At $1,999 the RTX 5090 is an easy yes for local AI. Nobody will sell you one at $1,999

The short answer

At its $1,999 MSRP the RTX 5090 is the best single card for local AI, full stop — 32GB of VRAM and measured llama.cpp generation of 102.7 t/s on Qwen3-14B at 16k context (hardware-corner.net, March…

The 32GB sweet spot is real: it runs 32B-class models at long context with headroom no 24GB card has. It still cannot hold a dense 70B at Q4 — that's ~40GB. 32GB is a bigger boat, not a different…

At $3,699, two used 3090s deliver 48GB for ~$1,700 less, and our live tracker shows rented 5090s at $0.69/hr on-demand — the street price buys ~5,400 rental hours, about 2.6 years of 40-hour weeks.

Buy it anyway if: you also do image/video generation (Blackwell's compute advantage there is larger than the LLM one), you want one quiet-ish card instead of a multi-GPU project, or you actually…

Skip it if: your models fit in 24GB (a ~$1,000 used 3090 does that job), or your ambitions are 70B+ (offload, dual cards, or unified memory do it better per dollar).

Worth it at MSRP. Not at street.

The measured performance is real and the 32GB long-context headroom is unique among consumer cards. But value is a function of price, and at $3,699 the same silicon scores lower than it did at…