If you want to run language models on your own machine, the first number to check is VRAM, and the second is whether the card speaks CUDA. Raw gaming speed matters far less than people expect. A fast card that cannot fit the model in memory is a slow card for this job. Below we sort five cards by the kind of buyer each one actually suits, using launch MSRPs, memory sizes and board power. Treat the prices as launch prices only, because the used and discounted market moves on its own schedule.
If you want the biggest models one card can hold
The GeForce RTX 5090 is the only consumer card with 32GB of VRAM, in this case GDDR7. That extra 8GB over the 24GB cards is the whole argument. It lets 32B-class models run comfortably with room to spare, and nothing else in a consumer slot goes higher. It launched at $1,999, and street prices frequently sit above that.
The cost goes beyond the price. Board power is 575W, which means a serious power supply and a case with real airflow. If you mostly game at 1440p, this card is overkill and you should stop reading this section. If local AI is your main hobby and the budget is north of $2,000, we think this is the endgame card, and the gaming performance comes along as a bonus.
If you want 24GB and have a preference about how to get it
Three cards land at 24GB, and they suit three different personalities.
The GeForce RTX 3090 (used) is our pick for most serious hobbyists. It launched at $1,499 in 2020, and it remains the cheapest reliable path to 24GB with full CUDA support. The community knowledge around it is deep, so nearly every setup guide you find will have been tested on one. It also supports NVLink, so two cards can pool to 48GB, a trick nothing newer at this price can do. The trade-offs are real: gaming performance is roughly upper-midrange now, board power is 350W from five-year-old silicon, and many units lived hard lives in mining rigs or render farms. Buy from a seller you can return to.
The GeForce RTX 4090 (used) is for the buyer who wants 24GB of CUDA plus the second-fastest gaming card available. It launched at $1,599, is now discontinued, and its used prices stay close to what it cost new because AI demand never cooled. It runs 32B-class models comfortably. Board power is 450W, and the 12VHPWR power connector is this generation's known weak point, so inspect it before you pay. Favor listings with the original box and a return window, since most come without warranty. The upside is that it holds value if you resell.
The Radeon RX 7900 XTX is the only way to get 24GB in a card you can still buy new for under $1,000, with a launch MSRP of $999. Ollama and llama.cpp run well on it, so if your work is inference, the AI story is solid. The catch is ROCm instead of CUDA: any CUDA-only tool locks it out, and training workflows are a weaker fit. It draws 355W and runs warm. For a buyer who wants a warranty, does mostly inference and also games at 4K, discounted units make this the sleeper value of the group.
If your budget is closer to $500
The GeForce RTX 5060 Ti 16GB launched at $429 and is the cheapest new NVIDIA card with 16GB of GDDR7. That is enough for 13B-class models to run without drama, with full CUDA support for every tool that matters. Board power is a modest 180W, so it drops into almost any existing system without a power supply upgrade.
Its gaming performance is midrange, and you are paying for the memory rather than the frames. That is fine, because the memory is the point. Make sure you buy the 16GB version. An 8GB variant exists under nearly the same name, and for local AI it defeats the entire purpose.
How we would actually choose
Start from the model size you want to run. If 13B-class models cover your needs, the 5060 Ti 16GB is the sensible buy and leaves money for everything else. If you want 32B-class models, you need 24GB, and the choice comes down to risk tolerance. Pick a used 3090 for the lowest cost of entry into CUDA, a used 4090 if you also want top-tier gaming, or the 7900 XTX if you want a new card with a warranty and can live with ROCm. Only jump to the 5090 if 32GB matters to you and your power supply can deliver 575W cleanly. Whichever you choose, check your PSU rating and case clearance before you order, because power and space are where most local AI builds stall.