diff --git a/src/dstack/_internal/core/backends/oci/compute.py b/src/dstack/_internal/core/backends/oci/compute.py
index ab9be4908b..f5f839771b 100644
--- a/src/dstack/_internal/core/backends/oci/compute.py
+++ b/src/dstack/_internal/core/backends/oci/compute.py
@@ -35,6 +35,13 @@
 from dstack._internal.core.models.resources import Memory, Range
 from dstack._internal.core.models.runs import JobProvisioningData, Requirements
 
+# Shape families that dstack's x86-64 OCI Marketplace images can run. GPU shapes are limited to
+# NVIDIA GPUs on x86-64 hosts because the images ship CUDA drivers only. Other GPU shapes from
+# the gpuhunt catalog are deliberately left out:
+# - BM.GPU.MI300X.*, BM.GPU.MI355X.* (AMD Instinct): no ROCm image is published, and
+#   `create_instance` picks the CUDA image for any GPU offer.
+# - BM.GPU.GB200.*, BM.GPU.GB300.* (Grace superchips): arm64 hosts, no arm64 image is published.
+# See https://docs.oracle.com/en-us/iaas/Content/Compute/References/computeshapes.htm
 SUPPORTED_SHAPE_FAMILIES = [
     "VM.Standard2.",
     "BM.Standard2.",
@@ -48,6 +55,13 @@
     "BM.GPU4.",
     "VM.GPU.A10.",
     "BM.GPU.A10.",
+    "BM.GPU.A100-v2.",
+    "BM.GPU.L40S.",
+    "BM.GPU.H100.",
+    "BM.GPU.H200.",
+    "BM.GPU.B200.",
+    "BM.GPU.B300.",
+    "BM.GPU.RTXPRO.",
 ]
 CONFIGURABLE_DISK_SIZE = Range[Memory](min=Memory.parse("50GB"), max=Memory.parse("32TB"))
 
diff --git a/src/dstack/_internal/core/backends/oci/configurator.py b/src/dstack/_internal/core/backends/oci/configurator.py
index 95696330e7..8fc6aa57c3 100644
--- a/src/dstack/_internal/core/backends/oci/configurator.py
+++ b/src/dstack/_internal/core/backends/oci/configurator.py
@@ -28,7 +28,9 @@
 )
 from dstack._internal.core.models.common import validate_extra_ignore, validate_json_extra_ignore
 
-# where dstack images are published
+# Regions where dstack publishes its VM images as OCI Marketplace community listings
+# (see scripts/oci_image_tools.py). Listings are regional, so instances can only be launched
+# in regions that have the images. Extend this set only together with publishing the images.
 SUPPORTED_REGIONS = frozenset(
     [
         "eu-frankfurt-1",
diff --git a/src/tests/_internal/core/backends/oci/test_compute.py b/src/tests/_internal/core/backends/oci/test_compute.py
new file mode 100644
index 0000000000..51644c7c3b
--- /dev/null
+++ b/src/tests/_internal/core/backends/oci/test_compute.py
@@ -0,0 +1,66 @@
+import pytest
+
+from dstack._internal.core.backends.oci.compute import _supported_instances
+from dstack._internal.core.models.backends.base import BackendType
+from dstack._internal.core.models.instances import (
+    InstanceOffer,
+    InstanceType,
+    Resources,
+)
+
+
+def make_offer(shape_name: str) -> InstanceOffer:
+    return InstanceOffer(
+        backend=BackendType.OCI,
+        instance=InstanceType(
+            name=shape_name,
+            resources=Resources(cpus=8, memory_mib=64 * 1024, gpus=[], spot=False),
+        ),
+        region="us-chicago-1",
+        price=1.0,
+    )
+
+
+class TestSupportedInstances:
+    @pytest.mark.parametrize(
+        "shape_name",
+        [
+            "VM.Standard2.1",
+            "BM.Standard.E5.192",
+            "VM.GPU3.1",
+            "BM.GPU4.8",
+            "VM.GPU.A10.1",
+            "BM.GPU.A10.4",
+            "BM.GPU.A100-v2.8",
+            "BM.GPU.L40S.4",
+            "BM.GPU.H100.8",
+            "BM.GPU.H200.8",
+            "BM.GPU.B200.8",
+            "BM.GPU.B300.8",
+            "BM.GPU.RTXPRO.8",
+        ],
+    )
+    def test_supported(self, shape_name: str):
+        assert _supported_instances(make_offer(shape_name))
+
+    @pytest.mark.parametrize(
+        "shape_name",
+        [
+            # Flex shapes need OCPU/memory configuration
+            "VM.Standard.E4.Flex",
+            "VM.Optimized3.Flex",
+            # Ampere A1 Arm CPUs and Grace superchips: no arm64 image
+            "VM.Standard.A1.Flex",
+            "BM.Standard.A1.160",
+            "BM.GPU.GB200.4",
+            "BM.GPU.GB300.4",
+            # AMD Instinct: no ROCm image
+            "BM.GPU.MI300X.8",
+            "BM.GPU.MI355X.8",
+            # Deprecated families
+            "BM.GPU2.2",
+            "VM.GPU2.1",
+        ],
+    )
+    def test_unsupported(self, shape_name: str):
+        assert not _supported_instances(make_offer(shape_name))