How to check a host before you pay one
Everybody on this market advertises dedicated cores, NVMe and instant deployment. Most of those claims are true in some narrow sense and misleading in a useful one. What follows is the method we would apply to ourselves, and then eight comparisons against the archetypes people actually weigh us against.
Five questions a specification table will not answer
Comparison tables are written by the party being compared. The questions below decide whether a provider is any good, and not one of them appears on a pricing page.
Price is the sixth question. It means something once the other five have answers.
Is a core a core?
Ask for the allocation ratio in writing. A host that runs dedicated cores knows its number and will give it to you; a host that runs burstable ones answers with a paragraph about fair scheduling. Both replies are informative.
What were the last three incidents?
An uptime percentage is a marketing figure until you can read the post-mortems underneath it. A status page with no red on it belongs either to an unusually quiet network or to an unusually busy communications policy.
Who can take the machine away, and how quickly?
The terms decide that, not the agent answering your ticket. Read the suspension triggers, the notice period, and what a payment dispute does to a running instance before you put anything on it.
What is kept about the account?
A real answer is a list of items with a duration beside each one. Anything shorter is a sentiment.
What does leaving cost?
Image export format, egress charges on the way out, whether a snapshot can be downloaded at all, and how real the refund window is. Migration is far easier to arrange while you are still happy.
Words doing more work than they appear to
None of the terms below are lies. Each has a technical meaning and a marketing meaning, and the gap between the two is where your money goes.
vCPU
Depending on the author, this means a whole physical core, one SMT thread, or a scheduling share of a thread. On oversold platforms a single core commonly backs eight to twenty tenants, so the figure you bought describes your position in a queue rather than a quantity of silicon.
NVMe
That is the interface, not the drive. Consumer NVMe bolted to a server board is still consumer NVMe: quick for a few seconds, then the write cache fills, latency climbs by an order of magnitude and stays there for as long as you keep writing.
Unmetered
Almost always metered somewhere, at a figure that lives in the acceptable use policy rather than next to the price. If the number is not written down it is not a limit; it is a discretion.
Up to 10 Gbit/s
The load-bearing words are the first two. Ten gigabits shared across a rack is a fair description and a poor basis for capacity planning, so ask the follow-up: is the port dedicated, and dedicated to what.
99.9% against 99.99%
One digit separates forty-three minutes of permitted monthly downtime from four. Read the credit clause immediately afterwards, because a credit you have to claim within 24 hours of an outage you slept through is decoration.
ECC memory
Present or absent, with nothing in between. Most desktop platforms cannot do it at all, so a plan advertising ECC on consumer silicon has an error somewhere in its description. Our Ryzen line does not have it and says so; the EPYC line runs registered ECC DDR5 because that platform was built for it.
Test it yourself, inside the refund window
The entry tier tells you about the platform, because it is the same platform. One evening of measurement is worth more than a month of reading review sites, most of which are affiliate pages with a rating widget bolted on.
All of this runs on the cheapest plan a host sells. That is what the cheapest plan is for.
- 01
Buy small, in the city you actually want
Order the entry plan at the site you would really use. Latency, route quality and stock levels differ by site far more than by plan, and the cheap instance is disposable once you are done with it.
- 02
Measure the idle machine first
Drive model, CPU model, port speed, route length and steal time, all read in the first five minutes. Do it before you install anything that could later be blamed for a number you do not like.
- 03
Then hold it at full load for ten minutes
Short benchmarks measure a burst credit balance and a boost window. Ten minutes on every core, watching the score settle, measures the machine you are renting.
- 04
Ask support one question with a right answer
Something specific and checkable: whether the IPv6 block is routed or proxied, what the fair-use figure is on your port, which kernel versions are supported. The response time and the accuracy are two separate results.
