The amount of time devoted to AI workloads on HPC systems has been steadily increasing. These workloads, however, have very different execution patterns compared to traditional HPC applications, and they put stress on the subsystems that traditional applications do not exercise as much. In fact, the two systems at the top of TOP500 list, Summit and Sierra, include architectural elements, such as AI-oriented GPUs and high-speed interconnects between the GPUs within node and across nodes, to support AI applications. But many other challenges remain, such as stringent storage system performance requirements and AI frameworks scalability issues. This break-out session is designed to review the efforts at the JLESC partnering institutions to enable efficient execution of AI workloads on the existing HPC systems and to identify research opportunities to address some of the challenges with this new use of HPC platforms. The session is co-organized by Volodymyr Kindratenko of NCSA and Michela Taufer of UTK and incudes presentations from RIKEN, JSC, UTK, and NCSA.