71.3% of chat queries could run locally, per intelligence per watt paper