Inference cost crossover
See when moving the open-model slice of your agent inference off a managed API and onto self-hosted GPUs — orchestrated by Kinesis — starts saving money. Adjust the assumptions below; the chart and verdict update live.
API (managed)Kinesis (high utilization)DIY reserved (idle penalty)
Log-log chart comparing managed API, Kinesis, and DIY inference costs. Crossover at 935.6M tokens/month.
At 2B tokens/month, self-hosting on Kinesis saves $1,706/month compared to the managed API. The crossover is at 935.6M tokens/month. Effective rates: API $2.00/1M vs Kinesis $0.40/1M.
2B
$2.0
$2.5/hr
2,500
70%
30%
$1,500
Defaults are illustrative. The x-axis represents only the movable open-model inference slice — frontier-reasoning calls stay on the managed API regardless. Privacy and sovereignty are separate forcing functions not modelled here.