First-class
NVIDIA (desktop / pro)
CUDA is the default path: Ollama, llama.cpp (CUDA), vLLM, SGLang. Best docs, GGUF ecosystem, and agent tooling. Prefer NVIDIA when the budget is for a coding desk GPU.
Local coding
ハードウェア帯
The lock is memory, not TFLOPS. Buy the band that fits weights plus KV for your real context—or keep paying the API.
Protocol
- VRAM / unified memory is the lock—not TFLOPS.
- Weights fitting the card ≠ target context fitting (KV tax).
- Local token stickers are ~$0; you still pay electricity and depreciation.
Memory bands
Accent column = USD reference range for the memory band (mid in parentheses).
| Band | USD ref (band) | Typical class | Max params (Q4 class) | Workloads |
|---|---|---|---|---|
| 16GB | $250–$750 ($450) | RTX 4060 / 5060 Ti 16GB class / Mac 16GB unified | 8B | Tab complete / short chat |
| 24GB | $1.4k–$3.2k ($2.2k) | RTX 4090 24GB / Mac 24–32GB unified | 32B | Tab complete / short chat · Repo coding agent |
| 32GB | $2k–$5.5k ($3.8k) | RTX 5090 32GB class / Mac 32–36GB unified | 40B | Tab complete / short chat · Repo coding agent · Long-context review |
| 48GB | $2.5k–$7k ($4.5k) | 48GB workstation / dual-24 class / Mac 64GB unified | 70B | Tab complete / short chat · Repo coding agent · Long-context review |
| 80GB | $6k–$14k ($9k) | 80GB-class pro card / Mac Studio 96–128GB | 120B | Tab complete / short chat · Repo coding agent · Long-context review |
| 128GB | $8k–$18k ($12k) | Mac Studio 128–192GB / multi-GPU desk | 180B | Tab complete / short chat · Repo coding agent · Long-context review |
| 512GB | $18k–$45k ($28k) | Mac Studio 512GB / multi-GPU rack | 480B | Tab complete / short chat · Repo coding agent · Long-context review |
Representative SKU reference prices (USD)
GPU-card or device-class streets. Mac rows are whole-device unified memory, not a discrete GPU sticker.
| Band | SKU / class | USD ref |
|---|---|---|
| 16GB | RTX 4060 Ti / 5060 Ti 16GB class (card) | $300–$550 |
| 16GB | Used RTX 3090 / 3080 Ti class (often 12–24GB) | $400–$900 |
| 24GB | RTX 4090 24GB (used/new street) | $1.5k–$2.8k |
| 24GB | Used dual-12GB / single-24GB workstation class | $1.2k–$2.5k |
| 32GB | RTX 5090 32GB (MSRP vs street — volatile) | $2k–$5.1k |
| 48GB | Dual used RTX 3090 (2×24GB) system delta | $2k–$3.5k |
| 48GB | Mac Studio / Pro ~64GB unified (device, not GPU-only) | $2.5k–$4.5k |
| 80GB | Pro / datacenter 80GB-class (A100 80 / H100 80 street) | $8k–$25k |
| 80GB | Mac Studio Ultra ~96–128GB unified (device) | $5k–$9k |
| 128GB | Mac Studio high unified config (device) | $7k–$15k |
| 128GB | Multi-GPU desk (2×32–48GB class) | $7k–$16k |
| 512GB | Mac Studio max unified / mini-cluster class | $15k–$40k |
NVIDIA · AMD · Intel · Apple
Local coding inference is CUDA/Metal-first. AMD and Intel can work in limited setups—read the “why” before you buy for that reason alone.
Limited support
AMD (Radeon / Instinct)
Usable, not first-class. ROCm coverage is uneven by card/OS; many coding apps and GGUF paths expect CUDA. llama.cpp Vulkan/ROCm builds can run, but expect slower bring-up, fewer “it just works” guides, and weaker vLLM/SGLang desk stories. Buy AMD only if you already own the card or accept tinkering—not as the cheapest path to a Cursor-like local agent.
Limited support
Intel (Arc / Xe / discrete)
Experimental for serious local coding agents. SYCL/oneAPI and llama.cpp backends exist, but Arc VRAM ceilings, driver maturity, and missing CUDA-centric tooling make Intel a poor primary buy for SWE-style agents. Fine for light experiments if the GPU is already in the machine; do not plan a coding workstation around Arc alone.
First-class (unified memory)
Apple Silicon
Not a discrete GPU vendor path—unified memory + MLX (or llama.cpp/Ollama). Expensive $/GB but simple ops. See band and SKU tables for device-class USD ranges.
TCO vs API
Illustrative break-even: monthly capex (3-year) plus electricity versus one 100K+20K coding job on OpenRouter.
| Band | Mid capex (class) | Break-even jobs / month |
|---|---|---|
| 16GB | $450 | 1112 |
| 24GB | $2.2k | 4970 |
| 32GB | $3.8k | 8497 |
| 48GB | $4.5k | 10040 |
| 80GB | $9k | 19961 |
| 128GB | $12k | 26575 |
| 512GB | $28k | 61848 |
Do not buy
- ≤16GB laptop forcing 70B weights.
- 24GB card expecting 480B-class coder downloads.
- Personal SWE buying H100 / cluster class for tab-complete.
Assumptions and sources
Hardware bands are classes, not street quotes. VRAM estimates use curated GGUF footprints plus size-class KV factors. API job cost uses OpenRouter primary listing when present. choose-ai-infrastructure-guide covers API hosts—not desk GPUs.
Hardware USD figures are curated mid-2026 reference ranges for planning. Verify current street price before buying.