The capacity study ended with two workloads that met the declared checks in three clean runs: a dense 64-client fight and a more spread-out 128-client fight on Highlands. The game's configured limit remains 64 players, including bots.
The larger result belongs to a particular research workload. Highlands spreads players and projectiles differently from Ironworks. Passing a Highlands free-for-all does not establish a 128-player Shogun match with full bot AI, a crowded Ironworks fight or Internet performance.
What the test clients actually did
A Windows machine with an i5-12500H ran the dedicated server. A separate Linux machine with a Ryzen 7 5800H generated real protocol clients over wired gigabit Ethernet. These automated clients exercised movement, weapons and networking. They did not each render a game window, mix audio or run a full server bot brain.
Before confirming a workload, I set limits for the time spent on each server simulation tick, the age of received positions, action feedback and the generator's input schedule. A recommendation needed three clean runs, including a five-minute active run, with no invalid client records or input-buffer handoffs. A handoff is a recovery change to the timing agreement that places client commands on the server's simulation schedule; the direct-LAN screen required that agreement to continue without one. Runs affected by unrelated machine activity stayed in the report with that qualification.
The two workloads that passed
Ironworks with 64 star users passed all three confirmations. Highlands with 128 clients in mixed weapon groups, with net, shadow and chi enabled, also passed three, including its long run.
Across the Highlands confirmations, worst-client position-age p99 ranged from 17.66 to 25.47 ms. This measures how old received positions were at the 99th percentile. Worst-client weapon-request feedback p99 ranged from 21.68 to 28.67 ms, measuring the wait for a result to the request.
Each client has its own p99—the value within which 99% of its samples fall—and the report takes the worst client's value. The results are not one pooled percentile across all clients, and weapon-request feedback does not include the time a projectile takes to fly to its target. The slowest 1% of samples can exceed p99.
The long Highlands run still recorded 614 late input commands and 13 late actions. It passed the declared numerical screen, but that did not make every event timely. Those counts remain part of the result.
Why the other runs did not raise the limit
At 160 clients on Highlands, two clean runs passed, but another needed two input-buffer handoffs. That broke the predeclared zero-handoff rule. A clean repeat of the dense workload at 80 clients also needed two handoffs. Other attempts overlapped unrelated automated tests and could not count as clean confirmations.
Doubling the per-client bandwidth allowance did not establish a reliable 128-client dense fight. One seed missed the action-feedback and position-age limits; the sustained run missed the generator's command interval and needed handoffs. At 192 clients, connections and traffic worked, but the generator could not maintain its input schedule. That left the result unqualified; it did not establish the server's absolute ceiling.
I also tested delay, packet loss and variation in arrival timing, or jitter, separately at 64 clients. All affected delivery, and the combined jitter-and-loss case was substantially worse than the direct local-network case. Those encounters remain on the netcode page, alongside both successful and failed capacity attempts.
Measuring equipment in its encounter
The feature study contains sixteen matched active/unused pairs, covering eight weapons and abilities on Ironworks and Highlands. Each pair used 64 clients and one random seed, comparing a feature in use with its matched unused case.
Grenades added about 9.78 megabits per second of server upload on Ironworks and 18.18 megabits per second on Highlands, relative to those unused cases. These are aggregate server rates, not a rate for each player. Respawns and replenished ammunition allowed different numbers of throws, so the difference cannot be converted into a fixed price per grenade user.
Adding isolated measurements together would have the same problem. A fast renderer, a quick simulation and a timely network test do not jointly certify a match they never ran together. Full Shogun AI, integrated game clients, more hardware and broader Internet conditions still need their own qualification.
The boundary from this study is specific: 64 remains the general reference; 128 is confirmed only for the declared, spread-out Highlands free-for-all research workload on the local network. Inspect the measurements or download their public summary.