What can I run on a NAS?

Why capability matters more than bays

Bay count decides how much storage you can fit; it says nothing about what the box can do. A cheap four-bay ARM unit holds plenty of drives yet cannot run a virtual machine, hardware-transcode 4K, or host a heavy Docker stack. A two-bay Intel box does all three. So if you are buying a NAS for what it enables rather than raw capacity, start from the app you care about and work back to the hardware, which is exactly what this matrix is for.

The ARM vs Intel fork

This is the one distinction that shapes everything. Intel and AMD (x86) models run a full container platform, host virtual machines and (on Intel chips with Quick Sync) hardware-transcode 4K video. ARM models (UGREEN’s DH line, QNAP’s TS-233) are cheaper, quieter and lower-power, but they cannot virtualize, cannot hardware-transcode, and run Docker only in a limited way; QNAP’s ARM box has no Container Station at all. Neither is “better”: ARM is right for storage, backups and direct-play media; Intel is right for a machine that doubles as a home server. Getting this wrong is the most expensive NAS mistake, because you cannot upgrade a CPU.

The RAM lever

After the CPU, memory decides how much you can stack. Light apps, Pi-hole, Vaultwarden, Home Assistant, run in a few hundred megabytes. But Immich’s machine-learning search wants 8 GB, a Minecraft world wants 4 GB to itself, and a virtual machine wants 16 GB of system RAM so it has a real slice without starving the NAS. Many entry models ship with 2–4 GB, so the question is not just what they ship with but how far you can upgrade, a model with SO-DIMM slots you can fill beats one with soldered memory the moment you get ambitious. Where a row below needs more RAM to run an app well, its model page links straight to the compatible modules.

The three AI columns

AI chat rates a private, local chatbot (Ollama with Open WebUI) running Llama 3.1 8B at the default Q4 with an 4k context (Ollama’s default), exactly what the AI NAS calculator opens with, so the two always agree. Speed here comes from memory bandwidth, not the NPU: most boxes that say Ltd fit the model but answer slowly, or need more RAM first. Photo AI is Immich’s face and object search, and whether the box can accelerate it (an Intel iGPU through OpenVINO, a Rockchip NPU through RKNN, an AMD GPU through ROCm) or leaves it to the CPU. Camera AI is Frigate’s object detection, which wants a supported detector rather than the CPU. Click a chat cell to open that model in the calculator.

Frequently asked questions