How many cores and how much memory for which job

This is the most asked question when building a new machine, and it has no single right answer, because the answer depends on the kind of work. Still, jobs fall into a few families, and each family has its own bottleneck.

Why giving too much hurts

Giving a machine more cores than it needs does not make it faster. Extra vCPU lengthens the queue on the server, and every time the machine wants to run it waits for more cores to free up at once. Memory behaves similarly: assigned memory counts as taken from the host even when it is not really used, and it is closed off to the other machines.

The kind of work sets the bottleneck

A media server usually wants transcoding power rather than cores; with hardware acceleration two cores are enough, without it no number of cores will do. A database is sensitive to disk latency: a fast disk and a separate IO thread make more difference than extra cores. Reverse proxies and small panel services ask for almost nothing and run for years on one core and little memory. AI and image work want the opposite: without a graphics card passed through they do not run meaningfully at all, and their memory appetite is high.

The starting rule

Give little, measure, then raise. Adding a core to a running machine is usually easy; taking one back means removing a resource that has become a habit. The same road applies to memory, and memory is the resource that runs out fastest on the host.

Do not decide without measuring

What a machine really needs shows up after it has run for a few days. The first day's guess is only a starting point; the real decision comes from a measurement taken under load.