The rack you would have to run
Owning the hardware is cheaper per core than anything we sell. It is cheaper on the condition that your time costs nothing, the machines stay full, and nothing important fails at an inconvenient hour. Two of those three are usually false.
The lines nobody puts in the spreadsheet
A comparison that stops at capital expenditure against monthly rent always favours ownership, because it omits most of the operating cost and all of the labour.
The hardware quote is the part everyone gets right.
Space and power, billed whether or not you use them
Rack space is sold by the unit and power by the circuit, contracted in advance. An idle machine costs the same as a busy one, so utilisation is your problem from the first day.
Transit, with a commit and a percentile
Bandwidth is bought as a monthly commitment plus overage measured at the ninety-fifth percentile. Getting that number wrong in either direction is expensive, and you sign for it before you know your traffic shape.
Cross-connects and address space
Every physical link is a monthly line item, and announcing your own addresses means registry paperwork, upstream authorisation and somebody who understands routing on call.
Filtering you cannot buy at small scale
Serious scrubbing capacity is priced for people who buy it in terabits. Below a certain size the honest options are an expensive third-party service or accepting that a large attack takes you off the internet.
Spares, and the hands to fit them
Drives, power supplies, memory and a spare board, sitting in a cupboard depreciating. Remote hands are billed in quarter hours and are not always awake when your machine is not.
Your time, which is the real cost
Firmware updates, a failed disk on a Sunday, a switch that needs replacing during a change freeze. None of it appears in a cost model, and all of it appears in your calendar.
What ownership genuinely gives you
This is not a straw man. There are properties of owned hardware that no rental arrangement reproduces, and for some organisations they decide the question.
Cost per core at scale
Once machines run at high, steady utilisation for three to five years, ownership wins on price and it wins clearly. Nobody who has done the arithmetic disputes this.
Total control of the stack
Firmware versions, drive models, network topology, physical access. If a security requirement reaches down to the board, only ownership satisfies it.
No per-gigabyte anything
Traffic within your own racks and across your own links is free, forever, at any volume. For petabyte-scale internal movement that is decisive.
Hardware you can specify exactly
Odd accelerators, tape, unusual drive counts, a specific NIC. We stock a catalogue; you can buy whatever exists.
An asset rather than a subscription
The machines are worth something at the end, and depreciation behaves differently to rent on a balance sheet. Your finance department may care about this more than your engineers do.
One 2U node, both ways
The right column is what a single rented machine includes. The left is the set of obligations that arrive with the equivalent board in your own rack.
| Line | Your own rack | Rented here |
|---|---|---|
| Hardware | Capital cost, depreciated over three to five years | None |
| Space and power | Monthly, per circuit, idle or not | Included |
| Transit | A commit plus ninety-fifth percentile overage | Unmetered under a published fair-use figure |
| Cross-connects | Per link, per month | Not your concern |
| DDoS filtering | Effectively unaffordable below a certain scale | Up to 12 Tbit/s, always in path |
| Spare drives | You stock them and you fit them | We stock them, and there is a hot spare per node |
| A dead power supply at 03:00 | Remote hands, billed, if somebody is awake | A ticket, 11 minute median response |
| A second city | A second contract, a second shipment, a second trip | A different site code at checkout |
| Cost per core at full utilisation | Lower | Higher |
| Cost per core at twenty percent utilisation | Much higher | Identical to full utilisation |
The last two rows contain the entire argument. Ownership is a bet on utilisation, and rented capacity is insurance against being wrong about it.
When your own rack is the better choice
Large, steady, predictable load
Dozens of machines at high utilisation, for years, with a growth curve you can forecast. That is the shape ownership was made for, and renting it is straightforwardly more expensive.
You already have the facility and the staff
If the room, the power and the people exist, the marginal cost of another node is small. The comparison changes completely once the fixed costs are already sunk.
Custody is a hard requirement
Some obligations demand physical control of the media, with a named person able to walk to the machine. No rental arrangement satisfies that, including ours.
Storage where drives dominate the bill
At the petabyte scale the drives are most of the cost and everything else is rounding. Our S-100 is €449 a month for 100 TB, which is excellent for one node and poor arithmetic for forty.
Hardware we do not stock
Specialist accelerators, tape libraries, particular network cards. We sell a catalogue rather than a purchasing department, and the catalogue is deliberately narrow.
The arrangement that usually wins
Most organisations that examine this properly end up with both, split along a line that is easy to describe.
Owned hardware carries the steady base load in one or two places, where utilisation is high and predictable and the capital is already committed. Rented capacity handles everything at the edges: the second and third region, disaster recovery, seasonal peaks, the experiment that might be deleted in March, and anything that needs to exist in a country where you will never sign a rack contract.
Thirty-four sites at a checkout is the part that is genuinely hard to reproduce. Reaching twenty-nine countries with your own equipment means twenty-nine contracts, twenty-nine shipments and twenty-nine relationships to maintain, and that is before anybody has served a request.
Fleet pricing exists above a certain number of machines. It does not beat ownership at very large scale, and we are not going to claim it does.
Questions
No. We do not run a colocation product, and every machine in the fleet is one we specified, bought and can replace from stock. Mixed ownership breaks that, so we decline it rather than do it badly.
Yes, on bare metal and on EPYC plans from the E-32 upward, at sites where the upstream permits it. Open a ticket with your object details and you will have an answer inside a day.
Then colocation is the product you want, and we are not it. That is a genuine requirement with a genuine market, and pretending otherwise would waste a month of your time.
Above a certain number of machines, yes, and it is negotiated rather than published. It is worth doing the ownership arithmetic first, because past a few dozen full machines the answer may not be us.
Rent the edges, own the middle
Thirty-four cities, no contracts, no shipments, no cross-connect invoices. Start with one machine in a city where you would never sign a rack lease and see what the split costs you.