On-Premise vs Cloud: What a Company Server Costs over Five Years

The bill for your own server has six lines, of which a quote usually shows only the first. Here are all six with public prices — and the tax-saving row that is often taken from the accounts rather than from the tax computation that actually applies.

“Nobody has their own server any more, we keep everything in the cloud” and “your own server pays for itself in five years” are two sentences you can hear in the same week from two people who both work with computers and who are both convinced. The interesting part is that almost none of them have calculated either option all the way through, because the bill is six lines, of which a quote usually shows only the first.

This article works both sides of on-premise vs cloud with public prices and also names the line most often added up wrongly — the tax saving taken from accounts depreciation, which is not always the tax line in the company's own state. This is not about which hosting plan a website needs; that question is answered in a separate article on shared hosting, a VPS and a cloud server. This is about company systems, one physical machine, and the question of whether to buy it or rent it.

On-premise vs cloud: what we are actually comparing, and the six lines of the bill

The comparison is between a machine the company buys and puts on its own premises, and an equivalent machine it rents from a data centre, looking five years ahead in both cases. The five-year window is not chosen at random — it is roughly the time after which the server is no longer under warranty and starts to demand a decision, and it is also the time in which the rental payments can add up to the same order as the purchase price. A shorter window always looks favourable to rental, a longer one always looks favourable to buying, so any such calculation has to be read together with its period.

There are six lines, and only the first usually makes it into a presentation. The hardware price. Electricity. The room and its cooling. Software licences, if they are needed at all. A person's time to keep the system, update it and respond when something trips on a Friday evening. And finally the warranty, with what happens after it ends. On the rented side the first three lines are inside the price, the fourth usually is not, the fifth depends on how much of the work is inside the contract, and the sixth simply disappears, because the hardware is not yours.

Each line below has a real figure from a public source, and where there is no honest figure, that is said plainly. Two of the six lines cannot be closed with public sources, and that is not a flaw in the calculation but the most important thing this kind of comparison can teach — namely that the largest uncertainty sits in exactly those lines people usually fill in by feel.

One more thing worth putting on the table before the arithmetic is that the two sides are not the same product, even if the specification matches. Buying a machine, the company obtains property it can reconfigure, move or sell, and at the same time takes on everything that can happen to it. Renting, it obtains a service with a defined limit of responsibility and loses the ability to treat the hardware as it sees fit. That difference does not appear in the table, because it is not in money, yet it is exactly what decides most of the real decisions, which is why it is honest to name it before the figures.

The hardware price you see at once, which is not the figure in the advertisement

The Dell PowerEdge T360 is a typical single-processor tower machine that a small company would buy for its server room. In Dell's US store it starts at 5,599 dollars, but that configuration is one quad-core Xeon E-2414, one 32-gigabyte memory stick, one two-terabyte disk and — this is the important part — one year of basic support. In Dell's Ireland configurator a similar, but more realistically built, machine with two power supplies, a RAID controller and forty months of support costs €10,982.22 excluding VAT.

The gap between these two figures is not a discount, but two different machines, and this is exactly where most false comparisons are born. Extending support from the basic level to one in which a technician arrives the next business day costs another €658.57 in the same configurator, and a four-hour response €1,333.66, and the longest term offered is five years. HPE, in its own shop, shows the same class both at 7,017 and at 3,590 dollars in two different places on its site, which is a useful reminder: a “from” price on servers is not a price, but an invitation to open the configurator.

The practical sequence is therefore the reverse of the usual one: first you write down what the machine has to do and how long it has to stay under warranty, then that is built in the configurator with every line included, and only the figure you obtain is the one you may compare with a rental price. On the purchase side there is one more line nobody sees at this point — what happens in the sixth year, when the warranty has ended and a disk fails on a Saturday night.

The configurator also has several lines that move the price more than the processor, and those are usually chosen by habit. Two power supplies in a hot-swap chassis, in the same Dell configuration, cost €507.12 more than a simpler pair and €714.63 more than a single non-redundant unit; swapping the RAID controller for a simpler model takes off €1,982.02. None of these decisions is right or wrong in itself, but the redundant supplies together with the RAID controller are about 23 percent of the machine's price, and they are worth taking deliberately, not left as the configurator stacked them by default.

