Most teams refreshing storage in 2026 still default to spinning disks "because they're cheaper per terabyte." Per raw terabyte they still are. Per useful IOPS, per kilowatt, per rack U, per administrator-hour, per outage-minute — they have not been cheaper for years.

The math changed around 2021 for mixed workloads and around 2023 for archive. The mental model from 2015 — "SSD for performance, HDD for capacity, tier accordingly" — is now usually wrong.

Where the crossover actually happened

Three lines crossed at different times.

Cost per IOPS. A 7200rpm enterprise drive sustains roughly 150 random IOPS at full saturation. A mainstream enterprise SSD sustains 50,000+. At list prices today, the per-IOPS cost of SSD is roughly 70 times lower than spinning disk. Any workload with a random-IO requirement — virtualization, databases, mail, virtualized desktops, file shares serving more than a handful of users — pays the bill for spinning disk in either purchased IOPS or perceived performance.

Cost per usable terabyte after redundancy. A 22TB enterprise HDD lists around $400-500. Same-tier enterprise SSDs land near $100-130 per TB at the 4TB-8TB sweet spot. That looks like a 5× cost advantage for HDD on its face. It isn't, after you account for:

  • Redundancy overhead — a RAID-6 of 22TB HDDs needs spare capacity for a 22TB rebuild that takes 18-30 hours. RAID-6 of SSDs rebuilds in 2-4 hours, so smaller spare pools work.
  • Deduplication and compression — typical 2-3× reduction is realistic on mainstream all-flash arrays. Most SSD-targeted file systems and SAN OSes lean on this hard; HDD systems do less because random reads needed for dedup hashing are expensive on spindle.
  • Density — 24 NVMe drives in 1U vs. 12 HDDs in 2U at similar usable capacity post-dedup. Power and cooling savings compound across a 5-year refresh.

After dedup, redundancy, and density, the per-usable-TB gap closes to roughly 1.5× — not 5×. And for workloads that need any performance, the gap inverts.

Power and cooling. A 22TB HDD pulls 6-9W under load. A mainstream enterprise SSD pulls 8-12W under load, but idles at near zero where HDDs idle at 4-5W. In a typical 60%-idle datacenter array, the SSD pulls less power than the HDD over a 24-hour window. Multiply by hundreds of drives and the OPEX line moves.

Where HDDs still win

Three remaining workloads.

Pure cold archive — write-once, read-rarely, capacity-bound, no performance requirement. Compliance archives, video surveillance footage retention past 30 days, backup landing zones for short-cycle restore. The TCO math still favors HDD for these, but tape often beats both.

Sequential bulk — media production scratch space, scientific dataset staging, anything where you write a 4GB file once and stream-read it three times. HDD is fine here. The cost-per-IOPS argument doesn't apply because you have no random IOPS.

Capacity at the long tail — past 100TB of usable capacity in a single tier, the per-TB delta is real enough to matter and the workload doesn't require IOPS. Object storage for unstructured data sits here.

For everything else — and "everything else" is most of what most teams run — SSD has been the cheaper answer for years.

RAID levels for SSD have shifted too

The RAID-level conversation looks different on flash.

RAID 5 — historically discouraged on HDD because of the URE (unrecoverable read error) math during rebuild. A 22TB drive at a 10^14 URE rate has a meaningful chance of throwing a read error somewhere in a full-array rebuild, killing the rebuild. On SSDs at 10^16 URE rates the math changes by two orders of magnitude. RAID-5 on flash is no longer reckless.

RAID 6 — the safe default for HDD arrays above 1TB. On flash it's still defensible for very large arrays but starts to look conservative. The "lose two drives at once" probability is much lower on flash because failure modes are different — flash drives fail predictably from write exhaustion rather than randomly from mechanical stress.

RAID 10 — was the performance-required default. On flash, the IOPS gap that justified the 50% capacity overhead is gone; RAID-5 or RAID-6 on flash typically delivers acceptable IOPS for everything except write-extreme workloads.

The honest summary: on flash, RAID-5 is fine for everyday work, RAID-6 is the conservative default at scale, RAID-10 is no longer the obvious choice for performance.

The teams that get this wrong

Two patterns.

The 7-year refresh. Teams that refresh storage on a fixed 7-year cycle, replacing spindle with spindle because "it worked before." The previous refresh's math is not this refresh's math. Every refresh deserves a fresh TCO model.

The hybrid array compromise. Teams that buy a hybrid (mixed SSD + HDD) array as a "best of both worlds" risk-mitigation move. In 2018 this was reasonable. In 2026 it almost always means paying for two storage stacks (tiering software, cache management, capacity planning across two tiers) when one tier of all-flash would have been simpler and cheaper at the TCO level.

The exception worth keeping: a hybrid where the SSD tier handles the working set and the HDD tier handles a real cold archive. If you can clearly identify which data is which, hybrid still works. If you can't, you'll end up with both tiers half-full of the wrong data.

The honest counsel

The cost-per-TB conversation is the wrong starting point. Start from cost-per-useful-IOPS, then add redundancy overhead, then add dedup, then add power, then add the 5-year refresh frequency. For most enterprise and mid-market workloads — virtualization hosts, databases, mail servers, virtual desktops, file servers serving more than 25 users — the all-flash answer is cheaper before you account for any performance benefit.

The hardware caught up. The mental model has to catch up too.