The Cheapest 24GB GPU for AI in 2026
$9.96/GB
Best value 24GB card: the Tesla P40 at $239
Live prices as of July 2026 · See all P40 listings
24GB of VRAM is the magic number for local AI. It runs 30B-class models at Q4, fits Mixtral 8x7B, and gives a 70B model enough room when you pair two cards. The question almost everyone asks: what's the cheapest way to get 24GB?
The short answer: the Tesla M40 is the cheapest 24GB card you can buy at around $125, but it's slow. The Tesla P40 at $239 is the one we actually recommend. Here's the full ranking with live prices.
Every 24GB GPU Ranked by Price
Sorted cheapest first. $/GB is the metric that matters when you're buying VRAM to run models, not to game.
| GPU | Used Price | $/GB | Bandwidth | The Catch |
|---|---|---|---|---|
| Tesla K80 | $60 | $2.50/GB | 480 GB/s | Two 12GB GPUs, not real 24GB. Avoid. |
| Tesla M40 | $125 | $5.21/GB | 288 GB/s | Cheapest real 24GB, but slowest by far. |
| Tesla P40 | $239 | $9.96/GB | 347 GB/s | Best overall value. No display out, needs cooling. |
| RX 7900 XTX | $620 | $25.83/GB | 960 GB/s | Fast + display out, but ROCm setup is fiddlier than CUDA. |
| RTX 3090 | $1,550 | $64.58/GB | 936 GB/s | Fastest 24GB consumer card, but the priciest here. |
The K80 and M40 win on raw $/GB, but both are traps in their own way. The K80 is two 12GB GPUs glued together, so no single model larger than 12GB will load on one half. The M40 is a genuine 24GB card, but at 288 GB/s its token generation is painfully slow. The P40 is the sweet spot: real 24GB, 20% more bandwidth than the M40, full CUDA 12 support, and a huge community.
Why $/GB Is the Number That Matters
For local inference, VRAM determines the largest model you can load, and a larger model is a smarter model. A $40 K80 looks unbeatable on paper, but if it can't actually run a 13B model on a single GPU, the low price is meaningless. That's why we rank by usable $/GB, not sticker price.
It's also why a used Tesla P40 at $239 beats a new RTX 4060 Ti 8GB for AI: the 4060 Ti is faster per token, but it physically cannot fit the models the P40 runs. Use the calculator to see what each card can load for your budget.
The Three Real Choices
- Absolute cheapest, speed doesn't matter: Tesla M40 at $125. Fine for batch jobs and overnight runs where you don't sit watching tokens appear.
- Best value overall: Tesla P40 at $239. The default recommendation for budget local AI. Solve cooling and you're set.
- You want display output and speed: RTX 3090 at $1,550, or the RX 7900 XTX at $620 if you're comfortable with ROCm. Both are real desktop GPUs you can also game on.
Skip the Tesla K80
It shows up first in any "cheapest 24GB" search, but the K80 is a dual-GPU card: 2× 12GB that the OS sees as two separate devices. A 13B model that needs more than 12GB won't load on one GPU, and splitting across both is slow and finicky. For modern LLMs it's the worst pick here despite the lowest price. See the full K80 review.
What 24GB Can Run
Any 24GB card on this list (except a single K80 GPU) can handle:
- Qwen 2.5 32B (Q4): fits at ~18GB with room for context. Excellent quality.
- Mixtral 8x7B (Q3-Q4): the MoE model fits with light quantization.
- Llama 3 14B (Q6-Q8): comfortable, high quality.
- Llama 3 8B (FP16): full precision in ~16GB.
- Deepseek Coder 33B (Q4): tight but workable for coding.
For 70B models you need ~40GB, which means two of these cards. See our 70B under $500 guide. It's built around dual P40s.
Verdict
Cheapest real 24GB: M40. Best 24GB: P40.
If you only care about the lowest possible price for 24GB of VRAM, the Tesla M40 at $125 wins, just accept that it's slow. For almost everyone else, the Tesla P40 at $239 ($9.96/GB) is the right call: it's only a little more, runs the same models meaningfully faster, and has far better software support.
Want display output or real speed? Step up to a used RTX 3090 or RX 7900 XTX. You'll pay more per GB, but you get a card you can also game on.
Compare Live 24GB Prices
GPUDojo ranks every GPU by $/GB VRAM with live used prices from eBay, Amazon, and Newegg.
See the Full RankingNew to used GPUs? Read our eBay buying guide first.