Hosting spec sheets blur three very different storage tiers behind the word “SSD”. At the bottom, spinning HDDs still lurk in budget offshore tiers — fine for cold archives, miserable for anything interactive. In the middle, SATA SSDs: flash chips bottlenecked by a ~550 MB/s interface and a command protocol (AHCI) designed in the optical-drive era, with a single queue 32 commands deep. At the top, NVMe: the same flash idea attached straight to PCIe lanes, speaking a protocol built for parallel silicon — up to 64K queues, each 64K deep, with throughput from ~3.5 GB/s (Gen3) to 7+ GB/s (Gen4) and access latency measured in tens of microseconds. For a multi-tenant VPS node, that gap is not a benchmark vanity metric; it decides whether your database commits stall when a neighbour gets busy.
The numbers that matter
Three figures tell the story. Sequential throughput: SATA SSDs cap around 550 MB/s; NVMe Gen3 reaches ~3,500 MB/s and Gen4 ~7,000 MB/s — a 6–13× gap. IOPS: the random 4K reads and writes that real workloads are made of run ~90–100K on good SATA but 500K–1M+ on NVMe, because the protocol can actually exploit flash parallelism. Latency: AHCI’s overhead puts SATA round-trips near 100µs; NVMe cuts the protocol cost to ~10µs. Databases, mail queues and package managers live and die by the last two numbers, not the first.
Where you feel it on a VPS
- Databases. MySQL/PostgreSQL commit latency tracks storage latency directly; under concurrency the IOPS ceiling becomes the requests-per-second ceiling.
- Web applications. Cold caches, composer/npm installs, log writes and session storage all hammer random I/O — the difference between a snappy and a sluggish admin panel.
- Mail servers. Queues are fsync-heavy by design; slow storage shows up as delivery lag — relevant for a mail VPS.
- Game servers. World saves and chunk loads stutter on slow disks; see game server VPS.
- The neighbour problem. On shared nodes, total IOPS is a pie. A SATA pie is small enough that one noisy tenant starves the table; an NVMe pie is an order of magnitude larger before contention bites.
Reading hosting spec sheets honestly
Marketing exploits the ambiguity hard, especially offshore. The patterns to decode: “SSD storage” with no interface named is almost always SATA; “NVMe-cached” means HDDs with a flash cache in front — fine until your working set exceeds the cache; tiny entry plans on “SSD” with NVMe reserved for premium lines is the classic offshore pattern (the established anti-DMCA brands still ship SATA or even HDD at entry, which our cheapest-offshore comparison footnotes per provider). And RAID matters: flash fails like any component, so look for RAID10 or equivalent redundancy behind the acronym — performance with no redundancy is a time bomb.
Does NVMe cost more?
It used to; the premium has mostly evaporated as Gen3 drives became commodity parts, and what remains is a choice about margins. We run NVMe RAID10 on every tier including the $3.50 entry plan — there is no SATA line to upsell from. When a host still charges a meaningful NVMe premium in 2026, you are usually paying for their procurement inertia, not for the silicon.
