AI applications are used in various applications by performing the training task T and inference task R of the machine learning (ML) on servers in data centers. Data processing and decision-making tasks are performed on edge nodes in the edge computing (EC) model of the IoT. Since data is locally processed on a small computer named edge node close to users, the responsiveness is improved, smaller energy is consumed, and private data is protected. In this paper, we newly propose an AEC (AI for Edge Computing) model where the inference task T is performed on edge nodes while the training task R is performed on servers. Parts of the inference task T are distributed on edge nodes, and a group of the edge nodes cooperates to process data from devices. Each edge node obtains output data by performing tasks on input data from devices and other edge nodes. Differently from the EC model, the inference task T changes as the training task R is performed, e.g., the execution time and output data change for the same input data. Each edge node exchanges output data with other edge nodes in addition to delivering the output data to devices. Here, an edge node has to find a succeeding edge node which can process the output data. We propose an algorithm to select an energy-efficient succeeding edge node under the condition where the inference task changes. In the evaluation, we show a succeeding edge node consuming smaller energy in the proposed algorithm.