Depends on the model, the size of the query and response, the hardware it’s running on (tpus are more efficient than GPUs), and many other factors. Also depends on the aircon.