AI Infrastructure
Capacity you provision for is capacity somebody has to predict.
The operation
Inference load is not smooth. It is diurnal, bursty, and it saturates against a ceiling you already paid for. The expensive failure is rarely the spike itself. It is provisioning for a spike that never arrives, every day for a quarter, or missing the one that did.
What changes
Forge reads the request load the fleet already produces and projects it forward, so the capacity decision is made against a number instead of a nerve.
Tested here
Tested on a production Azure Functions trace, a 300-function cohort at five-minute grain. Floor skill of +12.0 to +21.5% over persistence, and breach-risk Brier of 0.031 to 0.035 against a 0.049 base rate. This was the industry we predicted would fit before we had the data to check.