Electricity, where the band matters more than the unit price

Electricity prices for businesses are published by Eurostat, and there is one detail there that changes the result by more than the unit price: the prices are grouped by annual-consumption bands, and one server that draws fifty watts continuously uses about 0.44 megawatt-hours a year — that is, it falls into the very smallest band, not the one the news headlines write about. Those band prices are national. In the second half of last year, in the smallest band, excluding recoverable VAT, a business paid €0.0951 per kilowatt-hour in Norway, €0.1045 in Finland, €0.2450 in Latvia and €0.3258 in Germany; Eurostat published no United Kingdom value for that semester.

The gap between the smallest band and the middle band is large enough — in some countries almost twofold — to turn a well-meant calculation into a false one. If the figures are put in the right band, a server whose processor thermal design power is 55 watts costs about €46 a year in Norway, about €50 in Finland and about €157 in Germany, or on the order of €230 to €785 over five years. That is not much against the hardware price, and that is precisely why this line is worth calculating honestly rather than skipping: it is not the line that decides the choice, and a calculation in which it has been inflated loses credibility on the other lines as well.

Another error on this line is the power-supply rating: Dell's installation manual states outright that the thermal calculation is done on the power supply's rated wattage and that consumption should be checked with Dell Energy Smart Solution Advisor, so a 450-watt unit does not mean a 450-watt bill. Processor thermal design power in the same configurator ranges from 35 to 95 watts, and that is not consumption either, only the nearest public approximation the manufacturer gives.

A further reason to calculate this line yourself is that in Europe the prices differ by more than three times. In the same period, in the middle band, a business in Finland paid €0.0748 per kilowatt-hour, in Sweden €0.0970, but in Ireland €0.2552 and in Germany €0.2264, while the European Union average was €0.1837. That means the same calculation in Helsinki and in Frankfurt gives different answers, and that is exactly why data centres exist in the Nordic countries — but also that a comparison done in another country can simply be wrong in your situation.

Space and cooling, where the honest answer is “we do not know”

The cooling line is usually filled with a figure called PUE — the ratio of everything a data centre consumes to what the equipment itself consumes. In the Uptime Institute's annual survey, which last year included 681 data centres, that ratio averaged 1.54, and in facilities handed over in the last five years — 1.48. These figures are correct and useful, but they describe professionally built halls, not a cupboard at the end of the corridor with a split air-conditioning unit.

A cupboard with air-conditioning has no public PUE, and we will not invent one here, which in practice means two things. First, if the machine stands in an office, the cooling cost has to be measured yourself — the simplest way is to compare the electricity bill in summer and in winter. Second, the room has costs that are not electricity: space you would otherwise use, an uninterruptible power supply with batteries that themselves have to be replaced, and the fact that someone is responsible for the temperature even during holidays. These costs are not large, but they are not zero, and on the rented side they are not there at all.

There is also one space cost that cannot be expressed in money and is nevertheless real: noise and heat in an office where people work. Server fans under load are loud, and that is one of the commoner practical complaints in a small company that has put the machine next to desks. In a data centre this problem belongs to nobody, and at home people usually get used to it by finding it some kind of cupboard — which is then exactly the room that has neither a measured temperature nor a calculated cooling cost.

A person's time, which is on no price list

The fifth line is the one comparisons skip most often, because it has no public price: someone updates the operating system, checks that backups really restore, watches the disk's health and is reachable when something trips outside working hours. This work exists on both sides if the machine is unmanaged, and disappears only when it has been bought together with the server as a service.

It is honest to say that we did not find a public advertised hourly rate for a systems administrator in this market, and we will not invent one. What we can name is our own published rate for development work — €50 an hour — and here it serves only as a scale, not as a market price for administration. On that scale, two hours a month is €6,000 over five years, but five hours a month is already €15,000, that is, many times more than all the electricity and more than the built tower's price before licences.

It is exactly this line that most often turns the result upside down: on the purchase side it is invisible, because the work is done by a person who is already on the payroll, and so it appears in no table; on the rental side it is either the same invisible cost, or it is inside the price and therefore visible. A comparison in which one side shows this line and the other hides it is not a comparison — and that is the commonest way your own server looks cheaper in the table than it turns out in practice.

