diff --git a/tpu_sync/common/BUILD b/tpu_sync/common/BUILD
index 4283f121c..b04d05871 100644
--- a/tpu_sync/common/BUILD
+++ b/tpu_sync/common/BUILD
@@ -48,6 +48,28 @@ cc_test(
     ],
 )
 
+cc_library(
+    name = "accept_backoff",
+    srcs = ["accept_backoff.cc"],
+    hdrs = ["accept_backoff.h"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "@com_google_absl//absl/log",
+        "@com_google_absl//absl/strings",
+        "@com_google_absl//absl/time",
+    ],
+)
+
+cc_test(
+    name = "accept_backoff_test",
+    srcs = ["accept_backoff_test.cc"],
+    deps = [
+        ":accept_backoff",
+        "@com_google_absl//absl/time",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
 cc_library(
     name = "detached_thread_group",
     srcs = ["detached_thread_group.cc"],
diff --git a/tpu_sync/common/accept_backoff.cc b/tpu_sync/common/accept_backoff.cc
new file mode 100644
index 000000000..f57d385f6
--- /dev/null
+++ b/tpu_sync/common/accept_backoff.cc
@@ -0,0 +1,91 @@
+// Copyright 2026 Google LLC.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "tpu_sync/common/accept_backoff.h"
+
+#include <algorithm>
+#include <cerrno>
+#include <cstring>
+
+#include "absl/log/log.h"
+#include "absl/strings/string_view.h"
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+
+namespace tpu_raiden {
+namespace {
+
+constexpr absl::Duration kInitialBackoff = absl::Milliseconds(10);
+constexpr absl::Duration kMaxBackoff = absl::Milliseconds(500);
+
+}  // namespace
+
+AcceptErrorAction ClassifyAcceptError(int err) {
+  switch (err) {
+    // A signal arrived, or the peer went away before we accepted it. Both are
+    // routine; backing off would penalise healthy peers that reconnect often.
+    case EINTR:
+    case ECONNABORTED:
+    case EPROTO:
+      return AcceptErrorAction::kRetry;
+    // Out of descriptors, buffers or memory.
+    case EMFILE:
+    case ENFILE:
+    case ENOBUFS:
+    case ENOMEM:
+      return AcceptErrorAction::kBackoff;
+    // The descriptor is closed, or is no longer a listening socket. On Linux
+    // EINVAL is what shutdown() on a listening socket reports, which is how
+    // teardown unblocks accept().
+    case EBADF:
+    case EINVAL:
+    case ENOTSOCK:
+    case EOPNOTSUPP:
+      return AcceptErrorAction::kStop;
+    default:
+      // Unknown, so assume it can repeat. Backing off is the safe default:
+      // the worst case is a slower retry, not a spin.
+      return AcceptErrorAction::kBackoff;
+  }
+}
+
+absl::Duration NextAcceptBackoff(absl::Duration current) {
+  if (current <= absl::ZeroDuration()) return kInitialBackoff;
+  return std::min(2 * current, kMaxBackoff);
+}
+
+AcceptBackoff::AcceptBackoff(absl::string_view name) : name_(name) {}
+
+bool AcceptBackoff::OnError(int err) {
+  switch (ClassifyAcceptError(err)) {
+    case AcceptErrorAction::kRetry:
+      return true;
+    case AcceptErrorAction::kBackoff:
+      delay_ = NextAcceptBackoff(delay_);
+      LOG_EVERY_N_SEC(WARNING, 1)
+          << name_ << " accept() failed: " << std::strerror(err)
+          << "; retrying in " << delay_;
