Nothing between you and the board
No hypervisor of ours, no neighbour, no shared page cache, nothing arbitrating your interrupts. You get a management interface, a power button and the machine. What runs on it is between you and the silicon.
No hypervisor of ours, no noisy neighbour, no shared page cache. You get IPMI, a power button and the machine. What you run on it is between you and the silicon.
| Plan | Dedicated cores | Memory | NVMe storage | Port speed | Price | |
|---|---|---|---|---|---|---|
| BM-R 16C/32T · 2 × 2 TB NVMe | 16 | 128 GB | 4 TB | 10 Gbit/s | €239.00 /mo | Configure |
| BM-EMost taken 32C/64T · 4 × 3.84 TB NVMe | 32 | 256 GB | 15.36 TB | 25 Gbit/s | €529.00 /mo | Configure |
| BM-E2 Dual socket · 8 × 7.68 TB NVMe | 64 | 1024 GB | 61.44 TB | 40 Gbit/s | €1180.00 /mo | Configure |
Prices exclude VAT where applicable. Traffic: Unmetered, fair use.
Single tenant · IPMI over VPN · custom ISO on request · same-day rebuilds
What single tenancy is worth
Our virtual instances do not oversell anything, so the case for metal is not about getting the cores you paid for. It is about the resources that cannot be partitioned cleanly however honest the operator is.
The whole machine, including the parts nobody usually mentions.
Cache, memory bandwidth, PCIe root complexes, interrupt handling and the storage controller queue are all shared on a virtual host, no matter how carefully cores and gigabytes are divided. On metal they are not shared, because there is nobody to share them with. For a workload with a hard tail-latency requirement, that last few percent of variance is often the whole reason to be here.
Then there is control. Your kernel, your scheduler, your filesystem, your hypervisor, your partition table, your firmware settings. Nested virtualisation runs at native speed rather than as an emulation of an emulation. If you have ever read a tuning guide and found that the parameter you needed was one the host owns, this is the line that gives it back.
Your own hypervisor
Proxmox is a supported image and anything else can arrive as a custom ISO. Run your own virtual machines, resell them, nest them, break them at three in the morning; none of it is our business.
Firmware and boot
Access to the platform settings that matter, including memory and virtualisation options. If a change needs a hand at the console, the out-of-band interface gives you one over an authenticated tunnel.
Predictable tails
No steal time, no interrupt storm from a neighbour, no cache flushed by somebody else’s build. The ninety-ninth percentile stops being a mystery.
Storage laid out your way
Two to eight NVMe drives depending on machine, presented raw. Mirror them, stripe them, run them as separate devices, or hand them all to a filesystem with opinions. We do not impose a layout.
Who should not buy this
Metal costs more per unit of compute than a virtual instance and hands you every responsibility that comes with a physical machine. Both halves of that sentence are load-bearing.
Most people asking for metal want the EPYC line and one afternoon of tuning.
You want instant provisioning
Virtual instances take under a minute. A machine takes hours, occasionally a day when the configuration is unusual or the site is short of stock. Nothing about racking is fast, and any host claiming otherwise is handing you a virtual machine with a nice name.
You want to resize on Tuesday
Memory and disks are physical objects here. Growing means a new machine and a migration you plan yourself, not a reboot into a larger shape.
You do not have a redundancy plan
One machine is a single point of failure with no live migration behind it. If a four-hour outage would be a serious problem, you need two of something, and two virtual instances in different cities is usually the better version of that.
You are not comfortable at a console
Nobody else is administering this machine. There is no rescue hypervisor to snapshot you out of a bad kernel, only a serial console, an install-time ISO and your own competence.
You need it in fifteen cities
Nine sites carry metal. If your requirement is a footprint across the whole fleet, the Ryzen line reaches all thirty-four and costs less to spread out.
The three machines
We keep the range deliberately short. Three configurations we can hold in stock at nine sites beats twenty configurations that are always four days away.
One Ryzen build, one single-socket EPYC, one dual-socket monster.
The dual-socket machine is the one that needs a warning. Two sockets means two NUMA domains and a link between them, so an application that ignores topology can be slower on it than on the single-socket machine it replaced. Read the topology, pin deliberately, and it is the fastest thing on this page.
Every machine ships with more IPv4 addresses than the virtual lines carry, plus a routed IPv6 /64 at no cost. A /48 is available for anyone subnetting behind the machine, delegated rather than proxied.
| Machine | Silicon | Memory and disks | Bought for |
|---|---|---|---|
| BM-R | Ryzen 9950X, sixteen cores, thirty-two threads | 128 GB DDR5, two NVMe drives | Single-thread speed with nobody else on the box |
| BM-E | EPYC 9354P, thirty-two cores, sixty-four threads | 256 GB ECC DDR5, four NVMe drives | Your own hypervisor, or a large database with a latency budget |
| BM-E2 | Two EPYC 9354, sixty-four cores in total | One terabyte of ECC DDR5, eight NVMe drives | Dense virtualisation, analytics, anything memory-hungry |
Windows Server images are available on metal with a licence billed monthly. Everything else in the image catalogue is included, and a custom ISO costs nothing but the time it takes you to upload it.
Network, addresses and filtering
The port is yours and is not shared with the rack. Everything that protects a virtual instance protects a machine, because it happens at the edge rather than in a hypervisor.
Ten to forty gigabits, dedicated, with scrubbing permanently in path.