This line also includes one less obvious cost — replaceability: if one person knows the server and that person leaves or falls ill, then the knowledge of how the machine is put together leaves with them, and the next person starts with an investigation. The price of that outcome appears only once and usually in the worst possible week, so it cannot be calculated in advance; what can be done is to require that the configuration is written down and that restoring from backups is tested on a schedule, not from memory.

Licences are where your own server suddenly costs, or suddenly saves

If the system runs on Linux with PostgreSQL or MariaDB, there is no licence line at all, and the only cost is the person who keeps it. If the system requires Windows Server, the line appears and it is larger than it looks: you license by cores, the minimum is eight core licences per processor and sixteen per server even if there is only one processor, and on top of that each user or device needs a Client Access Licence (CAL). Microsoft's price list gives a recommended retail price of 1,176 dollars for sixteen core licences, but Dell, in the same configurator, sells the same product for €1,571.33.

The database licence is sharper still: SQL Server Standard edition on Microsoft's price list costs 3,945 dollars per two-core pack, and the minimum is four licences per physical processor, so the lowest possible bill is 7,890 dollars. In the physical-core model, Standard edition covers only the physical environment, not the virtual machines on it, and licensing by virtual environment with a minimum of four licences on each is a separate path, available only with Software Assurance or a subscription. The argument in favour of your own hardware comes from Windows Server, where a different rule applies: a sixteen-core licence assigned to the physical machine covers the host and two virtual machines, whereas four two-core cloud machines, each licensed separately, need thirty-two core licences, because the minimum there is eight per machine.

This is not a story and not a generalisation, but the manufacturer's licensing rule, which you can read in its own document. That is exactly why the licence line has to be calculated before everything else: if the answer is Linux, it is zero on both sides and decides nothing, but if the answer is Windows together with SQL Server, then licences can be the largest line in the whole five-year table and they can outweigh both the hardware price and the rental fee.

Client Access Licences are a separate line, easy to forget because it grows with the company. In the same Dell configurator one Windows Server user CAL costs €87.82, and a device CAL €68.88, so for a twenty-person office that is about €1,750 on top of the server and operating-system price. Windows Server requires these licences in addition to the core licences, not instead of them. SQL Server Standard edition also has a second model on Microsoft's price list — a server licence of 989 dollars plus a CAL of 230 dollars each, and against the four-core minimum of 7,890 dollars that is cheaper up to about thirty users. This threshold is not written anywhere as a recommendation; it is simple division you can do yourself.

The rental side with real prices, not a round figure

On the rental side the prices are public and you can put them on paper in the same minute. A Hetzner dedicated server AX42 with an eight-core processor, 64 gigabytes of memory and two NVMe disks costs €97.30 a month excluding tax, plus €1.70 for the address and €49 one-off setup, which over five years gives €5,989. OVH Advance-1 in an equivalent generation costs €159.99 a month with a setup fee of the same size, which over five years gives €9,759.39. Scaleway offers a similar machine at €89.99 a month.

Two notes belong with these figures, without which the comparison is not honest: the first is that these are unmanaged machines — the data centre guarantees that the hardware works and that the network is there, but the operating system is updated, the backups are checked and incidents are handled by your person. The second is that the cheapest list entries are not comparable machines — a five-euro server with two cores and four gigabytes of memory is an entirely different class, and putting it in the table against a ten-thousand-euro tower means comparing prices, not solutions.

Our own managed hosting and server rental starts from €45 a month, which over five years is €2,700, but that is a third product, not a cheaper version of the first two. That price also includes what, on an unmanaged machine, your person does: operating-system updates, security configuration, daily backups with thirty-day retention, monitoring around the clock and one named person. Comparing these three prices with one another, we are comparing a different amount of work, not different hardware.

On the rental side there are also two lines that sit in the small print of the price list and that can be decisive if the system sends a lot of data. The first is outbound traffic: on dedicated servers it is usually unlimited, but with a faster connection that is no longer so, and above a set volume it starts to cost per terabyte. The second is the setup fee, which with some providers equals a month's charge and this summer was raised. AWS's logic is different again — the first hundred gigabytes a month to the internet are free, and after that you pay per gigabyte, which means nothing to a website but means a great deal to data exchange between systems.

