diff --git a/local/AGENTS.md b/local/AGENTS.md
index abe76db26a..07c3289413 100644
--- a/local/AGENTS.md
+++ b/local/AGENTS.md
@@ -579,6 +579,26 @@ Practical consequences:
   second CPU platform.
 - Build only the amd64 ISO. There are no cross-arch ISOs.
 
+#### GPU driver + LLVM GPU-backend strategy
+
+GPU support is a separate axis from the amd64 CPU platform. Mesa driver lists
+(`-Dgallium-drivers`, `-Dvulkan-drivers`) and the LLVM GPU backends
+(`LLVM_TARGETS_TO_BUILD` in `recipes/dev/llvm21`) must stay in sync:
+
+| Vendor | Mesa driver(s) | LLVM GPU backend needed? |
+|---|---|---|
+| AMD | radeonsi (GL), radv (Vulkan) | **AMDGPU** (required) |
+| Intel | iris (GL), anvil (Vulkan) | No — NIR + Intel's own compiler |
+| virgl | virgl (GL) | No — virtualized (virtio-gpu) |
+| software | softpipe/llvmpipe (GL), lavapipe (Vulkan) | X86 (CPU JIT), not a GPU backend |
+| NVIDIA *(future)* | nouveau (GL), NVK (Vulkan) | **NVPTX** |
+
+`llvm21` builds `X86;AMDGPU;NVPTX`: X86 = amd64 CPU + software JIT, AMDGPU =
+AMD now, NVPTX = **included now** so adding NVIDIA later needs no (very
+expensive) LLVM rebuild. Do NOT drop Intel/virgl when touching GPU config —
+they need no LLVM backend but are first-class targets. When NVIDIA lands, add
+`nouveau`/`nvk` to the Mesa driver lists (LLVM is already ready).
+
 ## RELEASE MODEL (FORK — NOT OVERLAY)
 
 Red Bear OS is a FULL FORK, not an overlay. The repository diverges from upstream Redox through local/sources/<comp>/ forks (see "LOCAL FORK SUPREMACY POLICY" below) rather than through per-package overlay patches.
diff --git a/recipes/dev/llvm21/recipe.toml b/recipes/dev/llvm21/recipe.toml
index 7a5492fb3b..e487917ebd 100644
--- a/recipes/dev/llvm21/recipe.toml
+++ b/recipes/dev/llvm21/recipe.toml
@@ -69,19 +69,24 @@ set(ENV{CXXFLAGS} "")
 NATIVE_EOF
 
 # amd64-only (see local/AGENTS.md): the CPU target is X86 (= amd64/x86_64).
-# AMDGPU is the GPU codegen backend Mesa's radeonsi/radv (AMD GPU) drivers
-# require via `dependency('llvm', modules: [... amdgpu ...])` — without it the
-# host llvm-config the mesa build reads reports amdgpu(missing) and mesa fails
-# to configure. AMDGPU is a GPU backend, not a second CPU platform, so it does
-# not violate amd64-only.
+# GPU codegen backends are a separate axis from the CPU platform and are driven
+# by which Mesa drivers we ship:
+#   - AMDGPU  -> radeonsi (GL) + radv (Vulkan), AMD hardware GPUs. Required now:
+#     Mesa's `dependency('llvm', modules: [... amdgpu ...])` fails without it.
+#   - NVPTX   -> future NVIDIA support (nouveau/NVK). Included NOW so enabling
+#     NVIDIA later needs no (very expensive) LLVM rebuild.
+#   - Intel (iris/anvil) and virgl need NO LLVM GPU backend — Intel is NIR-based
+#     with its own shader compiler and virgl is virtualized; they are enabled via
+#     Mesa's gallium/vulkan driver lists, not an LLVM target.
+# None of these add a second CPU platform, so amd64-only still holds.
 case "${ARCH}" in
-    x86 | x86_64) LLVM_TARGETS_TO_BUILD="X86;AMDGPU";;
-    aarch64) LLVM_TARGETS_TO_BUILD="AArch64;AMDGPU";;
-    riscv64gc) LLVM_TARGETS_TO_BUILD="RISCV;AMDGPU";;
+    x86 | x86_64) LLVM_TARGETS_TO_BUILD="X86;AMDGPU;NVPTX";;
+    aarch64) LLVM_TARGETS_TO_BUILD="AArch64;AMDGPU;NVPTX";;
+    riscv64gc) LLVM_TARGETS_TO_BUILD="RISCV;AMDGPU;NVPTX";;
 esac
 
 if [ "${COOKBOOK_HOST_SYSROOT}" = "/usr" ]; then
-    LLVM_TARGETS_TO_BUILD="X86;AMDGPU"
+    LLVM_TARGETS_TO_BUILD="X86;AMDGPU;NVPTX"
 fi
 
 # This just build the LLVM library and tools just enough for Rust, to build the rest of LLVM see