What actually bottlenecks
Minecraft ticks on one thread. Single-core speed beats core count every single time, for every player count I've ever profiled. Chunk IO comes second — a big world with players teleporting around is a storage workload, and NVMe shows it. RAM matters only when you're genuinely swapping, or running large view distances with plugin counts that justify it.
The seven2-core rental with a slow base clock loses to the four-core box at 4.5 GHz more often than owners believe.
The checklist in order
- Profile. spark, an evening of real traffic, MSPT on the table.
- Settings. view, simulation, activation range — the three that beat hardware for free.
- Entity caps. Mob limits, hopper ranges, farm rules. One unchecked farm outlifts every other fix.
- Plugin audit. One bad plugin can cost more than all the good ones together.
- Then hardware — in this order: single-core speed, NVMe, RAM, and cores last of all.
Renting vs owning
Rent when you want DDoS filtering and zero operations — that's most networks, and the price is fair insurance. Own when you already debug boxes at night and the hardware cost beats twelve months of rent, because the real cost of owning is the 2am page that nobody else answers.
What I quote clients: a high-clock host with NVMe beats a big-core rent-a-server for the same money at every player count under two hundred. Buy the clock. Profile first, buy the clock second, and sleep better either way.
RnK is a Minecraft developer with six years in the game — plugins, server setups and performance work across Paper, Spigot and Folia backends. Three networks run day to day; the rest of the time is client work. The bench is open on Discord.