A few sysadmins I've talked to this year can't agree on whether RAID is worth setting up on NVMe drives anymore. Some swear by it. Others have quietly replaced it with software-defined storage and don't miss it. Both are right, depending on what they're running. RAID hasn't gone anywhere in 2026, but the way people use it has changed enough that the old advice doesn't always apply.
What RAID actually does
What is RAID? At its simplest, it combines several physical drives into one logical unit using striping, mirroring, or parity, so you get speed, fault tolerance, or some mix of both. RAID 1 mirrors data across drives. RAID 5 and 6 use parity so the array survives one or two drive failures. RAID 10 stripes mirrored pairs together for a middle ground between speed and safety.
IBM (News - Alert) researchers described this scheme in 1987. The math hasn't changed. Everything else has.
Old controllers, much faster drives
Hardware RAID cards were built for spinning disks that maxed out around 200 MB/s. A single NVMe drive can push several GB/s now, and a lot of legacy controllers can't calculate parity fast enough to keep up with four or five of them running at once. The RAID layer used to be the fast part of the stack. Now it's often the bottleneck.
Teams dealing with this tend to move to software RAID, where parity math runs on the CPU or a dedicated accelerator instead of an aging controller chip. NVMe-oF and zoned namespace (ZNS) drives are also showing up paired with RAID logic built specifically for flash. It cuts down on write amplification and gets more usable life out of each drive.
Why this matters for AI workloads
A GPU sitting idle while it waits on a data load is money burning for no reason. RAID 10 still shows up a lot on read-heavy training data since it skips parity math entirely. RAID 6 tends to protect the bigger, slower archive, things like model checkpoints and raw datasets, where losing two drives mid-training would wipe out days of compute.
Where Ceph and ZFS fit in
Distributed systems like Ceph and ZFS spread redundancy across nodes instead of drives in a single box, and at hyperscale that often replaces RAID outright. Most people aren't running hyperscale clusters, though. For a single server, an edge site, or a workstation doing local AI training, RAID is still simpler to set up and maintain than a distributed storage layer, and simple wins when someone has to fix it at 2am.
So is it still worth it?
Depends what you're running. A handful of NVMe drives under a training rig probably does better with RAID 10 or a software RAID setup tuned for flash. A fleet of nodes has likely already moved past RAID into something distributed. Both answers are fine. For a lot of setups, plain old RAID is still the one that makes sense.
