This is the most confusing question when renting a server. Buy too little and the server stalls; buy too much and you overpay every month.
In this article we’ll give concrete numbers instead of guesses: how much memory each job needs, and how to measure whether your current server’s memory is enough.
The short answer
| Workload | Recommended memory |
|---|---|
| Single WordPress site, low traffic | 2 GB |
| A few small sites, corporate site | 4 GB |
| Mid-sized e-commerce (WooCommerce, OpenCart) | 8 GB |
| Panel install (Plesk / cPanel) + a few customers | 8 GB |
| High-traffic e-commerce | 16 GB |
| Panel + many customer sites (reseller) | 16–32 GB |
| Separate database server | 16–64 GB |
| Game server (small community) | 4–8 GB |
| Game server (high player count) | 16 GB+ |
| Windows + remote desktop (RDP) | 8 GB |
| Your own virtualisation infrastructure | 64 GB+ |
| Running an AI model | Depends on the model — separate section below |
This table is for Linux installations. If you’ll install Windows the numbers change noticeably; we cover that in a separate section below.
What consumes memory
Three things consume memory on your server, and most people only think of the first.
1. The operating system. Linux uses around 300–500 MB on a server. On Windows this figure is 1.5–2 GB. But that’s consumption when the server is on and doing nothing; on its own it’s misleading for budgeting. We give realistic Windows figures in a separate section below.
2. The web server and application. Apache opens a separate process for each concurrent connection; as connections grow, memory use rises linearly. Nginx does the same job with far less memory. On the PHP side, each request takes up space up to its own memory limit.
3. The database. This is usually the biggest consumer. MySQL/MariaDB reserves the area set by innodb_buffer_pool_size and keeps tables there. If your database doesn’t fit in memory, every query goes to disk and the site slows noticeably.
Practical rule: memory equal to the size of your database should be reserved for the database. If you have a 4 GB database, a total of 8 GB on the server is a reasonable starting point.
If you’ll install Windows
Windows’s memory needs can’t be compared with Linux’s, and they vary significantly from edition to edition.
| Install | Working minimum | Comfortable use |
|---|---|---|
| Windows Server (2019 / 2022) | 4 GB | 8 GB |
| Windows 10 / 11 | 8 GB | 16 GB |
Server editions run noticeably lighter than desktop editions. Windows 10 and 11 carry background components that are useless on a server: store apps, search indexing, automatic update services, cloud sync. Individually a few hundred megabytes, together over a gigabyte. Windows Server does the same job with less memory. If you have the choice, prefer the server edition.
What really consumes memory is not the operating system but what you open on top of it. The moment you connect over remote desktop and open a browser, the math changes: Chrome reserves 300–500 MB per tab. Three tabs, an Excel file and an accounting program add up to 2–3 GB on their own.
That’s why a 4 GB Windows 10 server may look “enough” on paper but constantly falls into swap in practice. The result is what people describe as “the server is slow” — but the server isn’t slow, it’s out of memory.
Our practical advice: if you’ll install Windows, don’t go below 8 GB. If more than one person will connect to the same server over remote desktop, add 2 GB more per person.
How to tell if it’s enough
Don’t guess, measure. On a Linux server:
free -mLook at the available column in the output. Not free, available. Linux uses free memory for cache, so free always looks low and panics people. The real indicator is the available column.
What’s truly critical is swap usage:
free -m | grep SwapSwap is the method of using the disk as memory when memory runs short, and writing to disk is hundreds of times slower than memory. If swap usage is consistently high, your memory isn’t enough; this is the most common cause of server slowness.
Summary:
| Indicator | Meaning |
|---|---|
available above 20% of total memory |
You’re comfortable |
available below 10% |
You’re at the edge, plan an upgrade |
| Swap constantly in use | Out of memory, upgrade now |
Memory on dedicated servers
The table above applies to virtual servers. On the dedicated side the math changes at two points.
1. The floor is higher. Dedicated servers generally start at 16 GB; there’s no such thing as a 2 GB dedicated server. So the “should I get less” question loses its meaning here — it becomes “how far up should I go.”
2. Memory type comes into play. The Xeon series uses ECC memory. ECC corrects single-bit errors as they occur. On a database running 24/7 this can matter more than capacity: even with enough memory, a silent data corruption becomes a problem you notice months later.
A rough guide for dedicated-specific workloads:
| Workload | Recommended memory |
|---|---|
| Single application, medium load | 16–32 GB |
| GPU rendering / AI | 32–64 GB (VRAM is the real deciding factor) |
| High-player-count game server | 32 GB |
| Enterprise database | 64–128 GB, ECC |
| Virtualisation server (your own VMs) | 128 GB+, ECC |
| Many customer sites (reseller) | 64–128 GB |
If you’ll do virtualisation the math is different: add up the memory of each virtual machine you’ll run, add 4–8 GB for the hypervisor, and leave a 20% margin. If you’ll run ten 8 GB VMs, you should be looking at a system around 96–112 GB. Our dual-socket Xeon configurations go up to 512 GB for this need.
Memory for AI
If you’ll run an AI model the math works differently, and here we talk about two separate memories.
If you’ll run the model on the graphics card, the deciding factor is not system memory but the card’s own memory (VRAM). If the model doesn’t fit in VRAM, it won’t run. Rough measure: a 4-bit quantised model needs about half a byte per parameter. By that measure 8 GB of VRAM is for small models, 24 GB and above for large language models.
If you’ll run the model on the processor, the same math applies to system memory, but it runs much slower.
A graphics card is offered only on dedicated servers; it cannot be added to virtual servers.
Three common mistakes
1. Buying extra “just in case.” Renting three times the memory you need means paying every month for unused resources. Because memory can be upgraded later on virtual servers, there’s no gain in buying big from the start.
2. Sizing Windows like Linux. The general table above is for Linux. On Windows the floor is 8 GB, and if you’ll install Windows 10 / 11 the comfortable working limit rises to 16 GB. Looking at a figure calculated for Linux and ordering Windows is the most common cause of a slow server.
3. Panicking over the free column. On Linux, free memory is considered “wasted memory”; the system uses it as disk cache. When your application needs memory, that space is released instantly. The column you should look at is available.
If you realise it’s not enough
On virtual servers a plan upgrade is possible; when your memory falls short you can move to a larger plan. Disk capacity can be increased separately, independent of the plan.
On dedicated servers an upgrade requires a hardware change; the configuration can be changed at order time or afterwards, but it requires planning. That’s why leaving a little margin from the start makes more sense on a dedicated server than on a virtual one.
So the right strategy is to find your need by approaching from below: start with a reasonable plan, measure for a month, and upgrade if it genuinely falls short. This path is almost always cheaper than buying big at the start and overpaying for months.
You can compare plans and memory options on our virtual server page, and for memory-heavy jobs look at our dedicated server lineup.
Getting the RAM right is one part of the bigger picture; to see every step from CPU to disk, read our server rental guide.