Self-host vs API break-even calculator
Compare monthly GPU rental (cheapest public on-demand fit from our price table) with monthly inference API cost (sourced $/1M token rates). Throughput is your estimate. Prices checked Nov 9, 2026.
Break-even calculator
Estimate GPU rental vs inference API
- GPU / month
- Pick a model
- Cheapest on-demand fit
- API / month
- Pick a model
- At your token mix
- Break-even
- -
- Tokens/month where bills match
- GPU capacity
- -
- At your tokens/sec assumption
| Provider | API model | Input $/1M | Output $/1M | For your tokens | Confidence | Checked | Source |
|---|
Choose a model to compare GPU rental and API cost.
Static comparison
Cheapest GPU fit vs cheapest API, by model
Defaults: FP8 memory fit, 730 hours, 100M tokens/month at 50% input / 50% output, and the lowest blended API rate. Throughput is not assumed in this table.
| Model | Cheapest GPU fit | GPU / month | Cheapest API | API / month (100M) | Break-even tokens | Source |
|---|---|---|---|---|---|---|
| DeepSeek V3 | 5x MI300X Crusoe, $3.45/GPU-hr | $12,592.50 | DeepInfra $0.32 / $0.89 per 1M in/out | $60.50 | 20814.0M / mo | SourceChecked 2026-11-05 |
| Gemma 2 27B | 1x L40S Massed Compute, $0.97/GPU-hr | $708.10 | DeepInfra $0.27 / $0.27 per 1M in/out | $27.00 | 2622.6M / mo | SourceChecked 2026-11-05 |
| gpt-oss-120b | 2x A100 Thunder Compute, $1.09/GPU-hr | $1,591.40 | OpenRouter $0.02 / $0.09 per 1M in/out | $5.65 | 28166.4M / mo | SourceChecked 2026-11-05 |
| gpt-oss-20b | 1x L40S Massed Compute, $0.97/GPU-hr | $708.10 | OpenRouter $0.03 / $0.13 per 1M in/out | $8.00 | 8851.2M / mo | SourceChecked 2026-11-05 |
| Llama 3.1 405B | 7x A100 Thunder Compute, $1.09/GPU-hr | $5,569.90 | DeepInfra $0.80 / $0.80 per 1M in/out | $80.00 | 6962.4M / mo | SourceChecked 2026-11-05 |
| Llama 3.1 70B | 2x L40S Massed Compute, $0.97/GPU-hr | $1,416.20 | DeepInfra $0.40 / $0.40 per 1M in/out | $40.00 | 3540.5M / mo | SourceChecked 2026-11-05 |
| Llama 3.1 8B | 1x L40S Massed Compute, $0.97/GPU-hr | $708.10 | DeepInfra $0.02 / $0.04 per 1M in/out | $3.00 | 23603.3M / mo | SourceChecked 2026-11-05 |
| Llama 3.3 70B | 2x L40S Massed Compute, $0.97/GPU-hr | $1,416.20 | DeepInfra $0.10 / $0.32 per 1M in/out | $21.00 | 6743.8M / mo | SourceChecked 2026-11-05 |
| Llama 4 Scout | 2x A100 Thunder Compute, $1.09/GPU-hr | $1,591.40 | DeepInfra $0.10 / $0.30 per 1M in/out | $20.00 | 7957.0M / mo | SourceChecked 2026-11-05 |
| Mistral 7B | 1x L40S Massed Compute, $0.97/GPU-hr | $708.10 | DeepInfra $0.03 / $0.05 per 1M in/out | $4.10 | 17270.7M / mo | SourceChecked 2026-11-05 |
| Mixtral 8x7B | 1x A100 Thunder Compute, $1.09/GPU-hr | $795.70 | DeepInfra $0.54 / $0.54 per 1M in/out | $54.00 | 1473.5M / mo | SourceChecked 2026-11-05 |
| Qwen2.5 72B | 2x L40S Massed Compute, $0.97/GPU-hr | $1,416.20 | DeepInfra $0.36 / $0.40 per 1M in/out | $38.00 | 3726.8M / mo | SourceChecked 2026-11-05 |
| Qwen2.5 7B | 1x L40S Massed Compute, $0.97/GPU-hr | $708.10 | DeepInfra $0.04 / $0.10 per 1M in/out | $7.00 | 10115.7M / mo | SourceChecked 2026-11-05 |
| Qwen3 30B-A3B | 1x L40S Massed Compute, $0.97/GPU-hr | $708.10 | DeepInfra $0.12 / $0.50 per 1M in/out | $31.00 | 2284.2M / mo | SourceChecked 2026-11-05 |
| Qwen3 32B | 1x L40S Massed Compute, $0.97/GPU-hr | $708.10 | DeepInfra $0.08 / $0.28 per 1M in/out | $18.00 | 3933.9M / mo | SourceChecked 2026-11-05 |
FAQ
Self-host vs API questions
What does self-host mean here?
Self-host means renting public on-demand GPUs and running the model yourself. API means paying a hosted inference provider per million input and output tokens. Both sides use numbers from our sourced tables.
Where do tokens per second come from?
They do not come from a benchmark. tokens_per_second_per_gpu is your estimate, not a benchmark. Edit it. Capacity = tps x GPUs x hours x 3600.
How is break-even calculated?
Monthly GPU = $/GPU-hr x GPUs x hours (default 730). Monthly API = (tokens x input share / 1e6) x input $/1M + (tokens x output share / 1e6) x output $/1M. Break-even tokens/month = monthly GPU / blended $/token. Blended $/token = (input share x input $/1M + output share x output $/1M) / 1e6.
Which models appear?
Only models that appear in both data/models.csv and data/api_prices.csv with a high or medium confidence API rate, and that fit on a public on-demand GPU node we track.
Can you help me choose?
Send a GPU request with the model, tokens/month, latency target and budget. We reply with the public prices that match and the checks to run before you pay. Free, and we do not sell compute.
Need this capacity?
Tell us the GPU, count, term, region and budget. We reply with the public prices that match and the checks to run before you pay. Free, and we do not sell compute.