Amortize hardware over 3 years, add electricity, and compare per-1M-token cost against major cloud APIs. Pick the view that matches your usage. Methodology.
| GPU Setup | cost.col.vram | cost.col.tps | Annual Cost | Tokens/yr | $/1M tok | GPT-4o | GPT-4o-mini | Claude Sonnet 4 | Claude Haiku 4 | DeepSeek V3 | Qwen3-8B (Together) | Break-even |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Arc B580 12GB | 12GB | 35.4 tok/s | $166/yr | 372.1M/yr | $0.45 | $6.25 | $0.38 | $9.00 | $2.40 | $0.69 | $0.10 | 1628d |
| RTX 4060 Ti 16GB | 16GB | 22.4 tok/s | $238/yr | 235.5M/yr | $1.01 | $6.25 | $0.38 | $9.00 | $2.40 | $0.69 | $0.10 | 3689d |
| RTX 4090 24GB | 24GB | 78.2 tok/s | $730/yr | 822.0M/yr | $0.89 | $6.25 | $0.38 | $9.00 | $2.40 | $0.69 | $0.10 | 3241d |
| RTX 3090 24GB | 24GB | 72.6 tok/s | $653/yr | 763.2M/yr | $0.86 | $6.25 | $0.38 | $9.00 | $2.40 | $0.69 | $0.10 | 3123d |
| Mac mini M4 32GB | 32GB | 9.3 tok/s | $401/yr | 97.8M/yr | $4.10 | $6.25 | $0.38 | $9.00 | $2.40 | $0.69 | $0.10 | 14972d |