diff --git a/src/Empostor.Server/Extensions/ServiceProviderExtensions.cs b/src/Empostor.Server/Extensions/ServiceProviderExtensions.cs
index dff1eb1..0ad4709 100644
--- a/src/Empostor.Server/Extensions/ServiceProviderExtensions.cs
+++ b/src/Empostor.Server/Extensions/ServiceProviderExtensions.cs
@@ -9,7 +9,11 @@ namespace Empostor.Server
     {
         public static void AddEventPools(this IServiceCollection services)
         {
-            services.TryAddSingleton<ObjectPoolProvider>(new DefaultObjectPoolProvider());
+            // Larger object-pool retention so high-frequency packet allocation
+            // (MessageReader, event objects) is served from the pool instead of
+            // hammering the GC under load (e.g. 40+ players).
+            services.TryAddSingleton<ObjectPoolProvider>(
+                new DefaultObjectPoolProvider { MaximumRetained = 512 });
 
             services.AddSingleton(serviceProvider =>
             {
diff --git a/src/Empostor.Server/Net/Hazel/HazelConnection.cs b/src/Empostor.Server/Net/Hazel/HazelConnection.cs
index f998783..19cb2fb 100644
--- a/src/Empostor.Server/Net/Hazel/HazelConnection.cs
+++ b/src/Empostor.Server/Net/Hazel/HazelConnection.cs
@@ -4,6 +4,7 @@ using Empostor.Api;
 using Empostor.Api.Config;
 using Empostor.Api.Net;
 using Next.Hazel;
+using Next.Hazel.Udp;
 using Microsoft.Extensions.Logging;
 using Microsoft.Extensions.Options;
 
@@ -21,6 +22,16 @@ namespace Empostor.Server.Net.Hazel
             InnerConnection = innerConnection;
             innerConnection.DataReceived = ConnectionOnDataReceived;
             innerConnection.Disconnected = ConnectionOnDisconnected;
+
+            if (innerConnection is UdpConnection udp)
+            {
+                // Tolerate UDP packet loss under load (e.g. 40+ players). A few
+                // dropped reliable packets or keepalive pings are retransmitted
+                // for longer before the connection is declared dead, so a burst
+                // of loss no longer disconnects players.
+                udp.DisconnectTimeout = 10000;
+                udp.MissingPingsUntilDisconnect = 12;
+            }
         }
 
         public Connection InnerConnection { get; }
diff --git a/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnection.cs b/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnection.cs
index 4f9b12b..a1c2675 100644
--- a/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnection.cs
+++ b/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnection.cs
@@ -33,10 +33,17 @@ public abstract partial class UdpConnection : NetworkConnection
         _readerPool = readerPool;
         _stoppingCts = new CancellationTokenSource();
 
-        Pipeline = Channel.CreateUnbounded<byte[]>(new UnboundedChannelOptions
+        // Bound the inbound queue so a slow/overloaded connection cannot let
+        // packets accumulate without limit. Once full, the oldest queued packet
+        // is dropped (DropOldest) — reliable packets are retransmitted by the
+        // sender anyway and unreliable ones carry fresh state each tick, so this
+        // keeps memory flat under sustained load instead of letting it balloon
+        // until the GC pressure starts disconnecting everyone.
+        Pipeline = Channel.CreateBounded<byte[]>(new BoundedChannelOptions(1024)
         {
             SingleReader = true,
-            SingleWriter = true
+            SingleWriter = true,
+            FullMode = BoundedChannelFullMode.DropOldest,
         });
     }
 
@@ -46,8 +53,13 @@ public abstract partial class UdpConnection : NetworkConnection
 
     public Task StartAsync()
     {
-        // Store the task we're executing
-        _executingTask = Task.Factory.StartNew(ReadAsync, TaskCreationOptions.LongRunning);
+        // Run the read loop on the thread pool instead of a dedicated
+        // long-running thread per connection. ReadAsync is fully async (it
+        // awaits the pipeline channel), so it only occupies a thread-pool
+        // thread while there is actually data to process. With hundreds of
+        // connections this avoids hundreds of dedicated threads (1 MB stack
+        // each) and the resulting context-switch overhead.
+        _executingTask = Task.Run(ReadAsync);
 
         // If the task is completed then return it, this will bubble cancellation and failure to the caller
         if (_executingTask.IsCompleted) return _executingTask;
diff --git a/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnectionListener.cs b/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnectionListener.cs
index 47589f6..167b672 100644
--- a/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnectionListener.cs
+++ b/src/NextFast.Hazel-New/Next.Hazel/Udp/UdpConnectionListener.cs
@@ -1,5 +1,6 @@
 using System;
 using System.Collections.Concurrent;
+using System.Collections.Generic;
 using System.Net;
 using System.Net.Sockets;
 using System.Threading;
@@ -65,6 +66,20 @@ public class UdpConnectionListener : NetworkConnectionListener
         {
         }
 
+        try
+        {
+            // Under high load (many players / high tick rate) the OS-level UDP
+            // receive queue can overflow and silently drop packets; reliable
+            // retransmission / keepalive then misreads that as a dead connection
+            // and disconnects players. Use a large buffer so bursts are absorbed
+            // instead of dropped.
+            _socket.Client.ReceiveBufferSize = 4 * 1024 * 1024;
+            _socket.Client.SendBufferSize = 4 * 1024 * 1024;
+        }
+        catch (SocketException)
+        {
+        }
+
         _reliablePacketTimer = new Timer(ManageReliablePackets, null, 100, Timeout.Infinite);
 
         _allConnections = new ConcurrentDictionary<EndPoint, UdpServerConnection>();
@@ -79,10 +94,25 @@ public class UdpConnectionListener : NetworkConnectionListener
 
     private async void ManageReliablePackets(object? state)
     {
-        foreach (var kvp in _allConnections)
+        // Resend reliably-sent packets for every connection. Run them
+        // concurrently instead of serially awaiting each connection: with many
+        // players one slow connection used to stall retransmission for everyone
+        // else, which made packet loss snowball into mass disconnects.
+        var connections = new List<UdpServerConnection>(_allConnections.Values);
+        var tasks = new Task[connections.Count];
+        for (var i = 0; i < connections.Count; i++)
         {
-            var sock = kvp.Value;
-            await sock.ManageReliablePackets();
+            var conn = connections[i];
+            tasks[i] = conn.ManageReliablePackets().AsTask();
+        }
+
+        try
+        {
+            await Task.WhenAll(tasks);
+        }
+        catch
+        {
+            // ignored
         }
 
         try