Dedicated port speed
Ten gigabits on the Ryzen build, twenty-five on the single-socket EPYC, forty on the dual-socket machine. Fair use is published per port speed and enforced by conversation.
Filtering you cannot forget to switch on
Up to twelve terabits per second of edge scrubbing, always in path. There is no attack-time button, because a button you press during an attack is a button you press too late.
Out-of-band management
The management interface is reachable over an authenticated tunnel and has no public route. It is included, not an upsell, and it works when the production network does not.
Your own address space
Announce your own IP space at sites where the upstream permits it. Send the object details in a ticket and expect an answer within a day, including a plain no where the site cannot support it.
Reverse DNS is yours
Editable from the panel for every address on the machine, without a ticket and without a justification for the first four.
Where the machines are
Amsterdam, both Frankfurt floors, Zurich, Bucharest, London, Paris, New York and Dallas. That is the list, and it grows slowly because a machine nobody rents is an expensive ornament.
Nine sites, seven countries. Metal needs floor space that stays sold.
Dallas is the easiest site to get metal into at short notice: cheap power, a large floor and stock that is usually there. Zurich is the hardest, for the obvious reason, and capacity is rationed rather than advertised. Frankfurt’s newer floor runs current-generation silicon only, with nothing legacy racked behind it.
Delivery is measured in hours rather than seconds. A machine in stock with a standard image is usually yours the same working day. A custom ISO, an unusual disk layout or a machine that needs to be built takes longer, and support will give you a real date rather than an optimistic one.
- 01
Ask about stock first
A one-line ticket naming the site and the machine gets you an honest availability answer, usually within the median eleven minutes. It costs nothing and saves the awkward version of this conversation.
- 02
Order and pay in coin
Same checkout as everything else. An email address, an invoice, thirty-odd assets through OxaPay, and no identity documents at any point in the process.
- 03
Choose the install
A standard image goes on unattended. A custom ISO is mounted as virtual media, and you drive the install yourself from the console.
- 04
Take the console credentials
Out-of-band access arrives with the machine. Test it on day one, while nothing is wrong, rather than at midnight during your first bad kernel upgrade.
- 05
Rebuild whenever you like
Same-day rebuilds are included. Wipe it, reinstall it, change hypervisor twice in a week; none of it costs extra and none of it requires an explanation.
When a machine fails
A virtual instance is rebuilt on standby hardware and comes back with its data. A physical machine is a physical machine: when a component dies, somebody has to walk to the rack and change it.
No live migration. This is the trade you accepted at checkout.
Failed drives, memory modules, power supplies and fans are replaced from spares held on site. Typical time from detection to a running machine is a few hours during the working day, longer overnight at the smaller sites. Where the fault is a board or a processor, the usual answer is to move your drives into an identical chassis, which is faster than diagnosing it while you wait.
Everything above assumes your data is on the drives and the drives survived. If they did not, the recovery path is whatever you built for yourself. That is the honest shape of single tenancy, and it is why off-node backup exists as an option on this line at all.
Spares are held locally
Drives, memory and power supplies for every machine we sell sit in the same building as the machine. Nothing waits on a courier.
Console access during the fault
The out-of-band interface stays reachable while the production network is down, which is generally how you find out whether the problem is yours or ours.
Off-node backup is not optional in spirit
Nightly, encrypted before it leaves, written to a different city. On metal it is the only layer between a dead array and starting again.
The SLA still applies
99.99% contractual monthly uptime with automatic credits. Credits do not restore data, which is a sentence worth reading twice before you buy a single machine.
What we will and will not do for you
The boundary on metal is sharper than on the virtual lines, so it is worth writing down rather than discovering during an incident.
We are an infrastructure company and we are not good at pretending otherwise.
We do
Rack it, cable it, power it, cool it, keep it on the network, filter attacks at the edge, replace failed components from local stock, mount your ISO, rebuild it on request the same day, and answer tickets with a median of eleven minutes.
We can, for a fee
Build and maintain a hardened operating system baseline, patch it, and tell you when it drifts. Add layer-seven filtering rules with a human on the other end during an attack. Take nightly encrypted backups to another city.
We do not
Administer your applications, debug your kernel, tune your database, hold your encryption keys, or log in to your machine at all unless you ask us to and tell us what for.
Questions about this line
Hours, not seconds. A machine in stock with a standard image usually lands the same working day. Unusual layouts and custom installs take longer, and support will give you a real estimate before you pay rather than after.
Yes. Run whatever hypervisor you like and sell what you build on it. Your customers are your customers; we have no relationship with them and no interest in one. You remain responsible for what happens on the machine under the acceptable use policy.
Out-of-band console over an authenticated tunnel, with no public route to the management interface. Serial and remote video, power control, and virtual media for mounting an ISO. It is included with every machine.
In cores and memory, roughly. In behaviour, no: two sockets means two NUMA domains and a link between them. An application that pins threads sensibly does very well on it, and one that assumes a flat machine can be slower than on the single-socket build.
No. Every drive in the machine is ours, monitored by us, and replaced by us when it fails. Bringing hardware into a shared facility creates problems for everybody, including the customers in the next rack.
Ask. More memory, a different disk count or a specific network arrangement is sometimes possible and sometimes not, and you will get a straight answer with a lead time rather than a maybe.
Ask about stock before you order
One ticket naming the site and the machine gets an honest availability answer, usually inside eleven minutes. Payment is in coin, sign-up is an email address, and rebuilds are included for as long as you keep the machine.