Put two VPS plans side by side and the spec lines often look identical: 4 vCPU, 8 GB RAM, NVMe, 1 Gbit/s, unmetered. Then one of them runs your app smoothly and the other stalls every evening. The spec line isn't lying, exactly. It just leaves out the part that matters.
This is about comparing plans within the VPS category. If you're still deciding whether a VPS is the right kind of hosting at all, VPS vs shared hosting vs dedicated vs cloud is the place to start.
vCPU: a ceiling, not a guarantee
The word that hides the most. On most VPS platforms a vCPU is a slice of a physical core's time, scheduled alongside other customers' slices. The number in the plan is how many you can use at once during a burst. Whether you can hold all of them at 100% for hours is a different question, answered (if at all) in the fair-use policy.
We'll be specific about our own plans, because vague is how this goes wrong: the vCPU number is a peak ceiling for short bursts, not dedicated physical cores, and sustained full-load CPU may be throttled to keep things fair for everyone on the machine. For web apps, bots, agents and databases that's rarely noticeable. For hours of video encoding it is.
What to ask before buying: is the vCPU count guaranteed or burst, and what happens under sustained load? What to check after: the st (steal) column in vmstat. Steady double-digit steal means you're waiting for someone else's work.
RAM: the one honest number
Memory is the spec least open to interpretation. 8 GB is 8 GB. The only questions are whether the plan has enough for your workload and whether the provider oversells it, which you'll notice as unexplained swapping. If you're not sure how much you need, when you actually need 64 GB of RAM shows how to tell memory problems from disk and CPU problems.
NVMe: an interface, not a speed
NVMe is a storage interface, and it's much faster than the old SATA one. But the label says nothing about how many servers share the same drives, or whether each server has an I/O cap. A plan can say NVMe and deliver modest throughput under load. All our plans use NVMe; we'd still rather you test it than take our word for it.
Disk size matters more than people think. A database, logs and a couple of Docker images eat 15 GB faster than you'd guess, and plans only grow in place, so pick the disk you'll need in six months.
"1 Gbit/s": port speed, not throughput
This is the speed of the connection between your server and the switch. It isn't a promise that you'll always get a gigabit to anywhere. What you get depends on how busy the uplink is and on the path to the other end.
Our numbers, as a reference: Linux plans have a 1 Gbit/s port, Windows plans 150 Mbit/s. On a quiet line one Linux server gets about 940 Mbit/s each way; when several servers are busy at once, the line is split evenly between the ones actively sending. The details, including how to run a speed test that doesn't fool you, are in Traffic & speed.
"Unmetered": no quota, not no limits
Unmetered means nobody counts your terabytes and there's no slowdown after a monthly allowance. It doesn't mean unlimited speed: the port and the fair share above still apply. Read "unlimited bandwidth" with extra care. It usually means unmetered plus a fair-use clause, and the clause is where the real limit lives.
The lines that aren't on the spec sheet
Some of the biggest differences between plans aren't in the headline at all:
- IPv4 included or not. A plan without its own IPv4 can't host a website on IPv4. Here, Small at $8 is NAT and Small-IP at $16 is the same size with its own address; what one forwarded port can and can't do explains the difference in practice.
- Backups. Included, paid add-on, or your problem.
- Billing period. Annual plans here are 10% cheaper than paying monthly.
- Location. Ours are in Germany and Finland. Pick the one closest to your users or to the API your bot talks to.
- Resizing. Can you grow in place, and how long does it take? Can you shrink?
Comparing prices without fooling yourself
Dividing the price by gigabytes of RAM is tempting and mostly misleading. A plan that's cheap per gigabyte but has no IPv4, paid backups and a small disk can cost more than the "expensive" one once you add what you actually need. Build the same complete setup on each side, with the address, the backups and the disk you'll use, and compare the monthly total for that.
Check a new server in its first hour
Most providers let you try cheaply. Spend the first hour measuring instead of installing:
nproc && lscpu | grep 'Model name'
vmstat 1 30 # watch the st column
fio --name=rand --rw=randread --bs=4k --size=1G --runtime=30 --time_based --direct=1
iperf3 -c <a test server near you> -P 4 -t 20
Run the same commands at a busy hour in the evening. If the numbers fall apart, you've learned that cheaply.
Where this doesn't fit
Spec sheets can't tell you about support, about how a provider handles abuse reports, or about what happens when hardware fails. Benchmarks taken in one hour don't predict next month. And for some workloads none of this matters: a Telegram bot on Nano won't notice the difference between a good disk and a great one. Match the effort of comparing to the stakes. For a $3 bot, pick something reasonable and move on. For a production database, test.
Browse the VPS plans with the questions above in hand, and if you're switching from somewhere else, moving without downtime has the migration order.
Comments
No comments yet. Be the first.