diff --git a/docs/src/api/array.md b/docs/src/api/array.md
index 7d238d677..b6eb8848e 100644
--- a/docs/src/api/array.md
+++ b/docs/src/api/array.md
@@ -15,7 +15,7 @@ MtlVecOrMat
 
 ## Storage modes
 
-The Metal API has various storage modes that dictate how a resource can be accessed. `MtlArray`s are `Metal.PrivateStorage` by default, but they can also be `Metal.SharedStorage` or `Metal.ManagedStorage`. For more information on storage modes, see the official [Metal documentation](https://developer.apple.com/documentation/metal/resource_fundamentals/setting_resource_storage_modes).
+The Metal API has various storage modes that dictate how a resource can be accessed. `MtlArray`s are `Metal.SharedStorage` by default — zero-copy on unified-memory (Apple Silicon) GPUs — but they can also be `Metal.PrivateStorage` or `Metal.ManagedStorage`. On a discrete GPU, set the `default_storage` preference to `"private"` for best performance. For more information on storage modes, see the official [Metal documentation](https://developer.apple.com/documentation/metal/resource_fundamentals/setting_resource_storage_modes).
 
 ```@docs
 Metal.PrivateStorage
diff --git a/src/array.jl b/src/array.jl
index da6aa8a3c..8ad610f82 100644
--- a/src/array.jl
+++ b/src/array.jl
@@ -39,7 +39,7 @@ end
 
 `N`-dimensional Metal array with storage mode `S` and elements of type `T`.
 
-`S` can be `Metal.PrivateStorage` (default), `Metal.SharedStorage`.
+`S` can be `Metal.SharedStorage` (default) or `Metal.PrivateStorage`.
 
 See the Array Programming section of the Metal.jl docs for more details.
 """
@@ -190,14 +190,16 @@ See also `VecOrMat`(@ref) for examples.
 """
 const MtlVecOrMat{T,S} = Union{MtlVector{T,S},MtlMatrix{T,S}}
 
-# default to private memory
-const DefaultStorageMode = let str = @load_preference("default_storage", "private")
-    if str == "private"
-        PrivateStorage
-    elseif str == "shared"
+# Default to shared storage: zero-copy on unified-memory (Apple Silicon) GPUs and
+# valid on discrete GPUs too (`__init__` steers discrete-GPU users to `private`).
+# Override with the `default_storage` preference.
+const DefaultStorageMode = let str = @load_preference("default_storage", "shared")
+    if str == "shared"
         SharedStorage
+    elseif str == "private"
+        PrivateStorage
     else
-        error("unknown default storage mode: $default_storage")
+        error("unknown default storage mode: $str (expected \"shared\" or \"private\")")
     end
 end
 
@@ -475,9 +477,9 @@ function Adapt.adapt_storage(to::MtlArrayAdaptor{S}, xs::AbstractArray{T,N}) whe
 end
 
 """
-    mtl(A; storage=Metal.PrivateStorage)
+    mtl(A; storage=Metal.SharedStorage)
 
-`storage` can be `Metal.PrivateStorage` (default) or `Metal.SharedStorage`.
+`storage` can be `Metal.SharedStorage` (default) or `Metal.PrivateStorage`.
 
 Opinionated GPU array adaptor, which may alter the element type `T` of arrays:
 * For `T<:AbstractFloat`, it makes a `MtlArray{Float32}` for performance and compatibility
diff --git a/src/initialization.jl b/src/initialization.jl
index 9d6bb9680..28aa4b48e 100644
--- a/src/initialization.jl
+++ b/src/initialization.jl
@@ -79,6 +79,19 @@ function __init__()
         return
     end
 
+    # The default `SharedStorage` is optimal on unified-memory (Apple Silicon) GPUs
+    # and works everywhere, but on a discrete GPU `PrivateStorage` is usually faster.
+    # Steer those (rare) users to the preference, unless they already chose a default.
+    if DefaultStorageMode == SharedStorage &&
+            load_preference(Metal, "default_storage", nothing) === nothing &&
+            functional() && !device().hasUnifiedMemory
+        @warn """Metal.jl defaults to `SharedStorage`, which is optimal on unified-memory
+                 (Apple Silicon) GPUs. This GPU does not have unified memory, where
+                 `PrivateStorage` is usually faster. To switch the default, run
+                     using Preferences; set_preferences!(Metal, "default_storage" => "private")
+                 and restart Julia.""" maxlog = 1
+    end
+
     # ensure that operations executed by the REPL back-end finish before returning,
     # because displaying values happens on a different task
     if isdefined(Base, :active_repl_backend) && !isnothing(Base.active_repl_backend)
diff --git a/test/array.jl b/test/array.jl
index b473f0c40..08e72f0ff 100644
--- a/test/array.jl
+++ b/test/array.jl
@@ -72,6 +72,12 @@ end
     @test Adapt.adapt(MtlMatrix{ComplexF32, Metal.SharedStorage}, [1 2;3 4]) isa MtlArray{ComplexF32, 2, Metal.SharedStorage}
     @test Adapt.adapt(MtlArray{Float16}, Float64[1]) isa MtlArray{Float16}
 
+    # MtlArrays default to shared storage; explicit storage modes still resolve.
+    @test Metal.DefaultStorageMode === Metal.SharedStorage
+    @test Metal.is_shared(MtlArray{Int}(undef, 4))
+    @test Metal.is_shared(mtl([1.0f0, 2.0f0, 3.0f0]))
+    @test Metal.is_private(MtlArray{Int, 1, Metal.PrivateStorage}(undef, 4))
+
     # Test a few explicitly unsupported types
     @test_throws "MtlArray only supports element types that are stored inline" MtlArray(BigInt[1])
     @test_throws "Metal does not support Float64 values" MtlArray(Float64[1])