Disclosure: This post contains affiliate links. If you buy through them, we may earn a commission at no extra cost to you.
VPS pricing pages are designed to make comparison hard. The same “$5/mo” label can hide wildly different machines underneath. After years of running servers on all of them, here are the seven specs that actually determine whether a VPS is fast or a frustration — roughly in the order they tend to bite.
1. vCPU: shared vs. dedicated, and what “vCPU” even means
A vCPU is a thread, not a core. “2 vCPU” means two threads from a shared pool of physical cores — how much of a core you actually get depends on the neighbors.
Shared vCPU (the default on every $4–$20 plan) is burstable: fine for web, PHP, and MySQL workloads with spiky traffic. The thing to watch is steal time — run top and look at %st. Occasional steal is normal on shared hardware; sustained steal above ~10% means you’re on a crowded host or being throttled, and it’s time to move or upgrade.
Dedicated vCPU gives you full threads with no steal. It matters for CI runners, game servers, video encoding — anything that pins CPU for minutes at a time. Expect to pay 2–4x the shared price for the same nominal spec.
Rule of thumb: start shared. Move to dedicated when steal is chronic or you need predictable latency, not before.
Quick steal check on any new box: install stress-ng, run stress-ng --cpu $(nproc) --timeout 60s, and watch %st in top. On a healthy shared host it stays near zero even under your own load. If steal climbs while your workload is the only thing running, the host is oversold — pick a different region or provider, not a bigger plan on the same node.
2. RAM: the bottleneck that actually matters
On a typical LEMP/LAMP stack, RAM runs out long before CPU does. The InnoDB buffer pool, PHP-FPM workers, and the OS page cache all compete for it, and once the box starts swapping, response times fall off a cliff.
Practical minimums for a single-site LEMP stack: 1 GB is the floor (with page caching on), 2 GB is comfortable, 4 GB if MySQL and the app share the box. Ignore any plan that advertises swap as a feature — “includes 2 GB swap” is a confession that the RAM is undersized.
Watch it with free -m, and check dmesg for OOM-killer entries. One OOM kill is a warning; repeated ones mean you need a bigger plan, not more swap.
3. Storage: NVMe vs. “SSD” vs. network-attached
“SSD” on a pricing page can mean local SATA SSD, local NVMe, or network-attached block storage. These are not the same thing: network-attached volumes add latency and share IOPS with other tenants, which is exactly what you don’t want under a database.
Local NVMe is the gold standard for small VPS in 2026 — Hetzner’s CX line and Vultr’s High Frequency tier both ship it at budget prices. Community benchmarks (mnpfkx/vultr-cloud-benchmarks) measured roughly 118k 4k IOPS on Vultr’s High Frequency tier versus ~41–60k on standard tiers in the same price class: a 2x difference the pricing page will never tell you about.
When a provider won’t say whether the disk is local or what the IOPS look like, assume the worst and test it yourself with fio in the first hour — hourly billing makes this nearly free.
4. Bandwidth: the egress pricing trap
This is where the sticker price lies hardest. Most US providers include around 1 TB/month on their $5 plans (DigitalOcean, Vultr, Linode) and charge roughly $0.01/GB overage. Hetzner includes 20 TB on a €3.79 plan. Same internet, 20x difference in what’s included.
Do the math on your workload: 1 TB/month is about 3 Mbps sustained. A media-heavy site, a backup target, or anything serving downloads will blow past it. And “unlimited” bandwidth always comes with a fair-use policy — read it before you depend on it.
The standard mitigation is a CDN in front: caching static assets typically cuts origin egress by 80% or more, which usually makes the overage question moot.
5. Region and latency
Round-trips above ~100ms are perceptible, and TLS handshakes multiply every millisecond. Pick the region nearest your users, not nearest you.
Region count (Vultr’s 32 vs. Hetzner’s handful) only matters if your audience is global. EU-only audience? Take the cheaper box and don’t look back.
Test before you commit: spin up hourly-billed instances in two or three candidate regions, run mtr from your target networks, then destroy the losers. Total cost: cents. Total value: knowing instead of guessing.
6. Snapshots and backups
Snapshots are manual, full-disk, and fast to restore from — take one before every change, no exceptions. Automated backups typically cost ~20% of the instance price on DO/Vultr/Linode; Hetzner charges per-GB for snapshots, which works out cheaper for small boxes.
A snapshot in the same datacenter is a convenience, not a backup strategy. Keep an off-site copy of anything you can’t rebuild (database dumps to S3/B2 cost almost nothing).
And test your restore. An untested backup is a rumor.
7. Support quality
Below enterprise spend, support is tickets plus documentation on every provider — responses in hours, not minutes. The real differentiator is the docs: DigitalOcean’s tutorials remain the best in the industry and will save you more time than any support ticket, while Hetzner’s documentation is noticeably thinner.
Note what “managed” actually means: the provider patches the OS and the stack. It does not mean they debug your application. If you need someone to fix your WordPress, that’s a managed host or a sysadmin, not a $6 VPS with a “managed” label.
Three more things worth checking
IPv6
In 2026 there is no excuse for a provider without IPv6, and some budget tiers are IPv6-only to hit a price point (Vultr’s $2.50 plan, for instance). IPv6-only is fine for outbound workloads — scrapers, VPN exits, backup targets — but a non-starter for a public website unless you’re comfortable with NAT64/DNS64 gymnastics. Check what’s included before the price seduces you.
DDoS protection
Included L3/L4 mitigation is table stakes now — Hetzner, Vultr, and OVHcloud all bundle it. Confirm what “included” actually covers (usually volumetric attacks, not L7), and ask about null-routing policy: some providers null-route your IP for hours under attack, which is protection in roughly the sense that amputation prevents infection.
API and automation
If you ever want Terraform, Ansible, or a script managing infrastructure, the API is the product. Vultr, DigitalOcean, Linode, and Hetzner all ship solid REST APIs and Terraform providers — test the one-liners that matter to you (snapshot, rebuild, firewall rule) during the trial period. A provider whose API can’t do what its dashboard can will cost you an afternoon at the worst possible moment.
The 60-second checklist
Before you buy: shared vCPU is fine unless steal is chronic · 2 GB RAM for comfort · local NVMe if there’s a database · price the egress, not just the plan · region near your users, tested with mtr · snapshot before changes, off-site DB dumps · docs quality counts as a feature. Get those seven right and the provider logo barely matters. Revisit this checklist yearly — providers quietly change what “shared” and “SSD” mean.
Pricing and specs referenced above are mid-2026 figures and move around — check current pricing on the provider’s page before you buy. Benchmark numbers are cited from community sources linked inline, not our own testing.