diff --git a/sky/authentication.py b/sky/authentication.py
index 442c1abb372..d2d65835657 100644
--- a/sky/authentication.py
+++ b/sky/authentication.py
@@ -19,6 +19,7 @@
 is an exception, due to the limitation of the cloud provider. See the
 comments in setup_lambda_authentication)
 """
+import base64
 import copy
 import functools
 import os
@@ -261,6 +262,36 @@ def setup_gcp_authentication(config: Dict[str, Any]) -> Dict[str, Any]:
     return configure_ssh_info(config)
 
 
+# In Azure, cloud-init script must be encoded in base64. See
+# https://learn.microsoft.com/en-us/azure/virtual-machines/custom-data
+# for more information. Here we decode it and replace the ssh user
+# and public key content, then encode it back.
+def setup_azure_authentication(config: Dict[str, Any]) -> Dict[str, Any]:
+    _, public_key_path = get_or_generate_keys()
+    with open(public_key_path, 'r') as f:
+        public_key = f.read().strip()
+    for node_type in config['available_node_types']:
+        node_config = config['available_node_types'][node_type]['node_config']
+        cloud_init = (
+            node_config['azure_arm_parameters']['cloudInitSetupCommands'])
+        cloud_init = base64.b64decode(cloud_init).decode('utf-8')
+        cloud_init = cloud_init.replace('skypilot:ssh_user',
+                                        config['auth']['ssh_user'])
+        cloud_init = cloud_init.replace('skypilot:ssh_public_key_content',
+                                        public_key)
+        cloud_init = base64.b64encode(
+            cloud_init.encode('utf-8')).decode('utf-8')
+        node_config['azure_arm_parameters']['cloudInitSetupCommands'] = (
+            cloud_init)
+    config_str = common_utils.dump_yaml_str(config)
+    config_str = config_str.replace('skypilot:ssh_user',
+                                    config['auth']['ssh_user'])
+    config_str = config_str.replace('skypilot:ssh_public_key_content',
+                                    public_key)
+    config = yaml.safe_load(config_str)
+    return config
+
+
 def setup_lambda_authentication(config: Dict[str, Any]) -> Dict[str, Any]:
 
     get_or_generate_keys()
diff --git a/sky/backends/backend_utils.py b/sky/backends/backend_utils.py
index d1e3381a800..4f8d5f37c27 100644
--- a/sky/backends/backend_utils.py
+++ b/sky/backends/backend_utils.py
@@ -1156,10 +1156,12 @@ def _add_auth_to_cluster_config(cloud: clouds.Cloud, cluster_config_file: str):
     """
     config = common_utils.read_yaml(cluster_config_file)
     # Check the availability of the cloud type.
-    if isinstance(cloud, (clouds.AWS, clouds.Azure, clouds.OCI, clouds.SCP)):
+    if isinstance(cloud, (clouds.AWS, clouds.OCI, clouds.SCP)):
         config = auth.configure_ssh_info(config)
     elif isinstance(cloud, clouds.GCP):
         config = auth.setup_gcp_authentication(config)
+    elif isinstance(cloud, clouds.Azure):
+        config = auth.setup_azure_authentication(config)
     elif isinstance(cloud, clouds.Lambda):
         config = auth.setup_lambda_authentication(config)
     elif isinstance(cloud, clouds.Kubernetes):
diff --git a/sky/clouds/azure.py b/sky/clouds/azure.py
index 41f6862d9df..21731bdb0b3 100644
--- a/sky/clouds/azure.py
+++ b/sky/clouds/azure.py
@@ -134,37 +134,47 @@ def get_default_instance_type(
                                                          clouds='azure')
 
