Capacity planner
Your traffic in, a GPU fleet out: replicas to hold the peak, what it costs per month, and how much of it sits idle the rest of the time.
Loading GPU prices…
Couldn’t load GPU prices.
Used at average loadIdle, still billed
Estimate: output tokens/s per replica from memory bandwidth with batched serving, a fixed fleet sized for peak and billed 730 hours a month, cheapest offer per GPU in whole instances. Benchmark before you commit.