Tokens vs GPUs

Pay per token, or run the model on your own GPUs? See what each costs at your volume and how many tokens it takes to break even.

Live prices

Pay per token

–

Your own GPUs

–

Difference

–

Breakeven

–

Your workload

Model

Precision

BF16 FP8 INT4

Tokens per month

Input + output, all requests.

Share that is output

Own GPUs

Rent Buy

GPU

Use spot prices Cheaper, but can be interrupted

Average utilization

Traffic isn’t flat: you size for the peak and pay for the quiet hours too.

Prices and speed assumptions

API input ($/1M)

API output ($/1M)

Output tokens/s per instance

Estimated from memory bandwidth with batched serving. Measure yours with Load testing.

API prices are a sample of hosted list prices; set your provider’s.

Pay per token Own GPUs

Loading GPU prices…