diff --git a/packages/system/gpu-operator/examples/README.md b/packages/system/gpu-operator/examples/README.md
new file mode 100644
index 00000000..dac3703e
--- /dev/null
+++ b/packages/system/gpu-operator/examples/README.md
@@ -0,0 +1,69 @@
+# GPU operator — native pod workload on Talos (reference)
+
+The files in this directory are **not** templates. They are reference
+artifacts that document one working configuration for running GPU
+workloads directly in pods on a Talos-based Cozystack cluster, together
+with the DCGM metrics needed by the `gpu/gpu-performance` Grafana
+dashboard.
+
+The out-of-the-box `values-talos.yaml` for this package targets the
+sandbox (VFIO passthrough to KubeVirt VMs) scenario. The files here
+illustrate an alternative — running CUDA workloads in regular pods with
+the NVIDIA device plugin — and the workarounds it currently requires on
+Talos.
+
+## Files
+
+- [`values-native-talos.yaml`](./values-native-talos.yaml) — Cozystack
+  `Package` values that disable sandbox workloads, enable the device
+  plugin, point `hostPaths.driverInstallDir` at the staging location
+  used by the compat DaemonSet, and wire DCGM to the custom metrics
+  ConfigMap.
+- [`dcgm-custom-metrics.yaml`](./dcgm-custom-metrics.yaml) — `ConfigMap`
+  with a DCGM metrics CSV that adds profiling, ECC, throttling and
+  energy counters on top of the upstream defaults. Required by the
+  recording rules in `packages/system/monitoring-agents/alerts/gpu-recording.rules.yaml`
+  and by several panels in the `gpu/gpu-performance` dashboard.
+- [`nvidia-driver-compat.yaml`](./nvidia-driver-compat.yaml) — DaemonSet
+  that stages `libnvidia-ml.so.1` and `nvidia-smi` from the Talos glibc
+  tree into a path where the NVIDIA GPU Operator validator expects
+  them. See the "Why the compat DaemonSet exists" section below.
+
+## Why these are reference, not templates
+
+Shipping these as first-class templates would silently impose
+assumptions that do not hold for every user:
+
+- Whether the NVIDIA Talos system extension is installed on the nodes.
+- Whether GPUs are exposed directly to pods or passed through to VMs.
+- The exact path the installed driver ends up at (depends on the
+  extension version and Talos release).
+
+The sandbox-oriented `values-talos.yaml` remains the default. Operators
+who want native pod GPU workloads can start from this directory and
+adapt as needed.
+
+## Why the compat DaemonSet exists
+
+The NVIDIA GPU Operator validator checks for `libnvidia-ml.so.1` and
+`bin/nvidia-smi` in the path given by `hostPaths.driverInstallDir`.
+Talos installs them under `/usr/local/glibc/usr/lib/` and
+`/usr/local/bin/`, which the validator does not look at. Until upstream
+addresses [NVIDIA/gpu-operator#1687][1], the DaemonSet copies those
+files into a directory the validator does inspect and creates the
+`.driver-ctr-ready` flag file so the validator proceeds.
+
+[1]: https://github.com/NVIDIA/gpu-operator/issues/1687
+
+## How the dashboard and recording rules fit in
+
+- `dashboards/gpu/gpu-performance.json` expects `DCGM_FI_*` metrics,
+  including profiling series (`DCGM_FI_PROF_PIPE_TENSOR_ACTIVE`,
+  `DCGM_FI_PROF_GR_ENGINE_ACTIVE`) and throttling counters
+  (`DCGM_FI_DEV_POWER_VIOLATION`, `DCGM_FI_DEV_THERMAL_VIOLATION`).
+  These are only emitted when DCGM Exporter is started with the custom
+  CSV in `dcgm-custom-metrics.yaml`.
