‘Token unit price’ = card price divided by sustained aggregate generation speed. It’s the primary metric for how many tokens the same money extracts (speed measurements: gigagpu RTX 4090 8B benchmark cross-comparison; prices: Danawa actual listings / this site’s briefings).

🌐 · English · 中文 · Español · 한국어 · · English Hub

Card	Price (2026-09)	8B aggregate tok/s	Token unit price (KRW per 1,000 tokens)
RTX 3090 24GB (used)	2,490,000 KRW (Danawa used, median of 7 shops)	950 (AWQ, measured at b32)	₩2.62
RTX 4090 24GB (used)	3,560,000 won (this site’s briefing market price)	1,100 (FP8, measured at b32)	3.24 KRW
RTX 5090 32GB (new)	8,699,990 KRW (Danawa median)	1,700 (measured at b32)	5.12 won
RTX 5060 Ti 16GB	801,200 KRW is for the 8GB model — 16GB prices aren’t tracked	720 (measured at b32)	1.11 KRW (8GB basis)

The 3090’s real value is 24GB × scalability

• 48GB with two cards — the cheapest route that fits a 70B Q4_K_M (about 40GB). At a total of 4.98M KRW it’s cheaper than one 5090 (8.7M KRW), and the 3090×2 leads in aggregate (dual 3090 70B 17 tok/s, community standard).
• Power efficiency is 3.3 tok/J, nearly on par with the 5090 (3.4) — no loss on electricity.
• Market: Danawa used listings at 2.49 million KRW (7 stores); Daangn Market personal sales show 750,000 KRW, but that price has duds mixed in. The $325 Chinese wholesale ask is based on MOQ of 10+ units.
• Risk: mining and heat history — budget 100,000–200,000 won for fan bearings and thermal re-paste.

Electricity won’t cost much — detailed calculation atElectricity bill vs API.

Written 2026-09-28 · file sizes measured from live Hugging Face listings · speeds only from sourced public measurements and my own · no local runs ·View the full hub · Quantization ladder, 104 builds · Generation-speed chart, 47 models