The next-generation cloud computing systems are expected to be connected to the real world more tightly by massive amounts of sensors and actuators. Today's clouds, however, are not capable of handling massive sensor data or giving fast feedback to the real world because of the long latency and limited bandwidth of WANs. We propose a cloud architecture that allows fast response to the real world in spite of floods of sensor data. We also present a deep sensor-data aggregation mechanism for the proposed cloud architecture to compress sensor data by an order of magnitude. Our evaluation results show that the deep sensor-data aggregation at the edges of clouds is effective in reducing network traffic sent over cloud networks.