GPU 시세 · VRAM · 온디바이스 AI

Calculated with the 2026-09-28 rate of 1,356.04 KRW/USD (open.er-api), KEPCO’s residential low-voltage tariff (120.0 / 214.6 / 307.3 KRW/kWh tiered), and Claude’s official price list (Haiku 4.5 $1/$5, Sonnet 5 $2/$10 — a hike was planned for Sept 1 but $2/$10 is still the standard price). For local I assume a used 4090 (3.56M KRW) depreciated over 24 months + 70W system idle + 350W decode increment, aggregate 1,100 tok/s (gigagpu measured 3.4 tok/J).

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

Monthly usage (output/input) Electricity for local 4090 rigs Local total/month Haiku 4.5 API Sonnet 5 API
Light users 2M / 10M 5,481 + depreciation 148,333 = ₩153,815 27,121 KRW 54,242 won
Intermediate user 8M / 40M 5,545 + depreciation 148,333 = 153,878 KRW ₩108,483 216,966 KRW
Heavy agents 20M / 100M 5,672 + depreciation 148,333 = 154,005 won 271,208 won ₩542,416

Interpretation

  • Electricity always wins — even heavy users stay under 20,000 won/month. The real cost is card depreciation.
  • Even including depreciation, at 8M output tok/month local matches Haiku; at 20M local is 2.7x cheaper. Break-even is roughly 7~8M output tok/month.
  • If you want Sonnet 5-class quality, it’s a different story — to run that class locally you need Flash-Next 125B (M5 Max 128GB) class, where the machine cost lands before the electricity bill.
  • bestllmfor’s aggregation points the same way: for agent workloads the break-even vs Haiku is 800,000–2,000,000 tokens of daily output.

The card’s own token unit costUsed 3090 unit-price table, model selection isHubat .

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