How large could the compute demand for GenAI inference become? UBS discusses