GPU cost calculator for LLM inference and video generation
Which GPU serves your traffic for the least money? Enter your load. Every capacity number below was measured on a GPU we rented, under the same latency target, with the report one click away. No spec-sheet estimates.
These are our nodes. Yours will be different.
Hosts lower power limits, clocks sag under load, and your model and prompts aren't ours. We run the same measurement on your node and send a one-page report: health, capacity within your latency target, cost per million tokens at your price. Our first five teams get their first evaluation free, in exchange for an anonymized case study.
How the calculator works
LLM serving. Each GPU's capacity is the highest request rate at which p99 time to first token stayed ≤ 2 s and p99 time per output token ≤ 50 ms, on a chat-like workload (median 1,024 input and 256 output tokens). GPUs needed = peak traffic ÷ capacity, rounded up to whole GPUs. Monthly cost = GPUs × price per hour × 730. Your cost per million output tokens = monthly cost ÷ the output tokens your average load produces in a month. That's why low average load makes every GPU expensive: you pay for the peak and use it a fraction of the day.
Video generation. Each GPU's time is the measured generation time of a 15-second 1280×720 clip with a Wan2.2 14B image-to-video FP8 model (8 distilled steps) in ComfyUI. GPUs needed = clips per day × minutes per clip ÷ the GPU minutes you can keep busy per day.
Prices. "What we paid" is the price on the day of the test (mostly RunPod Secure Cloud). "Lowest list price" is today's lowest on-demand price in our price tracker for the same GPU and form factor. Some of those offers only come as 8-GPU machines, and GPUs the tracker doesn't list keep the test price. You can type any price into the table.
Limits. One node per GPU, one model per size, default serving settings. Quick-mode results run 15–20% higher than full mode. They're marked. Not included: storage, egress, engineering time, redundancy, or input-token pricing when comparing with APIs. Method: how we measure. The formula: cost per token.