+- `packages/system/monitoring-agents/alerts/gpu-recording.rules.yaml`
+  precomputes cluster-wide and per-namespace aggregations used by the
+  overview panels of the dashboard. The rules are safe to ship on any
+  cluster — they evaluate to empty series when DCGM is not scraped.
diff --git a/packages/system/gpu-operator/examples/dcgm-custom-metrics.yaml b/packages/system/gpu-operator/examples/dcgm-custom-metrics.yaml
new file mode 100644
index 00000000..14b89b5c
--- /dev/null
+++ b/packages/system/gpu-operator/examples/dcgm-custom-metrics.yaml
@@ -0,0 +1,82 @@
+# Custom DCGM Exporter metrics CSV. Referenced from
+# examples/values-native-talos.yaml via dcgmExporter.config.name.
+#
+# Extends the upstream default set with profiling counters, ECC, page
+# retirement, row remap, energy and throttling violations — everything
+# the gpu/gpu-performance dashboard and the GPU recording rules in
+# monitoring-agents expect.
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: dcgm-custom-metrics
+  namespace: cozy-gpu-operator
+data:
+  dcgm-metrics.csv: |
+    # Format
+    # If line starts with a '#' it is considered a comment
+    # DCGM FIELD, Prometheus metric type, help message
+
+    # Clocks
+    DCGM_FI_DEV_SM_CLOCK,  gauge, SM clock frequency (in MHz).
+    DCGM_FI_DEV_MEM_CLOCK, gauge, Memory clock frequency (in MHz).
+
+    # Temperature
+    DCGM_FI_DEV_MEMORY_TEMP, gauge, Memory temperature (in C).
+    DCGM_FI_DEV_GPU_TEMP,    gauge, GPU temperature (in C).
+
+    # Power
+    DCGM_FI_DEV_POWER_USAGE,              gauge, Power draw (in W).
+    DCGM_FI_DEV_TOTAL_ENERGY_CONSUMPTION, counter, Total energy consumption since boot (in mJ).
+
+    # PCIE
+    DCGM_FI_DEV_PCIE_REPLAY_COUNTER, counter, Total number of PCIe retries.
+
+    # Utilization (the sample period varies depending on the product)
+    DCGM_FI_DEV_GPU_UTIL,      gauge, GPU utilization (in %).
+    DCGM_FI_DEV_MEM_COPY_UTIL, gauge, Memory utilization (in %).
+    DCGM_FI_DEV_ENC_UTIL,      gauge, Encoder utilization (in %).
+    DCGM_FI_DEV_DEC_UTIL,      gauge, Decoder utilization (in %).
+
+    # Errors and violations
+    DCGM_FI_DEV_XID_ERRORS, gauge, Value of the last XID error encountered.
+
+    # Memory usage
+    DCGM_FI_DEV_FB_FREE,     gauge, Framebuffer memory free (in MiB).
+    DCGM_FI_DEV_FB_USED,     gauge, Framebuffer memory used (in MiB).
+    DCGM_FI_DEV_FB_RESERVED, gauge, Framebuffer memory reserved (in MiB).
+
+    # ECC (supported on datacenter-class GPUs such as A10)
+    DCGM_FI_DEV_ECC_SBE_VOL_TOTAL, counter, Total number of single-bit volatile ECC errors.
+    DCGM_FI_DEV_ECC_DBE_VOL_TOTAL, counter, Total number of double-bit volatile ECC errors.
+    DCGM_FI_DEV_ECC_SBE_AGG_TOTAL, counter, Total number of single-bit persistent ECC errors.
+    DCGM_FI_DEV_ECC_DBE_AGG_TOTAL, counter, Total number of double-bit persistent ECC errors.
+
+    # Retired pages
+    DCGM_FI_DEV_RETIRED_SBE,     counter, Total number of retired pages due to single-bit errors.
+    DCGM_FI_DEV_RETIRED_DBE,     counter, Total number of retired pages due to double-bit errors.
