026: Intel Arc B580 12GB at $249 ($20.75/GB, 456 GB/s) wins ultra-budget GGUF speed, MSI RTX 3060 12GB at $289 ($24.08/GB, 360 GB/s) wins zero-friction sub-$300 CUDA compatibility, and NVIDIA RTX 5060 Ti 16GB at $429 ($26.81/GB, 448 GB/s) unlocks 14B 32k coding and gpt-oss 20B. Skip every 8GB card.
Skip the $549 RTX 5070 12GB for serious local AI and buy the $749 RTX 5070 Ti 16GB whenever street pricing stays under $799. Qwen 2.5 14B Q4_K_M requires 8.9 GB of weights plus 6.0 GB of FP16 32k KV-cache and 1.5 GB CUDA overhead (16.4 GB total), crashing 12GB cards instantly while fitting comfortably inside 16GB at 13.4 GB with Q8_0 KV-cache at 62 tok/s.
Choose the $1,999 Apple Mac mini M4 Pro 48GB over a noisy discrete GPU tower if you run always-on 45W local coding agents like Qwen 3 Coder 30B-A3B (85+ tok/s in 4-bit MLX) or 70B Q4 models up to ~43 GB via sysctl wired-memory overrides. Build a discrete CUDA RTX rig instead whenever your workflow demands fast 64k prompt prefill or video diffusion.
Two used RTX 3090 24GB GPUs (~$1,360 combined) paired with an ATX 3.1 1200W PSU and shielded PCIe 4.0 x16 risers deliver 48GB of 936 GB/s VRAM, running unsplit Llama 3.3 70B Q4_K_M (42.5 GB) at ~21 tok/s for a quarter of workstation card pricing. Never use x1 mining risers if deploying vLLM tensor parallelism, and power-cap each 3090 to 280W via nvidia-smi.
Every power supply, shielded PCIe 4.0 riser, thermal phase-change pad, and NVMe scratch drive below links directly to an MPN-verified Amazon Prime listing with zero generic search clutter.
Target Workloads: Qwen 2.5 14B (32k Q8_0 KV), DeepSeek R1 14B, FLUX.1 Dev FP8/NF4
Quiet desktop node pairing a 180W RTX 5060 Ti 16GB with 64GB DDR5 system memory and headless framebuffer bypass for zero-OOM 14B coding agents.
Target Workloads: Gemma 4 31B Q4_K_M, Qwen 2.5 Coder 32B, Wan 2.1 14B 720p
Thermal-restoration, ATX 3.1 transient power, and structural anti-sag kit to turn a used ~$680 24GB RTX 3090 into a cool, reliable 32B server.
Target Workloads: Llama 3.3 70B Q4_K_M, DeepSeek R1 Distill 70B, Qwen 2.5 72B
Complete ATX 3.1 power, twin-axial shielded PCIe 4.0 x16 riser, 90-degree 12V-2x6, and 64GB DDR5 backbone for running two 24GB GPUs in TP=2.
Spin up an unquantized Llama 3.3 70B, Qwen 3 235B, or Wan 2.1 video pod in 30 seconds with billed-by-the-second SSH and Jupyter access.
Once Windows or Xorg reserves ~1.2 GB and CUDA allocates workspace buffers, an 8GB card leaves ~5.8 GB usable—barely holding an 8B Q4 model at short context and failing outright on 14B coders, FLUX.1 Dev, and Wan 2.1.
16GB is the baseline sweet spot for new retail GPUs (fitting 14B Q4_K_M models with 32k Q8_0 KV-cache plus FP8 image and video diffusion), while 24GB to 48GB unlocks dense 32B and 70B class models.
If your active GPU inference or fine-tuning time stays below 15–20 hours per month, renting an RTX 4090 ($0.39–$0.59/hr) or 80GB H100/A100 on RunPod or Vast.ai takes over three years to exceed local hardware depreciation.
Once model weights and KV-cache fit inside dedicated VRAM, batch-1 token speed depends almost exclusively on onboard GDDR6/GDDR7 memory bandwidth; PCIe generation mainly impacts initial disk-to-VRAM load time and multi-GPU tensor parallelism.