Training is a one-time bill. Inference is a subscription that compounds.
Training a model is expensive exactly once. Serving it costs money on every single request, forever — and usage grows. Explore the math below and find the crossover point yourself.
The core idea
A training run is a fixed capital cost: you pay it, and it never grows. Inference is an operating cost proportional to tokens served. If your product succeeds, tokens served rise month after month, so cumulative inference spend curves upward while the training bar stays flat. Sooner or later the curve crosses the bar — usually much sooner than people expect.
The simulator
Set your scenario. The chart plots cumulative spend over ten years and marks where inference overtakes training.
Worked examples
Three real-world shapes of this trade-off. Each preset animates the chart.
Check your intuition
FAQ
Why is training called a one-time cost?
Once a model checkpoint is produced, the compute spent training it does not recur. Retraining a new version is a new project, not an ongoing cost of serving the existing one.
What makes inference compound?
Inference cost scales with tokens served. If usage grows a fixed percentage each month, cumulative spend grows geometrically — the classic compounding curve.
Can inference never catch up?
Yes: with tiny traffic, low prices, or a huge training bill, the crossover can sit beyond ten years. The simulator shows that state explicitly.