🔭 The Local-AI Cost Observatory

What does it actually cost to run AI at home? Real sold prices from the used market, joined with community inference benchmarks. The metric that matters: tokens per second, per $100 spent.

2,189sold-price observations
13cards tracked
4benchmark results
2026-07-15last updated (UTC)

Value Rankings โ€” 7โ€“8B class models (click headers to sort)

CardMedian sold priceVRAM GB Gen speed (tok/s)tok/s per $100 โ–พW per tok/sVRAM GB per $100
NVIDIA RTX 3080 10GB ๐Ÿ†$335 n=122, 30d10121.3536.222.642.99
NVIDIA RTX 3090$1,199 n=182, 30d24158.1613.192.212.0
NVIDIA RTX 3090 Ti$1,292 n=212, 30d24โ€”โ€”โ€”1.86
NVIDIA RTX 3060 12GB$262 n=88, 30d12โ€”โ€”โ€”4.57
NVIDIA RTX 4060 Ti 16GB$450 n=109, 30d16โ€”โ€”โ€”3.56
NVIDIA RTX 4090$2,000 n=22, 30d24โ€”โ€”โ€”1.2
NVIDIA Tesla P40$285 n=151, 30d24โ€”โ€”โ€”8.42
NVIDIA Tesla P100 16GB$90 n=207, 30d16โ€”โ€”โ€”17.78
NVIDIA RTX 2080 Ti$260 n=235, 30d11โ€”โ€”โ€”4.23
AMD RX 7900 XTX$800 n=262, 30d24โ€”โ€”โ€”3.0
NVIDIA RTX A4000$792 n=318, 30d16โ€”โ€”โ€”2.02
NVIDIA Tesla V100 16GB$349 n=199, 30d16โ€”โ€”โ€”4.58
AMD Instinct MI50 16GB$219 n=82, 30d16โ€”โ€”โ€”7.31

Methodology (the honest part)

Prices are medians of real completed/sold used-market listings (currently eBay sold data), sanity-filtered, deduplicated by listing ID. n is the sample size; the window shows whether the median is from the last 30 days or all recorded history.

Performance is the best community llama-bench text-generation result (tg128) per card from the official llama.cpp CUDA and ROCm scoreboards โ€” all measured on the same model (Llama 2 7B Q4_0), so cards are directly comparable. Implausible submissions are auto-flagged and excluded.

Value = tok/s รท median price ร— 100. Efficiency = card TDP รท tok/s (lower is better; power cost matters if it runs all day).

Caveats we won't hide: VRAM ceilings matter more than speed if you want larger models โ€” a fast 10GB card can't run what a slow 24GB card can. Datacenter cards (P100, P40, MI50, V100) need cooling and mounting work. Prices move; we re-harvest twice monthly.

Have a card we're missing, or a llama-bench run to contribute?
Send it via the GitHub repo โ€” raw llama-bench output is all we need.
☠ Sister instrument: The Cloud-Death Registry โ€” which smart devices survive when the company loses interest.
If this instrument saved you money or research time, you can fuel it here โ€” it runs on one tiny container and stubbornness.