Where the tax argument does not work

The argument most often heard in favour of buying goes like this: the server is a capital asset, you write it off over five years, and that depreciation reduces profit and therefore tax. That is not automatically how the tax computation works, and the reason sits in the tax code rather than in the accounts: taxable profit is not the same as the profit in the accounts. Some systems tax distributed profit and have no depreciation deduction at company level at all; others add accounting depreciation back and grant a separate relief — tax depreciation, capital allowances, or whatever that state actually allows.

In practice that means buying a server is not itself a taxable event, but it also does not create a saving you can write into a five-year table from the accounts line. The tax line, if any, is a different calculation, with a different year-profile, and it is not the five-year straight line in the accounts. A table of rates copied from another country's income-tax rules, or from a sole trader's schedule, is not this company's computation, and a repealed schedule of rates is not a current one.

A five-year useful life for computer equipment does appear in accounting, as the period the company plans to use the asset, not as a tax rate. So if a quote or a presentation has a line “tax saving from depreciation”, it has to be asked more precisely: on the basis of which relief, and for which kind of taxpayer, in which state. If the answer is that state's tax depreciation or capital allowances, those are what you calculate; if the answer is the accounts depreciation line, that row is simply to be deleted from the table until someone has shown it is also the tax line.

From this it does not follow that buying is unfavourable — only that the advantage is elsewhere. A server is an investment in the business, and that is a cash-flow argument, not a tax saving from depreciation. The difference between those two formulations is that the first is true, while the second creates an expectation of a sum that the five-year table will not contain. In the accounts, depreciation is of course still calculated for the annual report; it simply does not reduce the tax payable in the way a slide titled “tax saving from depreciation” implies.

When your own server still wins

The first and clearest case is licensing, discussed above: if the load needs many cores and the software is licensed by them, then one physical machine, licensed once, can cost substantially less than the same capacity split across several virtual machines. This is not a rough consideration, but arithmetic on the manufacturer's published rules, and it has to be calculated before the platform is chosen, not after.

The second is a requirement about data residency, and here it is worth being precise. The GDPR restricts transfers of personal data out of the EEA, so a server in Germany or Finland meets it just as a server in Ireland or France does, and “the regulation requires a server in the UK” is not a correct sentence. A specific country can indeed be required by a contract or a tender specification, and then the source is that document, not the law in general. If the contract names a country, our own machines sit in Frankfurt, Helsinki or Riga, and renting in a named country means finding a provider with a data centre there.

The third is distance to the equipment: if the system drives a production line and response time is measured in milliseconds, then a machine on the shop floor next to the equipment is an engineering case, and we do not dispute that argument. It is also honest to say that we did not find a public document that would formulate this requirement numerically for the general case, so it is worth treating as an engineering consideration which the particular plant proves, not as a rule.

The fourth is a simple money case: if the server was bought two years ago and the warranty paid for five, then the remaining three years are already paid for and a move to rental simply throws them away. That is a valid cash-flow argument, only not a tax argument, and the two tend to be mixed up precisely because in the accounts they appear in the same table.

How to calculate your own figure, instead of taking someone else's

A calculation that decides something starts with the configuration, not with the price: write down how many cores, how much memory and how much disk the system actually needs today and how much it will need in three years, and only then look for prices — one in the manufacturer's configurator with every support line included, the other on a data-centre price list for the same specification. These two figures are the only ones that are directly comparable, and they usually sit closer to each other than expected.

Then add the other five lines on both sides, including the two that cannot be closed with a public source, and write down the assumption with which they were filled. If the result changes when one assumption is changed, then the decision is being made by that assumption, not by the server — and then it is worth testing the assumption, not choosing a platform. This is also the only place we recommend starting: not with the question “buy or rent”, but with the question “which line in this table is the largest”.

There is also one check that costs five minutes and guards against the largest error: take the finished table and ask in which line there is an assumption, not a price. If it turns out that the decisive line is a person's time or cooling, then the decision is not really about a server, but about how much work the company wants to keep in-house, and no price list answers that question. If the decisive line is licences, then the decision is about software, not about hardware, and that can be checked before anyone even chooses where the machine will stand.

If you have a configuration and need a short, comparable offer against it, with the work included and without hidden lines, write to us. If the decision has already been taken in favour of your own hardware, that changes nothing of the above — the six lines remain the same, only now they are all on your side of the table.