+      absl::SleepFor(delay_);
+      return true;
+    case AcceptErrorAction::kStop:
+      LOG(ERROR) << name_ << " stopped accepting connections: "
+                 << std::strerror(err);
+      return false;
+  }
+  return false;
+}
+
+void AcceptBackoff::OnSuccess() { delay_ = absl::ZeroDuration(); }
+
+}  // namespace tpu_raiden
diff --git a/tpu_sync/common/accept_backoff.h b/tpu_sync/common/accept_backoff.h
new file mode 100644
index 000000000..e5fef6b6c
--- /dev/null
+++ b/tpu_sync/common/accept_backoff.h
@@ -0,0 +1,72 @@
+// Copyright 2026 Google LLC.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef THIRD_PARTY_TPU_RAIDEN_TPU_SYNC_COMMON_ACCEPT_BACKOFF_H_
+#define THIRD_PARTY_TPU_RAIDEN_TPU_SYNC_COMMON_ACCEPT_BACKOFF_H_
+
+#include <string>
+
+#include "absl/strings/string_view.h"
+#include "absl/time/time.h"
+
+namespace tpu_raiden {
+
+// What an accept loop should do after accept(), or the poll() that guards it,
+// reports an error.
+enum class AcceptErrorAction {
+  // Routine, and the listener is healthy: retry straight away.
+  kRetry,
+  // The process is out of a resource. The connection that could not be
+  // accepted stays queued, so an immediate retry fails the same way and the
+  // loop spins at 100% CPU, starving the very threads that would free the
+  // resource. Sleep first.
+  kBackoff,
+  // The listening descriptor is not usable any more: leave the loop.
+  kStop,
+};
+
+AcceptErrorAction ClassifyAcceptError(int err);
+
+// Delay to wait after |current| (pass zero for the first backoff). Capped, so
+// a stop request is never delayed by more than one step.
+absl::Duration NextAcceptBackoff(absl::Duration current);
+
+// Backoff state for one accept loop. Not thread-safe: each loop owns one on
+// its own stack.
+class AcceptBackoff {
+ public:
+  // |name| identifies the listener in log messages.
+  explicit AcceptBackoff(absl::string_view name);
+
+  // Handles a failed accept(). Sleeps for errors worth retrying slowly, and
+  // returns false if the caller must leave its accept loop.
+  //
+  // Callers check their own stopping flag before calling this, so a kStop
+  // error here always means the descriptor died unexpectedly and is logged as
+  // an error rather than as routine shutdown.
+  bool OnError(int err);
+
+  // Resets the delay. Call after a successful accept().
+  void OnSuccess();
+
+  absl::Duration delay_for_testing() const { return delay_; }
+
+ private:
+  const std::string name_;
+  absl::Duration delay_ = absl::ZeroDuration();
+};
+
+}  // namespace tpu_raiden
+
+#endif  // THIRD_PARTY_TPU_RAIDEN_TPU_SYNC_COMMON_ACCEPT_BACKOFF_H_
diff --git a/tpu_sync/common/accept_backoff_test.cc b/tpu_sync/common/accept_backoff_test.cc
new file mode 100644
index 000000000..007a70e75
--- /dev/null
+++ b/tpu_sync/common/accept_backoff_test.cc
@@ -0,0 +1,90 @@
+// Copyright 2026 Google LLC.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "tpu_sync/common/accept_backoff.h"
+
+#include <cerrno>
+
+#include <gtest/gtest.h>
+#include "absl/time/time.h"
+
+namespace tpu_raiden {
+namespace {
+