+    DCGM_FI_DEV_RETIRED_PENDING, counter, Total number of pages pending retirement.
+
+    # Row remapping (not applicable to GDDR6 but left for datacenter GPUs)
+    DCGM_FI_DEV_UNCORRECTABLE_REMAPPED_ROWS, counter, Number of remapped rows for uncorrectable errors.
+    DCGM_FI_DEV_CORRECTABLE_REMAPPED_ROWS,   counter, Number of remapped rows for correctable errors.
+    DCGM_FI_DEV_ROW_REMAP_FAILURE,           gauge,   Whether remapping of rows has failed.
+
+    # Throttle / violation counters (crucial for SLA and tenant monitoring)
+    DCGM_FI_DEV_POWER_VIOLATION,       counter, Throttling duration due to power constraints (in us).
+    DCGM_FI_DEV_THERMAL_VIOLATION,     counter, Throttling duration due to thermal constraints (in us).
+    DCGM_FI_DEV_SYNC_BOOST_VIOLATION,  counter, Throttling duration due to sync-boost constraints (in us).
+    DCGM_FI_DEV_BOARD_LIMIT_VIOLATION, counter, Throttling duration due to board limit constraints (in us).
+    DCGM_FI_DEV_LOW_UTIL_VIOLATION,    counter, Throttling duration due to low utilization (in us).
+    DCGM_FI_DEV_RELIABILITY_VIOLATION, counter, Throttling duration due to reliability constraints (in us).
+
+    # NVLink — enable only for GPUs that actually have NVLink (A10 has none).
+    # DCGM_FI_DEV_NVLINK_BANDWIDTH_TOTAL, counter, Total number of NVLink bandwidth counters for all lanes.
+
+    # DCP (profiling) metrics — supported from Ampere onwards.
+    DCGM_FI_PROF_GR_ENGINE_ACTIVE,   gauge, Ratio of time the graphics engine is active.
+    DCGM_FI_PROF_SM_ACTIVE,          gauge, The ratio of cycles an SM has at least 1 warp assigned.
+    DCGM_FI_PROF_SM_OCCUPANCY,       gauge, The ratio of number of warps resident on an SM.
+    DCGM_FI_PROF_PIPE_TENSOR_ACTIVE, gauge, Ratio of cycles the tensor (HMMA) pipe is active.
+    DCGM_FI_PROF_DRAM_ACTIVE,        gauge, Ratio of cycles the device memory interface is active sending or receiving data.
+    DCGM_FI_PROF_PCIE_TX_BYTES,      counter, The number of bytes of active PCIe tx (transmit) data including both header and payload.
+    DCGM_FI_PROF_PCIE_RX_BYTES,      counter, The number of bytes of active PCIe rx (read) data including both header and payload.
diff --git a/packages/system/gpu-operator/examples/nvidia-driver-compat.yaml b/packages/system/gpu-operator/examples/nvidia-driver-compat.yaml
new file mode 100644
index 00000000..e4718e0a
--- /dev/null
+++ b/packages/system/gpu-operator/examples/nvidia-driver-compat.yaml
@@ -0,0 +1,72 @@
+# Workaround for https://github.com/NVIDIA/gpu-operator/issues/1687
+#
+# On Talos, the NVIDIA system extension installs driver files under
+# /usr/local/glibc/usr/lib/ and /usr/local/bin/. The GPU Operator
+# validator only looks for libnvidia-ml.so.1 and bin/nvidia-smi under
+# the path configured as hostPaths.driverInstallDir. This DaemonSet
+# stages the required files into /var/nvidia-driver on each node and
+# creates the .driver-ctr-ready flag so the validator proceeds.
+#
+# Paired with examples/values-native-talos.yaml, which sets
+# hostPaths.driverInstallDir to /var/nvidia-driver.
