Break-even index / RTX 3090 (used) / GPT-5.4 / Connecticut
Will a Used RTX 3090 rig beat GPT-5.4 in Connecticut?
Your rig pays for itself in 10 months — after that you bank roughly $82.99/mo versus the cloud. Clear win.
How to read it: OWN — the rig pays for itself inside 18 months. HYBRID — break-even lands between 18 and 36 months; real, but slow. RENT — it never pays off before the hardware is due for replacement. Every formula and threshold is public on the methodology page.
The numbers
| Build cost (Used RTX 3090 rig) | $1,200 ($1,100–$1,300) |
|---|---|
| Hardware amortized (3 yr) | $33.33/mo |
| Electricity in Connecticut ($0.311/kWh) | $14.93/mo (48 kWh) |
| Local total cost of ownership | $48.26/mo |
| GPT-5.4 at the same usage | $131/mo |
| Break-even | about 10 months |
Assumes the "coding assistant" preset: ~15M tokens/month, 3 h/day active, 5 h/day idle. GPT-5.4 priced at $2.5/M in, $15/M out (verified 2026-07-05). Rig ceiling: Up to ~32B models at Q4 — 70B is tight. Jargon: VRAM is the GPU memory that decides which models fit; Q4 is 4-bit quantization — a smaller, faster copy of a model with a slight quality cost.
See your verdict — then tweak every number →