diff --git a/source/slang/hlsl.meta.slang b/source/slang/hlsl.meta.slang
index f465fdd323..eeb990d3ef 100644
--- a/source/slang/hlsl.meta.slang
+++ b/source/slang/hlsl.meta.slang
@@ -26032,6 +26032,7 @@ struct DescriptorHandle<T:IOpaqueDescriptor> : IComparable
     /// Constructor for uint64_t handles
     [ForceInline]
     [require(spvBindlessTextureNV)]
+    [require(cuda)]
     __intrinsic_op($(kIROp_CastUInt64ToDescriptorHandle))
     __init(uint64_t handleValue);
 
@@ -26041,6 +26042,7 @@ struct DescriptorHandle<T:IOpaqueDescriptor> : IComparable
         __target_switch
         {
         case spvBindlessTextureNV:
+        case cuda:
             return (uint64_t)this == (uint64_t)other;
         default:
             return all(__vectorEql((uint2)this, (uint2)other));
@@ -26053,6 +26055,7 @@ struct DescriptorHandle<T:IOpaqueDescriptor> : IComparable
         __target_switch
         {
         case spvBindlessTextureNV:
+        case cuda:
             return (uint64_t)this < (uint64_t)other;
         default:
             let vthis = ((uint2)this);
@@ -26067,6 +26070,7 @@ struct DescriptorHandle<T:IOpaqueDescriptor> : IComparable
         __target_switch
         {
         case spvBindlessTextureNV:
+        case cuda:
             return (uint64_t)this <= (uint64_t)other;
         default:
             let vthis = ((uint2)this);
@@ -26089,6 +26093,7 @@ extension uint64_t
 {
     __intrinsic_op($(kIROp_CastDescriptorHandleToUInt64))
     [require(spvBindlessTextureNV)]
+    [require(cuda)]
     __init<T:IOpaqueDescriptor>(DescriptorHandle<T> bindless);
 }
 
diff --git a/tests/language-feature/descriptor-handle/desc-handle-cuda-uint64.slang b/tests/language-feature/descriptor-handle/desc-handle-cuda-uint64.slang
new file mode 100644
index 0000000000..86321cc4a2
--- /dev/null
+++ b/tests/language-feature/descriptor-handle/desc-handle-cuda-uint64.slang
@@ -0,0 +1,58 @@
+//TEST:SIMPLE(filecheck=CUDA): -target cuda -entry computeMain -stage compute
+
+// Test that DescriptorHandle can be constructed from uint64_t in CUDA.
+// In CUDA, DescriptorHandle<T> becomes transparent and maps directly to the underlying
+// resource type (CUtexObject/CUsurfObject which are ulonglong). This means:
+// 1. uint64_t -> DescriptorHandle conversion is a no-op cast
+// 2. DescriptorHandle -> uint64_t conversion is a no-op cast  
+// 3. Comparison methods (equals, lessThan, etc.) cannot be called directly because
+//    the handle type is erased to the resource type at compile time.
+
+// CUDA-DAG: ulonglong textureHandle
+// CUDA-DAG: tex2DLod
+// CUDA-DAG: void computeMain
+
+uniform uint64_t textureHandle;
+uniform uint64_t textureHandle2;
+
+// Cast uint64_t to DescriptorHandle and use it
+DescriptorHandle<Sampler2D> getTextureHandle(uint64_t h)
+{
+    return DescriptorHandle<Sampler2D>(h);
+}
+
+// Cast DescriptorHandle to uint64_t
+uint64_t getHandleValue(DescriptorHandle<Sampler2D> h)
+{
+    return (uint64_t)h;
+}
+
+// Test that we can compare handles using uint64_t comparisons
+// (since direct method calls aren't available in CUDA)
+bool compareHandles(uint64_t h1, uint64_t h2)
+{
+    DescriptorHandle<Sampler2D> handle1 = DescriptorHandle<Sampler2D>(h1);
+    DescriptorHandle<Sampler2D> handle2 = DescriptorHandle<Sampler2D>(h2);
+    // Convert back to uint64_t for comparison
+    return (uint64_t)handle1 == (uint64_t)handle2;
+}
+
+uniform RWStructuredBuffer<float4> output;
+
+[numthreads(1,1,1)]
+void computeMain()
+{
+    // Test construction from uint64_t - creates a texture handle
+    DescriptorHandle<Sampler2D> tex = DescriptorHandle<Sampler2D>(textureHandle);
+    
+    // Test conversion to uint64_t and back
+    uint64_t handleValue = (uint64_t)tex;
+    DescriptorHandle<Sampler2D> tex2 = DescriptorHandle<Sampler2D>(handleValue);
+    
+    // Test comparison via uint64_t conversion
+    bool same = compareHandles(textureHandle, textureHandle2);
+    
+    // Sample the texture using the handle
+    output[0] = tex->SampleLevel(float2(0.0), 0.0);
+    output[1] = float4(same ? 1.0 : 0.0, 0.0, 0.0, 0.0);
+}