Every game community eventually hits the same wall: the server that ran beautifully at 30 players starts stuttering at 70. The instinct is to add RAM. Sometimes that works. Often it does not, because the bottleneck is tick time, not memory — and doubling RAM on a CPU-bound server just doubles your bill. The honest way to size a game server is to define four separate budgets — CPU, memory, storage, and network — and match each one to the play style your community actually runs.
The numbers below assume Minecraft Java as the reference workload because it is the most common and the most demanding per player, with notes for Rust, Valheim, ARK, and CS2 where behaviour diverges sharply. If you are still deciding between machine classes, our dedicated server hosting comparison covers the hardware tiers these budgets map to.
Budget One: Vanilla Survival, 20–40 Players
This is the cheapest tier and the one most communities over-buy. A vanilla or lightly-plugged Minecraft server at 30 concurrent players needs roughly 4–6 GB heap, 2 vCPU at 3.0 GHz or better, and 20 GB of storage plus world growth. Budget $15–$30/month. Tick time on a decent Ryzen at 30 players sits around 12–18 ms — well inside the 50 ms budget, so you have 3× headroom before anything visible happens.
The critical spec is single-thread clock, not core count. Minecraft’s main thread will not use more than one core for world ticking; extra cores help only with async chunk loading and the OS. A 2-core 3.5 GHz machine outperforms a 4-core 2.2 GHz machine for this workload. Do not accept a vCPU allocation that is steal-heavy — check CPU steal time in your metrics; above 5% sustained, move hosts.
Budget Two: Modded / Plugin-Heavy, 40–100 Players
This is where cost jumps. A 150-mod Forge pack (All the Mods 9, FTB Skies) uses 10–14 GB heap at 20 players and grows non-linearly. At 60 players you should plan 24–32 GB heap, 4–6 vCPU at 3.8 GHz+, and 500 GB–1 TB NVMe for pack files plus world. Budget $80–$180/month for a properly dedicated box. MSPT on a well-tuned ATM9 server at 50 players lands around 25–35 ms; at 80 players with heavy automation, expect 40–48 ms and accept occasional tick spikes.
Memory here is not optional padding — modded worlds genuinely hold hundreds of thousands of loaded chunks and block entities in RAM. But be precise: allocating 64 GB heap to a Java server with a 30 GB live set causes garbage-collection pauses of 200 ms+, which players feel as rubber-banding. Set heap at 1.5× live set, choose G1GC with a 200 ms max pause target, and leave 4–8 GB free for the OS page cache. On a dedicated box, that page cache is what makes world saves fast.
Budget Three: Public Competitive / High Concurrency, 100–250 Players
Public servers with 150+ players and open registration need a fundamentally different posture: multiple cores, high memory bandwidth, and aggressive sharding. Plan 8–16 vCPU at 3.5 GHz+, 64–128 GB RAM, and 2 × 1 TB NVMe for mirroring. Budget $200–$450/month. You will almost always run this as multiple instances behind a proxy (Velocity or BungeeCord) rather than a single world — which is what actually scales, not a bigger single server.
For CS2 at 128-tick, the budget is different again: a single 128-tick CS2 competitive server needs about 1.5–2 cores at 4.0 GHz+ and 2 GB RAM, and serves 10 players. Scaling to 20 concurrent match servers means 20–30 cores and 48 GB RAM — this is where a 32-core EPYC earns its price. A Rust server at 200 players wants 8–12 GB RAM and 4 cores at 4.2 GHz+, and world loads spike storage hard; NVMe with 500K+ IOPS is mandatory.
Budget Four: The Storage and Backup Budget Nobody Plans
Communities forget that storage grows. A Minecraft world with active exploration adds 5–20 GB per month; a modded world with dimensional travel can add 60 GB. Rust wipes monthly but the map is 4–8 GB per wipe. ARK saves balloon to 30–40 GB. Over 12 months, a “500 GB is plenty” assumption becomes a full disk at 3 a.m. Plan primary storage at 2× your year-one estimate, and budget for off-box backups at 1× primary.
Backups also cost IOPS, not just bytes. A full 200 GB world archive compresses and writes in 6–15 minutes on NVMe; on SATA SSD it can take 40. Run backups during off-peak hours and verify that your provider’s snapshot mechanism does not pause the VM — some block-storage snapshots introduce 2–5 second I/O freezes, which shows up in-game as a mass disconnect.
Reading Tick Time Instead of Guessing
The single most useful diagnostic is mean ticks per second and MSPT. Run /tps and /mspt, or use Spark’s profiler. If MSPT is consistently above 40, you are one player cohort away from visible lag. If TPS drops below 19.5 during scheduled events, your budget is undersized for peak, not average — and peaks, not averages, are what players remember.
Track three numbers weekly: peak concurrent players, p95 MSPT, and disk latency. If p95 MSPT rises 30% while peak players rise 10%, your workload has changed shape — usually a new plugin, a redstone farm, or an entity-dense region. That is a tuning problem, not a hardware problem, and buying RAM will not fix it.
Where Communities Overspend
The two most common overspends are RAM-for-CPU and enterprise-for-community. Buying 128 GB to run a 30-player vanilla server wastes $80/month. Buying dual-Xeon enterprise hardware to host a 60-player modded server wastes another $150/month and delivers worse single-thread performance than a $120 Ryzen box. The right spend is a fast consumer-class CPU, adequate RAM at 1.5× live set, and enterprise NVMe.
The second overspend is over-provisioning capacity for a peak that occurs twice a year. Size for your 95th percentile concurrency, not your launch-day spike, and plan a burst path — either a second instance you spin up for events or a provider that allows short-term upgrades. Horizontal scaling with a proxy is almost always cheaper than vertical scaling once you pass 150 players.
Sizing Table and Decision Rule
Summarised: vanilla 30 players = 6 GB / 2 vCPU / $20. Modded 50 players = 24 GB / 4 vCPU / $110. Public 150 players = 64 GB / 8 vCPU across instances / $250. CS2 tick farm = cores over RAM, 30 cores for 20 servers / $400. Use these as a floor, then add 30% headroom and validate with a 72-hour load test using your real world and plugin set.
The decision rule is simple: if MSPT is your problem, buy clock speed; if GC pauses are your problem, right-size heap and tune GC; if world loads are your problem, buy IOPS; if disconnects cluster during saves, buy off-box backup throughput. Start with the dedicated server plans we track and map your play style to one of the four budgets rather than buying the biggest number on the page.



Leave a Reply
You must be logged in to post a comment.