Breakeven analysis
| Metric | Value |
|---|
About this calculator & data sources
What it answers: Given a GPU fleet, what token throughput, cost per million tokens, and gross margin can you expect — and what utilisation do you need to break even at your chosen price?
The model: Tokens/day = GPUs × tokens/sec/GPU × utilisation × 86,400. Cost/Mtok = (GPU $/hr × GPUs × 24) ÷ (tokens/day ÷ 1e6). Revenue/GPU-hr = (tokens/sec/GPU × util × 3,600 ÷ 1e6) × $/Mtok. Margin = (revenue − cost) ÷ revenue.
Critical caveat: The "DeepSeek fleet average" throughput (~14M tokens/sec/GPU) includes massive prefix caching and long shared contexts across a 128-GPU cluster. A single request sees ~80K uncached tokens/sec/GPU. The dropdown lets you pick which reality you're modelling.
Sources: DeepInfer production figures (155T tokens/day per 128-GPU cluster, $38/GPU-hr revenue) from
doc/inference_deployment_qa.md in deepseek-ai/DeepInfer. GPU rental prices from public cloud list pricing (B200 ~$1.89, H200 ~$1.50, H100 ~$1.10, owned amortised ~$0.60). All editable via GPU presets.