- 05
Diarise the last day of the window
A refund window is the only bargaining power you get for free, and it expires quietly. Decide on day five, not in the second month when the instance has become load-bearing.
| Claim on the page | Ten-minute test | A pass looks like |
|---|---|---|
| Dedicated vCPU | Busy-loop every core for ten minutes, then read st in top and steal in vmstat. | Steal time at zero throughout. A dedicated core that shows steal was never dedicated. |
| NVMe storage | fio, 4k random write, queue depth 32, direct I/O, five minutes without a pause. | Latency flat after the cache fills. A cliff at minute two means consumer drives. |
| 99.99% uptime | Ask for twelve months of incident history and the exact credit formula. | Public post-mortems, and credits applied automatically rather than on request. |
| 10 Gbit/s port | iperf3 to two unrelated off-net endpoints, plus one long single-stream download. | The rated speed under sustained load, not a burst that dies after the first gigabyte. |
| Unmetered traffic | Find the fair-use figure in the acceptable use policy. Search for the number. | A published figure per port, and a written description of what happens if you pass it. |
| DDoS protection included | Ask which layers are filtered, whether filtering is always in path, and when an address gets null-routed. | Always-on scrubbing with a stated threshold, rather than a button someone presses after twenty minutes. |
| No KYC | Search the terms for verification, identity, proof of address and at our discretion. | No document clause at any threshold, including refunds and large top-ups. |
| Instant deployment | Order the smallest plan and time it from settled payment to a login prompt. | Minutes. A queue labelled manual review is the tell, and it usually means a human is reading your order. |
| Enterprise hardware | Read /proc/cpuinfo and the NVMe model string, then look up the release year of both. | Current or previous generation silicon, and ECC where ECC was advertised. |
| Money-back guarantee | Read the exclusions before paying, not after. | A window measured in days, no reason required, and a stated route for a crypto refund. |
Run every one of these on us. If steal time is anything other than zero on a plan we sold with dedicated cores, that is a fault, and we would rather have the ticket than the benefit of the doubt.
The eight comparisons
Every page below argues against an archetype. Naming companies dates a page within a quarter and tempts everyone into quoting a rival’s worst plan against their own best one, which is a rigged demonstration that any reader can spot.
Each one includes a section on where the other side wins. That is the part worth reading.
Oversold burstable hosts
Where the core count is nominal and steal time is the specification nobody prints. Cheap, legitimate, and wrong for anything with a deadline.
Large public cloud
Superb engineering on a pricing model that punishes small deployments and egress. Also the place where your legal name is a prerequisite for compute.
Hosts that demand documents
Verification is usually the payment rail talking, not malice. It still leaves a scan of your passport in a ticket system for years.
Card-only hosts
A card carries a name, an address and a reversal right. Each of those changes how a host has to behave towards you.
The sub-five-euro tier
We sold one until 2021 and deleted it. The page explains what that decision cost, what it fixed, and when a cheap plan is still the correct purchase.
Renting whole machines
We sell both, so this comparison is easy to write honestly. Most people should take the virtual machine; those who should not usually already know why.
The low end of the offshore market
Offshore is a jurisdiction, not a spell. The cheap end of it is normally a resold instance on somebody else’s upstream account.
Your own rack
Cheaper per core once the machines are full, and only then. The page lists the cost lines that never make it into the spreadsheet.
About these pages
Two reasons, neither of them noble. Specifications and prices move faster than we revise pages, so a named comparison is stale within a quarter. The other reason is that vendor-written head-to-heads are usually rigged, and readers have learned to discount them entirely.
For our own purchasing, yes. We do not publish the numbers, because one instance of one provider measured on one night is an anecdote with a chart drawn around it, and we would be selecting the nights.
Price at the bottom of the market, managed data services, autoscaling, regional breadth against a hyperscaler, permissiveness against the far end of the offshore market, and cost per core against your own rack once it is full. There is a page for each.
Order the smallest plan, measure it properly, and take the refund if it disappoints. Seven days, no reason required, and crypto goes back in the asset you paid with.
Measure first. Decide second.
The entry Ryzen plan is €29 a month and runs on the same nodes as everything above it. Buy one, run the tests on this page, and keep it only if the numbers hold.