Best External SSDs and NAS for Storing AI Models and Datasets
4 min read · Updated Aug 10, 2026
Once you are running more than two or three local models, storage becomes the bottleneck nobody warns you about. A single 70B model can be 40GB or more on disk, and a modest collection of quantized variants adds up to hundreds of gigabytes fast. A 2TB NVMe external SSD over Thunderbolt or USB4 is the right first upgrade for a single machine. A 4-bay NAS with at least 8TB usable is the right call once you are sharing models or datasets across more than one device.
Key takeaways
- Expect 4–40GB per quantized model variant, and most people end up with a dozen "just in case" downloads they forget about.
- Read/write speed matters at load time: an NVMe SSD over USB4 or Thunderbolt loads a 20GB model in seconds, a spinning hard drive can take over a minute.
- A NAS is overkill for a single-machine setup, buy one only when you need to share models or datasets across more than one machine, or want automated backup.
- RAID protects against a drive failure, not against accidentally deleting a dataset, keep a real backup regardless.
- Check the sustained write speed, not the "up to" marketing number, many external SSDs throttle hard after the first few gigabytes on a large model download.
Quick picks
| Storage | Capacity | Price tier | Best for |
|---|---|---|---|
| NVMe external SSD (e.g. Samsung T9-class) | 2TB | $150–180 | single-machine model storage, fastest load times |
| NVMe external SSD (e.g. Samsung T9-class) | 4TB | $300–350 | heavy model hoarders, still single machine |
| 2-bay NAS (e.g. Synology DS224+) | 8–16TB usable | $300 diskless + drives | small home lab, automated backup |
| 4-bay NAS (e.g. Synology DS923+) | 16–32TB usable | $550 diskless + drives | shared model and dataset storage across a household or small team |
| Internal NVMe upgrade (if a slot is free) | 2–4TB | $120–280 | cheapest per-GB option if you have the slot |
How fast storage actually needs to be for local inference
Storage speed matters when a model loads into memory, not while it is generating tokens, which runs entirely from RAM or VRAM once loaded. A slow drive makes the wait before your first prompt longer. It does not slow down generation itself. This trips people up constantly when they blame a slow SSD for slow inference and the real bottleneck is memory bandwidth (see the RAM upgrade guide for that half of the picture).
The opinion I will defend
A story: the day I ran out of disk mid-download
In April 2025, halfway through pulling a 70B quantized model overnight to test against an 8B for an extraction job, my laptop's internal drive filled up. I had forgotten I was still holding eleven older model variants from earlier experiments, together well over 300GB I had not touched in months. I deleted what I did not need, the download finished, and I started keeping models on a dedicated external SSD after that instead of the internal drive I also needed for everything else. Model hoarding is real, and it sneaks up faster than you would guess.
Backing up datasets: the part people skip
Model weights are free to re-download if a drive fails. A dataset you spent three weekends scraping and cleaning by hand is not. Follow the 3-2-1 rule for anything irreplaceable: three copies, on two different types of media, with one copy off-site or in the cloud. Treat model weights and datasets differently, they have very different replacement costs.
# Check how much space your local models are actually using
du -sh ~/.ollama/models“Model weights are free to re-download. The dataset you spent three weekends cleaning is not. Back up accordingly.”
Frequently asked questions
Frequently asked questions
Do I need a NAS for local LLMs?
Not for a single machine. A NAS earns its keep once you are sharing models or datasets across more than one device, or want automated backup you do not have to think about.
How much storage do local models actually use?
Expect 4–8GB for small quantized models up to 40GB or more for large quantized variants. A working collection of a dozen models easily reaches several hundred gigabytes.
Is USB4 fast enough, or do I need Thunderbolt?
USB4 and Thunderbolt 4 share the same underlying bandwidth ceiling on most current drives, either is fast enough for model loading. Thunderbolt matters more for daisy-chaining multiple devices or driving external displays.
Can I run models directly from an external drive?
Yes, an NVMe external SSD is fast enough to run models directly from without copying them to internal storage first, which is exactly why it is the right first purchase.
What RAID level should a home NAS use for AI datasets?
RAID 1 for a 2-bay NAS, RAID 5 or 6 for 4-bay or larger, depending on how many drive failures you want to tolerate at once. Remember RAID is not a backup on its own.