Self-Host vs API Break-Even Calculator
Is self-hosting a cheaper LLM worth it at your token volume? Calculate the monthly cost of rented GPU vs API, VRAM requirements, and the exact break-even point.
| GPU | VRAM | $/hr on-demand | $/hr spot | Monthly (OD) | Est. tok/s |
|---|---|---|---|---|---|
| RTX 3090 ↗affiliate link | 24 GB | $0.44 | $0.29 | $321 | ~180 t/s |
| RTX 4090 ↗affiliate link | 24 GB | $0.74 | $0.49 | $540 | ~180 t/s |
| L40S 48GB ↗affiliate link | 48 GB | $1.25 | $0.89 | $913 | ~180 t/s |
| A100 40GB SXM ↗affiliate link | 40 GB | $1.89 | $1.24 | $1380 | ~180 t/s |
| A100 80GB SXM ↗affiliate link | 80 GB | $2.49 | $1.79 | $1818 | ~180 t/s |
| H100 80GB SXM ↗affiliate link | 80 GB | $3.29 | $2.49 | $2402 | ~180 t/s |
Compare providers on the model comparison table or route cheaper tasks to a less expensive model: Model Routing Calculator.
GPU prices verified 2026-08-19 · LLM prices verified 2026-08-20 · Methodology
How to Use
Enter your workload
Set monthly input and output token volumes, and whether you need 24/7, business hours, or batch-only availability.
Choose the API you'd replace
Select which cloud API model you're currently using or comparing against. The API monthly cost is computed from live pricing.
Choose your self-host setup
Pick the open-weight model and quantization level. Choose rented GPU cloud or owned hardware. The VRAM needed is computed automatically.
See the winner
The calculator shows which option is cheaper at your volume, the monthly savings, and the token volume at which self-hosting breaks even.
When self-hosting beats the API
Self-hosting a quantized open-weight model becomes cheaper than cloud APIs at a specific token volume. The math: API cost = (tokens_in × P_in + tokens_out × P_out) / 1M. Self-host cost is roughly fixed per month (GPU hours × rate). Where the two lines cross is the break-even point.
Key insight: at 24/7 uptime, rented GPU compute is a fixed monthly bill. API costs scale linearly with tokens. Low-volume workloads almost always favor the API. High-volume, continuous workloads favor self-hosting — especially with quantized models on cheaper GPUs.
How to Cut This Cost
Use quantized models (Q4_K_M) — they fit on fewer GPUs and run 2–4× faster than FP16, dramatically lowering the token volume needed to break even.
Use spot/interruptible GPU instances (RunPod Spot, Vast.ai) rather than on-demand. Typically 40–70% cheaper for batch inference workloads.
Run at high GPU utilization (>80%). Idle GPU time is pure cost. Batch requests and schedule heavy jobs during off-peak hours.
Share a GPU node across multiple models using frameworks like vLLM with model sharding or llama.cpp server multi-model support.
FAQ
Have more questions? Contact us
Last verified: 2026-08-20 · methodology · data sources