     def _get_image_config(self, gen_version, instance_type):
+        # TODO(tian): images for Azure is not well organized. We should refactor
+        # it to images.csv like AWS.
         # az vm image list \
         #  --publisher microsoft-dsvm --all --output table
-        # nvidia-driver: 495.29.05, cuda: 11.5
-
-        # The latest image 2022.09.14/2022.08.11/22.06.10/22.05.11/
-        # 22.04.27/22.04.05 has even older nvidia driver 470.57.02,
-        # cuda: 11.4
+        # nvidia-driver: 535.54.03, cuda: 12.2
+        # see: https://github.com/Azure/azhpc-images/releases/tag/ubuntu-hpc-20230803
+        # All A100 instances is of gen2, so it will always use
+        # the latest ubuntu-hpc:2204 image.
         image_config = {
             'image_publisher': 'microsoft-dsvm',
-            'image_offer': 'ubuntu-2004',
-            'image_sku': '2004-gen2',
-            'image_version': '21.11.04'
+            'image_offer': 'ubuntu-hpc',
+            'image_sku': '2204',
+            'image_version': '22.04.2023080201'
         }
 
-        # ubuntu-2004 v21.10.21 and v21.11.04 do not work on K80
-        # due to an NVIDIA driver issue.
+        # ubuntu-2004 v21.08.30, K80 requires image with old NVIDIA driver version
         acc = self.get_accelerators_from_instance_type(instance_type)
         if acc is not None:
             acc_name = list(acc.keys())[0]
             if acc_name == 'K80':
-                image_config['image_version'] = '21.08.30'
-
-        # ubuntu-2004 does not work on A100
-        if instance_type in [
-                'Standard_ND96asr_v4', 'Standard_ND96amsr_A100_v4'
-        ]:
-            image_config['image_offer'] = 'ubuntu-hpc'
-            image_config['image_sku'] = '2004'
-            image_config['image_version'] = '20.04.2021120101'
+                image_config = {
+                    'image_publisher': 'microsoft-dsvm',
+                    'image_offer': 'ubuntu-2004',
+                    'image_sku': '2004-gen2',
+                    'image_version': '21.08.30'
+                }
+
+        # ubuntu-2004 v21.11.04, the previous image we used in the past for
+        # V1 HyperV instance before we change default image to ubuntu-hpc.
+        # In Azure, all instances with K80 (Standard_NC series), some
+        # instances with M60 (Standard_NV series) and some cpu instances
+        # (Basic_A, Standard_D, ...) are V1 instance. For these instances,
+        # we use the previous image.
         if gen_version == 'V1':
-            image_config['image_sku'] = '2004'
+            image_config = {
+                'image_publisher': 'microsoft-dsvm',
+                'image_offer': 'ubuntu-2004',
+                'image_sku': '2004',
+                'image_version': '21.11.04'
+            }
+
         return image_config
 
     @classmethod
@@ -251,13 +261,15 @@ def make_deploy_resources_variables(
         # This script will modify /etc/ssh/sshd_config and add a bash script
         # into .bashrc. The bash script will restart sshd if it has not been
         # restarted, identified by a file /tmp/__restarted is existing.
+        # Also, add default user to docker group.
         # pylint: disable=line-too-long
         cloud_init_setup_commands = base64.b64encode(
             textwrap.dedent("""\
             #cloud-config
             runcmd:
               - sed -i 's/#Banner none/Banner none/' /etc/ssh/sshd_config
-              - echo '\\nif [ ! -f "/tmp/__restarted" ]; then\\n  sudo systemctl restart ssh\\n  sleep 2\\n  touch /tmp/__restarted\\nfi' >> /home/azureuser/.bashrc
+              - echo '\\nif [ ! -f "/tmp/__restarted" ]; then\\n  sudo systemctl restart ssh\\n  sleep 2\\n  touch /tmp/__restarted\\nfi' >> /home/skypilot:ssh_user/.bashrc
+              - usermod -aG docker skypilot:ssh_user
             write_files:
               - path: /etc/apt/apt.conf.d/20auto-upgrades
                 content: |
diff --git a/tests/test_smoke.py b/tests/test_smoke.py
index aecd5371de0..cab52ab2f75 100644
--- a/tests/test_smoke.py
+++ b/tests/test_smoke.py
@@ -727,6 +727,7 @@ def test_gcp_stale_job_manual_restart():
 @pytest.mark.no_scp  # SCP does not support num_nodes > 1 yet
 def test_env_check(generic_cloud: str):
     name = _get_cluster_name()
+    total_timeout_minutes = 25 if generic_cloud == 'azure' else 15
     test = Test(
         'env_check',
         [
@@ -734,6 +735,7 @@ def test_env_check(generic_cloud: str):
             f'sky logs {name} 1 --status',  # Ensure the job succeeded.
         ],
         f'sky down -y {name}',
+        timeout=total_timeout_minutes * 60,
     )
     run_one_test(test)
 
