Estimate the emissions of third-party AI API calls from model and request characteristics when the hardware is not accessible