Interactive Explainer

Training is a one-time bill. Inference compounds forever.

Training a model is expensive, but you pay it once. Every request users make afterward costs money — and that spend grows with usage. Explore the math behind why inference is set to dwarf the training market.

The Core Idea

Training cost is a fixed capital expense: one tall bar, paid up front. Inference is an operating expense that recurs on every single request. Even a modest per-request cost, multiplied by millions of daily requests and monthly growth, becomes an exponential curve that eventually towers over the training bar.

Training: paid onceInference: compounds with usagegrows every month

The Simulator

Set your assumptions. The chart plots cumulative spend over ten years and marks the month inference overtakes training.

Cumulative training (flat)Cumulative inference

Worked Examples

Three real-world scale scenarios. Each animates the simulator to new assumptions.

Indie fine-tune

$50k training run, cheap tokens, a few thousand daily requests, steady 5% growth.

Mid-size API product

$20M training, $2 per 1M tokens, 2M daily requests, 12% monthly growth.

Frontier-scale model

$3B training, premium pricing, hundreds of millions of daily requests, 6% growth.

Check Your Intuition

Four quick questions on fixed versus recurring costs.

FAQ

Why is training a one-time cost?

Training happens before deployment: you pay for compute once to produce the model weights. Retraining exists, but each run is a discrete capital expense, not a per-user cost.

Why does inference compound?

Every request consumes compute proportional to tokens processed. As users and usage grow month over month, spend multiplies — a classic exponential curve against a flat line.

Can inference never catch up?

Yes — if usage is near zero or per-token prices collapse faster than usage grows, cumulative inference can stay below a very large training bill for a decade. The simulator shows this state explicitly.

Enjoy this tool? Build your own with Super