"SSD hosting" stopped being a differentiator years ago; now the question is which kind. The three storage technologies in datacenters today differ by factors of ten to a thousand on the numbers that matter for a website, and the spec-sheet word "SSD" hides which one you are getting.
The three technologies
HDD — spinning magnetic platters read by a moving arm. Every random read waits for the platter to rotate and the arm to seek: 5–10 milliseconds. Capacity is cheap (20 TB drives exist), which is why HDDs still run backups and archives.
SATA SSD — flash memory behind the same SATA interface designed for hard drives in 2003. No moving parts, so a random read takes about 0.1 ms, but the interface caps throughput at roughly 550 MB/s and handles one command queue of 32 operations.
NVMe SSD — flash memory connected directly to the CPU over PCIe, with a protocol (NVMe) designed for flash rather than inherited from disks. Latency around 0.02 ms, throughput of 3,000–7,000 MB/s, and 64,000 queues of 64,000 commands each — built for many things happening at once.
The numbers
| HDD | SATA SSD | NVMe SSD | |
|---|---|---|---|
| Random read latency | 5–10 ms | ~0.1 ms | ~0.02 ms |
| Random 4K IOPS | ~150 | ~90,000 | ~700,000+ |
| Sequential read | ~200 MB/s | ~550 MB/s | 3,000–7,000 MB/s |
| Cost per TB | lowest | mid | highest |
The throughput row is the one people quote, and it is the least relevant. Websites do not stream large files; they do thousands of tiny reads.
Why a website is the worst case for slow storage
Loading one WordPress page touches, roughly: 200 PHP files (cached by OPcache after the first hit, but still checked), 20–50 database queries each reading a few index and data pages, session and transient reads, and log writes. Each is a small random access. Multiply by concurrent visitors and a busy site issues thousands of random reads per second.
At 150 IOPS, an HDD serves one visitor's page load in a queue behind everyone else's. At 90,000, SATA SSD makes the disk stop being the bottleneck for most sites. At 700,000 with 5× lower latency, NVMe makes the disk effectively invisible — the database's query time becomes CPU and memory, not I/O.
You feel it most in three places:
- Database-heavy pages — search results, WooCommerce category pages with filters, admin dashboards.
- Uncached traffic spikes — when a page cache is cold, every request hits disk at once.
- Backups and migrations — copying a 20 GB site is minutes on NVMe, an hour on HDD.
Measure it yourself
On any Linux server, fio shows random-read IOPS in about a minute:
sudo apt install fio -y
fio --name=randread --ioengine=libaio --rw=randread --bs=4k \
--numjobs=4 --size=1G --runtime=30 --direct=1 --group_reportingLook at the IOPS= figure. Under 1,000 is a disk or a badly oversold SSD; tens of thousands is SATA; hundreds of thousands is NVMe with headroom. On a VPS the number also reflects how many neighbours share the drive, which is exactly why it is worth running.
A quicker sanity check for latency:
ioping -c 10 /var/wwwSub-0.1 ms is flash; multiple milliseconds is spinning disk or heavy contention.
When NVMe is not worth paying for
- Static sites behind a CDN. Almost nothing reads the disk.
- Archive and backup storage. Capacity matters, speed does not; HDD is the right tool, and it is what our offsite backup copies live on.
- A site that is slow for other reasons. If PHP is on version 7.4 with no OPcache, NVMe will not rescue it. Fix the stack first; speed up WordPress on Plesk is the order to do it in.
For everything else — a database-backed site with real traffic — NVMe is the biggest single hardware upgrade available, and the price gap to SATA SSD has all but closed.
What "NVMe hosting" should mean
Ask the host two questions: is the site's data on NVMe, or only the OS? And is it a RAID array or a single drive? Data on a single NVMe drive is fast until it fails. Every VPSPioneer shared plan and managed VPS stores customer data on mirrored NVMe with daily backups to separate storage; see RAID levels explained for why the mirror matters.