diff --git a/src/modules/TrafficManagementModule.cpp b/src/modules/TrafficManagementModule.cpp
index ea5054ded..30a298019 100644
--- a/src/modules/TrafficManagementModule.cpp
+++ b/src/modules/TrafficManagementModule.cpp
@@ -1092,6 +1092,12 @@ bool TrafficManagementModule::shouldRespondToNodeInfo(const meshtastic_MeshPacke
     if (!sendResponse)
         return true;
 
+    // Checked here, once a reply would actually go out, so declined requests do not consume the budget.
+    if (!directResponseAllowed(getFrom(p), clockMs())) {
+        TM_LOG_DEBUG("NodeInfo direct response throttled for 0x%08x", getFrom(p));
+        return false;
+    }
+
     meshtastic_MeshPacket *reply = router->allocForSending();
     if (!reply) {
         TM_LOG_WARN("NodeInfo direct response dropped: no packet buffer");
@@ -1144,6 +1150,43 @@ bool TrafficManagementModule::shouldRespondToNodeInfo(const meshtastic_MeshPacke
     return true;
 }
 
+bool TrafficManagementModule::directResponseAllowed(NodeNum requestor, uint32_t nowMs)
+{
+    // Reached from the packet path and from runOnce, so the throttle state needs the same lock as the cache.
+    concurrency::LockGuard guard(&cacheLock);
+
+    if (lastDirectResponseMs != 0 && (nowMs - lastDirectResponseMs) < kDirectResponseGlobalMs)
+        return false;
+
+    DirectResponseThrottleEntry *slot = nullptr;
+    for (size_t i = 0; i < kDirectResponseTrackedRequestors; i++) {
+        if (directResponseSeen[i].requestor == requestor) {
+            if ((nowMs - directResponseSeen[i].lastReplyMs) < kDirectResponsePerRequestorMs)
+                return false;
+            slot = &directResponseSeen[i];
+            break;
+        }
+    }
+
+    if (slot == nullptr) {
+        // Unseen requester: take a free slot, otherwise reuse the least recently used one.
+        slot = &directResponseSeen[0];
+        for (size_t i = 0; i < kDirectResponseTrackedRequestors; i++) {
+            if (directResponseSeen[i].requestor == 0) {
+                slot = &directResponseSeen[i];
+                break;
+            }
+            if (directResponseSeen[i].lastReplyMs < slot->lastReplyMs)
+                slot = &directResponseSeen[i];
+        }
+    }
+
+    slot->requestor = requestor;
+    slot->lastReplyMs = nowMs;
+    lastDirectResponseMs = nowMs;
+    return true;
+}
+
 bool TrafficManagementModule::isMinHopsFromRequestor(const meshtastic_MeshPacket *p) const
 {
     int8_t hopsAway = getHopsAway(*p, -1);
diff --git a/src/modules/TrafficManagementModule.h b/src/modules/TrafficManagementModule.h
index f4a1d5954..a3619fc0a 100644
--- a/src/modules/TrafficManagementModule.h
+++ b/src/modules/TrafficManagementModule.h
@@ -293,6 +293,19 @@ class TrafficManagementModule : public MeshModule, private concurrency::OSThread
 
     bool shouldDropPosition(const meshtastic_MeshPacket *p, const meshtastic_Position *pos, uint32_t nowMs);
     bool shouldRespondToNodeInfo(const meshtastic_MeshPacket *p, bool sendResponse);
+
+    // Replies go to the requesting packet's unauthenticated `from`, so space them per requester to bound
+    // what any one node can be made to receive, plus a global floor on airtime.
+    static constexpr uint32_t kDirectResponsePerRequestorMs = 60'000UL;
+    static constexpr uint32_t kDirectResponseGlobalMs = 1'000UL;
+    static constexpr size_t kDirectResponseTrackedRequestors = 8;
+    struct DirectResponseThrottleEntry {
+        NodeNum requestor;
+        uint32_t lastReplyMs;
+    };
+    DirectResponseThrottleEntry directResponseSeen[kDirectResponseTrackedRequestors] = {};
+    uint32_t lastDirectResponseMs = 0;
+    bool directResponseAllowed(NodeNum requestor, uint32_t nowMs);
     bool isMinHopsFromRequestor(const meshtastic_MeshPacket *p) const;
     bool isRateLimited(NodeNum from, uint32_t nowMs);
     bool shouldDropUnknown(const meshtastic_MeshPacket *p, uint32_t nowMs);