diff --git a/.buildkite/test_areas/misc.yaml b/.buildkite/test_areas/misc.yaml
index 64f5d8047f..de5b235e20 100644
--- a/.buildkite/test_areas/misc.yaml
+++ b/.buildkite/test_areas/misc.yaml
@@ -219,6 +219,7 @@ steps:
     - pytest -v -s -m 'cpu_test' v1/core
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/test_serial_utils.py
+    - pytest -v -s v1/worker/test_cpu_worker_kv_budget.py
     - pytest -v -s v1/test_kv_cache_spec_registry.py
     - pytest -v -s v1/cudagraph/test_cudagraph_manager.py
     - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
diff --git a/tests/v1/worker/test_cpu_worker_kv_budget.py b/tests/v1/worker/test_cpu_worker_kv_budget.py
new file mode 100644
index 0000000000..e0493d7278
--- /dev/null
+++ b/tests/v1/worker/test_cpu_worker_kv_budget.py
@@ -0,0 +1,70 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""The CPU worker's automatic KV cache sizing must report the right cause.
+
+A worker whose own footprint exceeds its reservation and a node that is short
+of free memory need opposite remedies, so they get separate errors.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+
+from vllm.platforms import current_platform
+from vllm.utils.cpu_resource_utils import MemoryNodeInfo
+
+# Importing the CPU worker patches torch and the GPU buffer utilities for the
+# whole process (vllm/v1/worker/cpu/shm.py), so the import is deferred into
+# the tests and they run only on the CPU platform.
+pytestmark = pytest.mark.skipif(
+    not current_platform.is_cpu(), reason="CPU worker test; needs the CPU platform"
+)
+
+GiB = 1 << 30
+
+
+def _worker(monkeypatch, *, reserved, rss, total=8 * GiB, available=4 * GiB):
+    import vllm.v1.worker.cpu_worker as cpu_worker_mod
+    from vllm.v1.worker.cpu_worker import CPUWorker
+
+    worker = CPUWorker.__new__(CPUWorker)
+    worker.requested_cpu_memory = reserved
+    worker.cache_config = SimpleNamespace(kv_cache_memory_bytes=None)
+    monkeypatch.setattr(CPUWorker, "_should_warm_up_model", lambda self: False)
+    monkeypatch.setattr(
+        cpu_worker_mod,
+        "get_allowed_cpu_list",
+        lambda: [SimpleNamespace(numa_node=0)],
+    )
+    monkeypatch.setattr(
+        cpu_worker_mod,
+        "get_memory_node_info",
+        lambda node: MemoryNodeInfo(total_memory=total, available_memory=available),
+    )
+    process = SimpleNamespace(memory_info=lambda: SimpleNamespace(rss=rss))
+    monkeypatch.setattr(cpu_worker_mod.psutil, "Process", lambda pid: process)
+    return worker
+
+
+def test_footprint_above_reservation_asks_for_a_larger_reservation(monkeypatch):
+    # 0.76 GiB already in use against a 0.16 GiB reservation (2% of 8 GiB).
+    worker = _worker(monkeypatch, reserved=int(0.16 * GiB), rss=int(0.76 * GiB))
+    with pytest.raises(ValueError) as exc:
+        worker.determine_available_memory()
+    msg = str(exc.value)
+    assert "increase `--gpu-memory-utilization` to more than 0.10" in msg
+    assert "--kv-cache-memory-bytes" in msg
+    # Neither the misleading remedy nor a negative KV size.
+    assert "other processes" not in msg
+    assert "-0." not in msg
+
+
+def test_short_node_memory_still_asks_to_free_memory(monkeypatch):
+    worker = _worker(monkeypatch, reserved=6 * GiB, rss=1 * GiB, available=2 * GiB)
+    with pytest.raises(ValueError, match="Reduce CPU memory used by other processes"):
+        worker.determine_available_memory()
+
+
+def test_budget_that_fits_sizes_the_kv_cache(monkeypatch):
+    worker = _worker(monkeypatch, reserved=3 * GiB, rss=1 * GiB)
+    assert worker.determine_available_memory() == 2 * GiB
diff --git a/vllm/v1/worker/cpu_worker.py b/vllm/v1/worker/cpu_worker.py
index b103215d65..22b8618f3b 100644
--- a/vllm/v1/worker/cpu_worker.py
+++ b/vllm/v1/worker/cpu_worker.py
@@ -234,10 +234,24 @@ class CPUWorker(Worker):
         else:
             consumed_memory = psutil.Process(os.getpid()).memory_info().rss
             requested_memory_for_kv = int(self.requested_cpu_memory - consumed_memory)
-            if (
-                requested_memory_for_kv <= 0
-                or requested_memory_for_kv > available_memory
-            ):
+            if requested_memory_for_kv <= 0:
+                # The worker's own footprint already fills its reservation, so
+                # freeing memory elsewhere cannot help: only a larger
+                # reservation or an explicit KV cache size can.
+                in_use = math.ceil(consumed_memory / memory_status.total_memory * 100)
+                raise ValueError(
+                    f"The worker already uses {format_gib(consumed_memory)} GiB "
+                    "(model weights and runtime), more than the "
+                    f"{format_gib(self.requested_cpu_memory)} GiB of CPU memory "
+                    f"reserved for it on node {cpu_core.numa_node}, leaving none "
+                    "for the KV cache. On the CPU backend, the "
+                    "`--gpu-memory-utilization` flag controls the fraction of CPU "
+                    "memory reserved (despite its name). To resolve: increase "
+                    f"`--gpu-memory-utilization` to more than {in_use / 100:.2f}, "
+                    "or set the KV cache size explicitly with "
+                    "`--kv-cache-memory-bytes` or VLLM_CPU_KVCACHE_SPACE."
+                )
+            if requested_memory_for_kv > available_memory:
                 raise ValueError(
                     f"Available memory on node {cpu_core.numa_node} "
                     f"({format_gib(available_memory)}/"