diff --git a/ext/ArrayInterfaceGPUArraysCoreExt.jl b/ext/ArrayInterfaceGPUArraysCoreExt.jl
index 79bbf606..09990f05 100644
--- a/ext/ArrayInterfaceGPUArraysCoreExt.jl
+++ b/ext/ArrayInterfaceGPUArraysCoreExt.jl
@@ -2,7 +2,7 @@ module ArrayInterfaceGPUArraysCoreExt
 
 using Adapt
 using ArrayInterface
-using LinearAlgebra: lu
+import LinearAlgebra
 import GPUArraysCore
 
 ArrayInterface.fast_scalar_indexing(::Type{<:GPUArraysCore.AbstractGPUArray}) = false
@@ -19,8 +19,16 @@ function ArrayInterface.restructure(x::GPUArraysCore.AbstractGPUArray, y)
     reshape(Adapt.adapt(ArrayInterface.parameterless_type(x), y), Base.size(x)...)
 end
 
+# Build the `LU` directly rather than calling `lu` on an adapted array, matching the CPU
+# methods in ArrayInterface.jl. Going through `lu` only works for backends that define their
+# own, so it breaks on JLArrays (which downstream packages use to test GPU paths without a
+# GPU) and on Metal, see #501 and #467.
 function ArrayInterface.lu_instance(A::GPUArraysCore.AbstractGPUMatrix{T}) where {T}
-    lu(Adapt.adapt(ArrayInterface.parameterless_type(A), ones(T, 0, 0)))
+    noUnitT = typeof(zero(T))
+    luT = LinearAlgebra.lutype(noUnitT)
+    ipiv = similar(A, LinearAlgebra.BlasInt, 0)
+    info = zero(LinearAlgebra.BlasInt)
+    return LinearAlgebra.LU{luT}(similar(A, 0, 0), ipiv, info)
 end
 
 # Doesn't do much, but makes a gigantic change to the dependency chain.