+TEST(ClassifyAcceptErrorTest, RoutineErrorsRetryImmediately) {
+  EXPECT_EQ(ClassifyAcceptError(EINTR), AcceptErrorAction::kRetry);
+  EXPECT_EQ(ClassifyAcceptError(ECONNABORTED), AcceptErrorAction::kRetry);
+}
+
+TEST(ClassifyAcceptErrorTest, ResourceExhaustionBacksOff) {
+  EXPECT_EQ(ClassifyAcceptError(EMFILE), AcceptErrorAction::kBackoff);
+  EXPECT_EQ(ClassifyAcceptError(ENFILE), AcceptErrorAction::kBackoff);
+  EXPECT_EQ(ClassifyAcceptError(ENOBUFS), AcceptErrorAction::kBackoff);
+  EXPECT_EQ(ClassifyAcceptError(ENOMEM), AcceptErrorAction::kBackoff);
+}
+
+TEST(ClassifyAcceptErrorTest, UnusableDescriptorStops) {
+  EXPECT_EQ(ClassifyAcceptError(EBADF), AcceptErrorAction::kStop);
+  EXPECT_EQ(ClassifyAcceptError(ENOTSOCK), AcceptErrorAction::kStop);
+  // What shutdown() on a listening socket reports, so teardown relies on it.
+  EXPECT_EQ(ClassifyAcceptError(EINVAL), AcceptErrorAction::kStop);
+}
+
+TEST(ClassifyAcceptErrorTest, UnknownErrorsBackOffRatherThanSpin) {
+  EXPECT_EQ(ClassifyAcceptError(EIO), AcceptErrorAction::kBackoff);
+}
+
+TEST(NextAcceptBackoffTest, GrowsFromZero) {
+  const absl::Duration first = NextAcceptBackoff(absl::ZeroDuration());
+  EXPECT_GT(first, absl::ZeroDuration());
+  EXPECT_GT(NextAcceptBackoff(first), first);
+}
+
+TEST(NextAcceptBackoffTest, CapsLowEnoughNotToDelayTeardown) {
+  absl::Duration delay = absl::ZeroDuration();
+  for (int i = 0; i < 30; ++i) {
+    delay = NextAcceptBackoff(delay);
+  }
+  EXPECT_EQ(delay, NextAcceptBackoff(delay));
+  EXPECT_LE(delay, absl::Seconds(1));
+}
+
+TEST(AcceptBackoffTest, KeepsLoopRunningForRoutineErrors) {
+  AcceptBackoff backoff("test");
+  EXPECT_TRUE(backoff.OnError(EINTR));
+  // Routine errors must not slow the loop down.
+  EXPECT_EQ(backoff.delay_for_testing(), absl::ZeroDuration());
+}
+
+TEST(AcceptBackoffTest, LeavesLoopWhenDescriptorIsUnusable) {
+  AcceptBackoff backoff("test");
+  EXPECT_FALSE(backoff.OnError(EBADF));
+}
+
+TEST(AcceptBackoffTest, GrowsDelayWhileFailingAndResetsOnSuccess) {
+  AcceptBackoff backoff("test");
+
+  EXPECT_TRUE(backoff.OnError(EMFILE));
+  const absl::Duration first = backoff.delay_for_testing();
+  EXPECT_GT(first, absl::ZeroDuration());
+
+  EXPECT_TRUE(backoff.OnError(EMFILE));
+  EXPECT_GT(backoff.delay_for_testing(), first);
+
+  backoff.OnSuccess();
+  EXPECT_EQ(backoff.delay_for_testing(), absl::ZeroDuration());
+}
+
+}  // namespace
+}  // namespace tpu_raiden
diff --git a/tpu_sync/kv_cache/BUILD b/tpu_sync/kv_cache/BUILD
index 88cfeaca3..b88cf5d19 100644
--- a/tpu_sync/kv_cache/BUILD
+++ b/tpu_sync/kv_cache/BUILD
@@ -833,6 +833,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         ":kv_cache_manager_base",
+        "//tpu_sync/common:accept_backoff",
         "//tpu_sync/common:detached_thread_group",
         "//tpu_sync/rpc:raiden_service_cc_proto",
         "@com_google_absl//absl/log",
diff --git a/tpu_sync/kv_cache/kv_cache_listener.cc b/tpu_sync/kv_cache/kv_cache_listener.cc
index ea75971f1..ab28b1c36 100644
--- a/tpu_sync/kv_cache/kv_cache_listener.cc
+++ b/tpu_sync/kv_cache/kv_cache_listener.cc
@@ -30,6 +30,7 @@
 