@@ -1134,6 +1136,7 @@ def test_scp_job_queue():
 @pytest.mark.no_kubernetes  # Kubernetes not support num_nodes > 1 yet
 def test_job_queue_multinode(generic_cloud: str):
     name = _get_cluster_name()
+    total_timeout_minutes = 30 if generic_cloud == 'azure' else 15
     test = Test(
         'job_queue_multinode',
         [
@@ -1164,6 +1167,7 @@ def test_job_queue_multinode(generic_cloud: str):
             f'sky logs {name} 7 --status',
         ],
         f'sky down -y {name}',
+        timeout=total_timeout_minutes * 60,
     )
     run_one_test(test)
 
@@ -1428,6 +1432,7 @@ def test_tpu_vm_pod():
 @pytest.mark.no_scp  # SCP does not support num_nodes > 1 yet
 def test_multi_hostname(generic_cloud: str):
     name = _get_cluster_name()
+    total_timeout_minutes = 25 if generic_cloud == 'azure' else 15
     test = Test(
         'multi_hostname',
         [
@@ -1438,6 +1443,7 @@ def test_multi_hostname(generic_cloud: str):
             f'sky logs {name} 2 --status',  # Ensure the job succeeded.
         ],
         f'sky down -y {name}',
+        timeout=total_timeout_minutes * 60,
     )
     run_one_test(test)
 
@@ -1573,6 +1579,12 @@ def test_azure_start_stop():
 @pytest.mark.no_kubernetes  # Kubernetes does not autostop yet
 def test_autostop(generic_cloud: str):
     name = _get_cluster_name()
+    # Azure takes ~ 7m15s (435s) to autostop a VM, so here we use 600 to ensure
+    # the VM is stopped.
+    autostop_timeout = 600 if generic_cloud == 'azure' else 250
+    # Launching and starting Azure clusters can take a long time too. e.g., restart
+    # a stopped Azure cluster can take 7m. So we set the total timeout to 70m.
+    total_timeout_minutes = 70 if generic_cloud == 'azure' else 20
     test = Test(
         'autostop',
         [
@@ -1587,7 +1599,7 @@ def test_autostop(generic_cloud: str):
             f's=$(sky status {name} --refresh); echo "$s"; echo; echo; echo "$s"  | grep {name} | grep UP',
 
             # Ensure the cluster is STOPPED.
-            'sleep 250',
+            f'sleep {autostop_timeout}',
             f's=$(sky status {name} --refresh); echo "$s"; echo; echo; echo "$s"  | grep {name} | grep STOPPED',
 
             # Ensure the cluster is UP and the autostop setting is reset ('-').
@@ -1605,7 +1617,7 @@ def test_autostop(generic_cloud: str):
             f'sky autostop -y {name} -i 1',  # Should restart the timer.
             'sleep 45',
             f's=$(sky status {name} --refresh); echo "$s"; echo; echo; echo "$s" | grep {name} | grep UP',
-            'sleep 250',
+            f'sleep {autostop_timeout}',
             f's=$(sky status {name} --refresh); echo "$s"; echo; echo; echo "$s"  | grep {name} | grep STOPPED',
 
             # Test restarting the idleness timer via exec:
@@ -1616,11 +1628,11 @@ def test_autostop(generic_cloud: str):
             f'sky exec {name} echo hi',  # Should restart the timer.
             'sleep 45',
             f's=$(sky status {name} --refresh); echo "$s"; echo; echo; echo "$s"  | grep {name} | grep UP',
-            'sleep 250',
+            f'sleep {autostop_timeout}',
             f's=$(sky status {name} --refresh); echo "$s"; echo; echo; echo "$s"  | grep {name} | grep STOPPED',
         ],
         f'sky down -y {name}',
-        timeout=20 * 60,
+        timeout=total_timeout_minutes * 60,
     )
     run_one_test(test)
 
