· AMBER team
Why we cap the number of instances per host
The arithmetic behind a node that runs at 61% CPU instead of 95%, and why that number is the product rather than a missed target.
When we opened the first node, the obvious way to price a game server was to look at what the large panels charge and go slightly below. That is a race to the bottom with a predictable end: to stay profitable at those prices, a host has to put more instances on the same machine than the machine can serve well.
We went the other way. Each host gets a fixed instance budget, calculated once from the templates it will run and the worst-case load we are willing to accept. In practice that means average CPU utilisation around 61% and memory around 48%.
What the headroom actually buys
A Minecraft server is not a steady load. Chunk generation, a player entering a region that nobody has visited since the last restart, a modpack starting up — each of those spikes. If a host is already at 90%, the spike becomes someone else's lag. The headroom is not wasted capacity; it is the buffer that absorbs the spikes.
The second thing it buys is honesty about memory. A 4 GB plan means 4 GB that the instance can use, not 4 GB shared with a neighbour who happens to be idle. When you see 4 GB in the panel, that is a reservation, not a hope.
What it costs
Instances per host is the single biggest lever on price. Halving the number roughly doubles the hardware cost per instance. We compensate by not paying for things that do not make the server better: no affiliate programme, no sponsored videos, no sales team, a panel that does one job.
The result is a price that is lower than the large panels but not the lowest on the market — the cheapest offers are usually the ones with the most overselling behind them, and that is exactly what we are trying to avoid.
How we decide a node is full
Sustained load above 70% for two weeks is the trigger. New orders for that node go to the next one in the region, and we start planning a new host. A node can look “not full” by disk and still be full by CPU; we count the latter.
If you are on a node that fills up, nothing changes for you. The point of the cap is that the experience stays the same when the node is busy.