 #include "absl/log/log.h"
 #include "absl/status/status.h"
+#include "tpu_sync/common/accept_backoff.h"
 #include "tpu_sync/kv_cache/kv_cache_manager_base.h"
 #include "tpu_sync/rpc/raiden_service.pb.h"
 
@@ -137,15 +138,19 @@ KVCacheListener::~KVCacheListener() {
 }
 
 void KVCacheListener::ListenerLoop() {
+  AcceptBackoff backoff("KVCacheListener");
   while (!stopping_) {
     sockaddr_in6 client_addr{};
     socklen_t client_len = sizeof(client_addr);
     int client_fd = accept(
         server_fd_, reinterpret_cast<sockaddr*>(&client_addr), &client_len);
     if (client_fd < 0) {
+      const int err = errno;
       if (stopping_) break;
+      if (!backoff.OnError(err)) break;
       continue;
     }
+    backoff.OnSuccess();
 
     connection_threads_.Spawn(
         [this, client_fd] { ConnectionWorker(client_fd); });
diff --git a/tpu_sync/kv_cache/reshard/BUILD b/tpu_sync/kv_cache/reshard/BUILD
index 6aad0ba93..fbb6e58ed 100644
--- a/tpu_sync/kv_cache/reshard/BUILD
+++ b/tpu_sync/kv_cache/reshard/BUILD
@@ -44,6 +44,7 @@ cc_library(
     copts = RESHARD_COPTS,
     features = ["-use_header_modules"],
     deps = [
+        "//tpu_sync/common:accept_backoff",
         "//tpu_sync/common:detached_thread_group",
         "@com_google_absl//absl/log",
         "@com_google_absl//absl/status",
diff --git a/tpu_sync/kv_cache/reshard/framed_rpc.cc b/tpu_sync/kv_cache/reshard/framed_rpc.cc
index 208b68a37..9255e7c26 100644
--- a/tpu_sync/kv_cache/reshard/framed_rpc.cc
+++ b/tpu_sync/kv_cache/reshard/framed_rpc.cc
@@ -34,6 +34,7 @@
 #include "absl/strings/string_view.h"
 #include "absl/time/clock.h"
 #include "absl/time/time.h"
+#include "tpu_sync/common/accept_backoff.h"
 
 namespace tpu_raiden {
 namespace kv_cache {
@@ -252,15 +253,19 @@ void FramedServer::Stop() {
 }
 
 void FramedServer::AcceptLoop() {
+  AcceptBackoff backoff("FramedServer");
   while (!stopping_) {
     sockaddr_storage client_addr{};
     socklen_t client_len = sizeof(client_addr);
     int client_fd = accept(
         server_fd_, reinterpret_cast<sockaddr*>(&client_addr), &client_len);
     if (client_fd < 0) {
+      const int err = errno;
       if (stopping_) break;
+      if (!backoff.OnError(err)) break;
       continue;
     }
+    backoff.OnSuccess();
     if (stopping_) {
       close(client_fd);
       break;
diff --git a/tpu_sync/transport/lib/BUILD b/tpu_sync/transport/lib/BUILD
index c16adcf7b..29d59814b 100644
--- a/tpu_sync/transport/lib/BUILD
+++ b/tpu_sync/transport/lib/BUILD
@@ -176,6 +176,7 @@ cc_library(
         ":histogram",
         ":raw_buffer_transport_delegate",
         ":transport_adapter",
+        "//tpu_sync/common:accept_backoff",
         "//tpu_sync/common:detached_thread_group",
         "//tpu_sync/core:numa_thread_pool",
         "//tpu_sync/transport:buffer_push_task",
diff --git a/tpu_sync/transport/lib/raw_buffer_transport.cc b/tpu_sync/transport/lib/raw_buffer_transport.cc
index 2e69327c9..1f1b50142 100644
--- a/tpu_sync/transport/lib/raw_buffer_transport.cc
+++ b/tpu_sync/transport/lib/raw_buffer_transport.cc
@@ -51,6 +51,7 @@
 #include "absl/synchronization/blocking_counter.h"
 #include "absl/synchronization/mutex.h"
 #include "absl/types/span.h"
+#include "tpu_sync/common/accept_backoff.h"
 #include "tpu_sync/core/numa_thread_pool.h"
 #include "tpu_sync/transport/buffer_push_task.h"
 #include "tpu_sync/transport/lib/transport_adapter.h"
@@ -469,14 +470,20 @@ RawBufferTransport::RegisterPspPeer(uint32_t client_spi,
 }
 