+apiVersion: apps/v1
+kind: DaemonSet
+metadata:
+  name: nvidia-driver-compat
+  namespace: cozy-gpu-operator
+  labels:
+    app: nvidia-driver-compat
+spec:
+  selector:
+    matchLabels:
+      app: nvidia-driver-compat
+  template:
+    metadata:
+      labels:
+        app: nvidia-driver-compat
+    spec:
+      hostPID: true
+      priorityClassName: system-node-critical
+      tolerations:
+        - operator: Exists
+      initContainers:
+        - name: create-driver-tree
+          image: busybox:1.37
+          command:
+            - sh
+            - -c
+            - |
+              set -e
+              GLIBC_LIB="/host/usr/local/glibc/usr/lib"
+              SRC_BIN="/host/usr/local/bin"
+              DST="/host/var/nvidia-driver"
+
+              mkdir -p "$DST/bin"
+
+              if [ -f "$GLIBC_LIB/libnvidia-ml.so.1" ]; then
+                cp -f "$GLIBC_LIB/libnvidia-ml.so.1" "$DST/libnvidia-ml.so.1"
+                echo "Copied libnvidia-ml.so.1"
+              fi
+
+              if [ -f "$SRC_BIN/nvidia-smi" ]; then
+                cp -f "$SRC_BIN/nvidia-smi" "$DST/bin/nvidia-smi"
+                chmod +x "$DST/bin/nvidia-smi"
+                echo "Copied nvidia-smi"
+              fi
+
+              mkdir -p /host/run/nvidia/validations
+              touch /host/run/nvidia/validations/.driver-ctr-ready
+              echo "Created driver-ctr-ready flag"
+              echo "Done"
+          securityContext:
+            privileged: true
+          volumeMounts:
+            - name: host-root
+              mountPath: /host
+      containers:
+        - name: pause
+          image: registry.k8s.io/pause:3.10
+      volumes:
+        - name: host-root
+          hostPath:
+            path: /
diff --git a/packages/system/gpu-operator/examples/values-native-talos.yaml b/packages/system/gpu-operator/examples/values-native-talos.yaml
new file mode 100644
index 00000000..86e436c4
--- /dev/null
+++ b/packages/system/gpu-operator/examples/values-native-talos.yaml
@@ -0,0 +1,45 @@
+# Cozystack Package values for running GPU workloads natively in pods
+# on Talos. This is the counterpart to values-talos.yaml, which targets
+# the sandbox (VFIO passthrough) scenario.
+#
+# Prerequisites:
+#   - NVIDIA Talos system extension installed on GPU nodes.
+#   - examples/nvidia-driver-compat.yaml deployed to stage driver files
+#     where the gpu-operator validator looks for them.
+#   - examples/dcgm-custom-metrics.yaml applied so DCGM Exporter exports
+#     the full set of metrics used by the dashboard and recording rules.
+apiVersion: cozystack.io/v1alpha1
+kind: Package
+metadata:
+  name: cozystack.gpu-operator
+spec:
+  variant: default
+  components:
+    gpu-operator:
+      values:
+        gpu-operator:
+          # The compat DaemonSet stages libnvidia-ml.so.1 and nvidia-smi
+          # under /var/nvidia-driver. Point the validator at that path.
+          hostPaths:
+            driverInstallDir: "/var/nvidia-driver"
+          # Disable the sandbox path — workloads run in pods, not VMs.
+          sandboxWorkloads:
+            enabled: false
+          vfioManager:
+            enabled: false
+          # The Talos extension provides the driver and runtime hooks,
+          # so the operator's own toolkit and driver components must be
+          # switched off to avoid conflicting installations.
+          toolkit:
+            enabled: false
+          devicePlugin:
+            enabled: true
+          # Export full set of DCGM metrics using the ConfigMap in
+          # examples/dcgm-custom-metrics.yaml.
+          dcgmExporter:
+            serviceMonitor:
+              enabled: true
+              interval: "15s"
+              honorLabels: true
+            config:
+              name: dcgm-custom-metrics