@@ -1629,6 +1641,10 @@ def test_autostop(generic_cloud: str):
 @pytest.mark.no_scp  # SCP does not support num_nodes > 1 yet. Run test_scp_autodown instead.
 def test_autodown(generic_cloud: str):
     name = _get_cluster_name()
+    # Azure takes ~ 13m30s (810s) to autodown a VM, so here we use 900 to ensure
+    # the VM is terminated.
+    autodown_timeout = 900 if generic_cloud == 'azure' else 240
+    total_timeout_minutes = 90 if generic_cloud == 'azure' else 20
     test = Test(
         'autodown',
         [
@@ -1640,23 +1656,23 @@ def test_autodown(generic_cloud: str):
             'sleep 45',
             f's=$(sky status {name} --refresh); echo "$s"; echo; echo; echo "$s"  | grep {name} | grep UP',
             # Ensure the cluster is terminated.
-            'sleep 200',
+            f'sleep {autodown_timeout}',
             f's=$(SKYPILOT_DEBUG=0 sky status {name} --refresh) && echo "$s" && {{ echo "$s" | grep {name} | grep "Autodowned cluster\|terminated on the cloud"; }} || {{ echo "$s" | grep {name} && exit 1 || exit 0; }}',
             f'sky launch -y -d -c {name} --cloud {generic_cloud} --num-nodes 2 --down tests/test_yamls/minimal.yaml',
             f'sky status | grep {name} | grep UP',  # Ensure the cluster is UP.
             f'sky exec {name} --cloud {generic_cloud} tests/test_yamls/minimal.yaml',
             f'sky status | grep {name} | grep "1m (down)"',
-            'sleep 240',
+            f'sleep {autodown_timeout}',
             # Ensure the cluster is terminated.
             f's=$(SKYPILOT_DEBUG=0 sky status {name} --refresh) && echo "$s" && {{ echo "$s" | grep {name} | grep "Autodowned cluster\|terminated on the cloud"; }} || {{ echo "$s" | grep {name} && exit 1 || exit 0; }}',
             f'sky launch -y -d -c {name} --cloud {generic_cloud} --num-nodes 2 --down tests/test_yamls/minimal.yaml',
             f'sky autostop -y {name} --cancel',
-            'sleep 240',
+            f'sleep {autodown_timeout}',
             # Ensure the cluster is still UP.
             f's=$(SKYPILOT_DEBUG=0 sky status {name} --refresh) && echo "$s" && echo "$s" | grep {name} | grep UP',
         ],
         f'sky down -y {name}',
-        timeout=20 * 60,
+        timeout=total_timeout_minutes * 60,
     )
     run_one_test(test)
 
@@ -1797,6 +1813,7 @@ def test_cancel_ibm():
 
 
 # ---------- Testing use-spot option ----------
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances
@@ -1818,6 +1835,7 @@ def test_use_spot(generic_cloud: str):
 
 
 # ---------- Testing managed spot ----------
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances
@@ -1851,6 +1869,7 @@ def test_spot(generic_cloud: str):
     run_one_test(test)
 
 
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances
@@ -1890,6 +1909,7 @@ def test_spot_pipeline(generic_cloud: str):
     run_one_test(test)
 
 
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances
@@ -1913,6 +1933,7 @@ def test_spot_failed_setup(generic_cloud: str):
     run_one_test(test)
 
 
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances
@@ -2105,6 +2126,7 @@ def test_spot_pipeline_recovery_gcp():
     run_one_test(test)
 
 
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances
@@ -2326,6 +2348,7 @@ def test_spot_cancellation_gcp():
 
 
 # ---------- Testing storage for managed spot ----------
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances
@@ -2381,6 +2404,7 @@ def test_spot_tpu():
 
 
 # ---------- Testing env for spot ----------
+@pytest.mark.no_azure  # Azure does not support spot instances
 @pytest.mark.no_lambda_cloud  # Lambda Cloud does not support spot instances
 @pytest.mark.no_ibm  # IBM Cloud does not support spot instances
 @pytest.mark.no_scp  # SCP does not support spot instances