 void RawBufferTransport::ListenerLoop() {
+  AcceptBackoff backoff("RawBufferTransport");
   while (!stopping_) {
     DCHECK(IsValidSocket(server_fd_));
     struct pollfd pfd;
     pfd.fd = server_fd_;
     pfd.events = POLLIN;
     int ret = poll(&pfd, 1, 50);
-    if (ret <= 0) {
+    if (ret == 0) continue;  // Idle tick: nothing waiting to be accepted.
+    if (ret < 0) {
+      const int err = errno;
       if (stopping_) break;
+      // poll() reports the same descriptor and resource errors as accept(),
+      // so the two share one policy.
+      if (!backoff.OnError(err)) break;
       continue;
     }
 
@@ -485,9 +492,12 @@ void RawBufferTransport::ListenerLoop() {
     int client_fd = accept(
         server_fd_, reinterpret_cast<struct sockaddr*>(&client_addr), &clilen);
     if (client_fd < 0) {
+      const int err = errno;
       if (stopping_) break;
+      if (!backoff.OnError(err)) break;
       continue;
     }
+    backoff.OnSuccess();
     if (require_psp_tcp_ && !PspEnabled(client_fd)) {
       close(client_fd);
       LOG_EVERY_N_SEC(ERROR, 1)
diff --git a/tpu_sync/weight_sync/BUILD b/tpu_sync/weight_sync/BUILD
index 1a52a9188..208e99bed 100644
--- a/tpu_sync/weight_sync/BUILD
+++ b/tpu_sync/weight_sync/BUILD
@@ -56,6 +56,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         ":tiling_utils",
+        "//tpu_sync/common:accept_backoff",
         "//tpu_sync/common:detached_thread_group",
         "//tpu_sync/common:trace",
         "//tpu_sync/core:host_memory_allocator",
diff --git a/tpu_sync/weight_sync/weight_synchronizer_listener.cc b/tpu_sync/weight_sync/weight_synchronizer_listener.cc
index b33e39a74..2cf807d71 100644
--- a/tpu_sync/weight_sync/weight_synchronizer_listener.cc
+++ b/tpu_sync/weight_sync/weight_synchronizer_listener.cc
@@ -30,6 +30,7 @@
 
 #include "absl/container/flat_hash_map.h"
 #include "absl/log/log.h"
+#include "tpu_sync/common/accept_backoff.h"
 #include "tpu_sync/rpc/raiden_service.pb.h"
 #include "tpu_sync/weight_sync/weight_synchronizer_base.h"
 
@@ -100,6 +101,7 @@ WeightSynchronizerListener::~WeightSynchronizerListener() {
 }
 
 void WeightSynchronizerListener::ListenerLoop() {
+  AcceptBackoff backoff("WeightSynchronizerListener");
   while (!stopping_) {
     sockaddr_in6 client_addr{};
     socklen_t client_len = sizeof(client_addr);
@@ -107,9 +109,12 @@ void WeightSynchronizerListener::ListenerLoop() {
         accept(server_fd_.load(), reinterpret_cast<sockaddr*>(&client_addr),
                &client_len);
     if (client_fd < 0) {
+      const int err = errno;
       if (stopping_) break;
+      if (!backoff.OnError(err)) break;
       continue;
     }
+    backoff.OnSuccess();
 
     connection_threads_.Spawn(
         [this, client_fd] { ConnectionWorker(client_fd); });