diff --git a/drivers/usbdev/cdcncm.c b/drivers/usbdev/cdcncm.c
index d528d8f0da4c2..d892c563deb60 100644
--- a/drivers/usbdev/cdcncm.c
+++ b/drivers/usbdev/cdcncm.c
@@ -1302,8 +1302,18 @@ static int cdcncm_send(FAR struct netdev_lowerhalf_s *dev, FAR netpkt_t *pkt)
     }
   else
     {
-      work_queue(ETHWORK, &self->delaywork, cdcncm_transmit_work, self,
-                 MSEC2TICK(CDCNCM_DGRAM_COMBINE_PERIOD));
+      /* Defer the transmit to the work thread so the caller is not blocked
+       * on wrreq_idle, but do so with ZERO delay.  A non-zero delay here is
+       * quantized to the system tick: MSEC2TICK(1) rounds *up* to a full
+       * tick (10ms at the default 100Hz), so the reply to every single
+       * datagram (ping, TCP ACK, one-MSS HTTP segment) was being held for
+       * 10-20ms, dominating the USB-NIC round-trip time.  Delay 0 wakes the
+       * work thread immediately (sub-ms) while still coalescing any further
+       * datagrams that the network stack appends in the same TX burst before
+       * the worker gets to run.
+       */
+
+      work_queue(ETHWORK, &self->delaywork, cdcncm_transmit_work, self, 0);
     }
 
   return OK;