Public cloud, priced out line by line
Large clouds are remarkable engineering wrapped in a pricing model that makes small deployments expensive and large ones difficult to forecast. There is also the matter of who has to know your legal name before you are allowed to compute anything.
The bill has more lines than the machine
Nothing here is hidden. It is all documented, published and correct — and it is spread across enough meters that the monthly total surprises competent people every quarter.
Compute is the line everyone compares. It is rarely the line that grows.
Egress, per gigabyte, one direction
A moderately busy site shifting 20 TB a month can pay more for the traffic than for the server underneath it, at rates that have barely moved in a decade. Traffic here is unmetered under a fair-use figure we publish per port.
Storage, three times over
Capacity is one meter, provisioned throughput is another, and operations are a third. Our NVMe arrives attached to the plan, with no operations counter and no separate class of disk to choose wrongly.
Addresses billed by the hour
An idle public address usually costs money by the hour in a large cloud. Every instance we sell includes at least one IPv4 and a routed IPv6 /64, and the /64 costs nothing because it is not scarce.
Support as a percentage of spend
A response time with a number attached is generally a paid tier, priced against your bill. Our median first response is 11 minutes on every plan, including the €29 one, at every hour.
The parts that are genuinely better
Managed databases with automated failover, object storage spread across regions, queues and event buses that somebody else patches at three in the morning. Those are real products, they are good, and we do not sell them.
Identity is part of the price
Compute at that scale is a regulated financial relationship. It cannot be otherwise, and the consequences follow from the payment rail rather than from anybody’s intentions.
A card is an identity document with a chip in it
Opening an account requires a verified payment method in a legal name, an address that matches it, and often a phone number that receives a code. That set of facts is now permanently attached to whatever you run.
Automated risk scoring closes accounts
Fraud systems tuned for a population of millions occasionally take a project offline over a billing anomaly, and the appeal path is a form. The larger the platform, the less the process resembles a conversation.
Records outlive the workload
Billing history, access logs and support transcripts are retained for years for perfectly ordinary accounting reasons. Deleting the instance does not delete the account, and closing the account does not delete the records.
What we require, in full
An email address that can receive mail. Not a name, not a company, not a phone number, not a billing address, at no order value and at no point afterwards.
A like-for-like month
The right column is our actual catalogue price. The left is a shape rather than a figure, because published cloud rates differ by region and change more often than this page does.
| Line item | Large public cloud | Paragon |
|---|---|---|
| 8 dedicated cores, 32 GB RAM | A compute instance, on demand | R-8 at €54 a month |
| 400 GB of fast block storage | Billed by capacity, then by provisioned throughput | Included in the plan |
| 20 TB of egress | Per gigabyte, and often the largest line on the bill | Unmetered under a published fair-use figure |
| One static IPv4 address | Hourly charge for as long as it is allocated | Included |
| IPv6 | Available, region dependent | A routed /64 on every instance, at no cost |
| DDoS filtering | Base tier included, the meaningful tier is extra | Up to 12 Tbit/s, always in path, included |
| Support with a stated response time | A paid plan, often a share of spend | Included, 11 minute median |
| Identity required to buy | Legal name and a verified payment method | An email address |
| Time from payment to root | A minute or two | 47 seconds, median |
| Managed database with failover | Yes | No |
| Autoscaling to a thousand nodes | Yes | No |
The last two rows are the honest ones. If your architecture depends on either, the rest of this page is an interesting read rather than a recommendation.
When the hyperscaler is the right answer
Five situations where moving to us would be a downgrade dressed up as a saving.
You need the managed services, not the servers
A database that fails over without you, object storage with durability no single rack can offer, a queue nobody has to patch. Buying those from us means building and operating them yourself, and that is a full-time job you may not want.
Your load varies by three orders of magnitude
Ticketing, live events, tax deadlines, election nights. Paying by the second for two hours of enormous capacity beats renting the peak for a whole month, and it is not close.
The paperwork is the requirement
Where a customer contract demands specific audited certifications, a large provider holds them. We claim none, and we are not going to imply otherwise on a comparison page.
You need a region we do not have
Thirty-four sites across twenty-nine countries is a lot for an independent operator and modest against a hyperscaler footprint. When your users sit somewhere we are not, latency beats loyalty every time.
The team already fluently knows it
Institutional knowledge is a real asset. Rewriting working infrastructure to save a few hundred euro a month is usually a bad trade once you price the engineering weeks properly.
The arrangement most people settle on
This does not have to be a divorce. The split that works keeps the stateful, managed things where they are and moves the predictable, egress-heavy compute somewhere flat-rate.
Steady-state web and application servers, build farms, media processing, game servers, VPN and relay nodes and anything that moves a lot of bytes are the obvious candidates. They run at a fairly constant utilisation, they benefit from fast single-threaded cores, and per-gigabyte egress punishes them hardest.
What tends to stay behind is the managed data layer and anything with a compliance attachment. That is a reasonable outcome, and a much smaller bill than the one you started with. Nobody has to make a speech about repatriation.
If you want a sanity check on the split before committing, open a ticket with your traffic profile and core counts. We will tell you which parts we would not bother moving.
Questions
Yes, and plenty of people do, with a plain distribution installed on instances you control. There is no managed control plane, which means no per-cluster fee and no upgrade you did not schedule.
Yes, documented, covering provisioning, rebuild, snapshots and DNS. It is a small API, deliberately, and it does not change every quarter.
You get an email suggesting a dedicated 40 Gbit/s port at €45 a month, or a second instance in a second site. What you do not get is a retroactive per-gigabyte bill, because there is no meter to read.
We sign a one-page SLA with automatic credits, and we will answer a security questionnaire. We do not hold audited certifications, and if your procurement process requires one then the answer is no rather than a maybe.
Move the boring half of the bill
Eight dedicated Zen 5 cores, 32 GB of DDR5, 400 GB of NVMe and unmetered traffic for €54 a month. Try it with one service, compare a full billing cycle, and move the rest only if it wins.