Managed GPU infrastructure · Amsterdam · Frankfurt · Singapore

The compute is yours.
The data-centre layer is ours.

We run GPU racks for our customers in Europe and Asia-Pacific: power, cooling, uplink, spares and 7 × 24 on-site staffing. Monthly uptime goes into the contract, and missing it credits you automatically. The hardware can be bought, rented by the GPU-hour, or already yours.

Typical rack section · schematic

Your compute— GPU nodes you bought or rent by the term.

The layer we run— space, A/B power, cold aisle, uplink, spares, staffing.

Rack-ready
5 – 10 days
On-site staffing
7 × 24
Part swap
Spare on site +4 h
Contracted uptime
≥ 99.9%
Zones
AMS · FRA · SIN

The layer we take over

What we operate

When one of these six fails, your training should not stop and your people should not be woken up.

PWR

Space and power

A/B utility feeds, UPS and diesel behind them, 12 – 15 kW per rack. Minutes of power loss are reported monthly at rack PDU level.

COOL

Cooling

Contained cold aisle, intake at 22 ± 3 °C and 40 – 60% RH, temperature curve included in the monthly report. We own the intake side; temperatures inside the chassis follow your load profile.

NET

Network

2 × 100 G uplink, direct peering with AMS-IX / DE-CIX / SGIX, BGP with your own IP range. We commit to the rack-to-uplink-port segment.

OPS

Hardware on call

Someone is on site 7 × 24. A dropped GPU, a stopped fan or a tripped PDU is not your 3 a.m. problem: fitted and power-on tested within 4 hours of the spare arriving. Spare arrival time depends on supply and is written into the contract annex.

ASSET

Racking and asset records

Unboxing inspection photographed, cabling, asset labels applied. The delivery list is a CSV you can import straight into your CMDB: serial number, warranty start and end, burn-in report link.

RPT

Monthly report and credits

One per month: uptime, minutes of power loss, intake temperature, ticket response times. Below the contracted figure the credit is automatic and deducted from the following month's invoice — you do not file a claim. Credits are capped at that month's hosting fee for that rack; the measurement basis and the exclusions are in the terms.

What one node looks like, and the seven places we watch

Inside one node · orthographic

Front and top views of a 4 U GPU node. Seven positions are marked: which layer is your asset, and the six we watch every day. Model and configuration are in the plan we send you.

Fig. 01 — 4 U GPU node · front & top

04010203482.6 mm · 19" · 4 U060507Top view · front intake → rear exhaustPSU APSU B
Orthographic drawing at 19″ rack dimensions. It shows construction, not a render.
  1. Front drive baysSystem and local cache drives, hot-swappable. Swapping a drive does not stop the node.
  2. Perforated intake faceFaces the cold aisle, intake at 22 ± 3 °C. Open area decides whether this node can absorb a full 15 kW rack.
  3. Status and identityOn-site staff locate by light. The dashed box beside it is the asset label position: serial number and warranty dates go there.
  4. Rack ears · 19″Standard hole pattern. Racking changes nothing structurally, including for hardware you bring in.
  5. GPU modules × 8The yellow layer is yours. High-speed inter-GPU links, 72 hours at full load before it leaves the factory; anything short of spec is not racked.
  6. Fan wallA stopped fan raises a ticket on its own — you do not have to notice. Fitted and power-on tested within 4 hours of the spare arriving.
  7. Dual power modulesOne A feed and one B feed. Losing either does not stop the node. Minutes of power loss reported monthly at PDU level.

On-demand cloud, your own build, or us

Three ways to get compute

All three work. The dividing lines are term and headcount: a few days, take the first; a facilities team, take the second; otherwise talk to us.

On-demand cloud

Someone else's machine, rented by the second

Billed by the second, fastest to start, right for an experiment that runs a few days. The trade-off: the machine is not yours, popular configurations queue, and you do not control neighbours or oversubscription.

Above roughly three months of continuous running, usually the highest unit cost (measured against published on-demand rates and our range pricing; your own discount and utilisation decide the real figure)
Build your own

You carry all of it

Maximum control. The trade-off: rack contracts, power upgrades, cooling work, spares inventory and 7 × 24 staffing are all mandatory, and it usually takes 6 – 12 months before anything runs.

On top of the hardware you absorb facility, power and headcount cost
Hand it to usUs

The compute is yours, the operations are ours

The machine can be an asset you bought or one you rent by the term. It runs in our racks, with uptime, swap response and credit percentages written into the contract. Racked 5 – 10 days from signature and hardware arrival.

Billed per rack · month, power metered separately, 30 days' notice to exit

Three ways to work with us

Buy · rent · host

Purchase · SupplyRental · RentalsHosting · Care
Entry scale and priceFrom one complete machine, priced by configuration; the range narrows with volumeFrom 8 GPUs · 24 hours, from $0.95 / GPU-hourFrom one rack, from $1,650 / rack · month, power metered separately
Minimum termOne-off, no term24 hours3 months
Who operates itYou. Inside warranty we swap parts; we do not come to siteYou. During the term we replace the machine rather than repair itUs. Power, cooling, network, part swaps, 7 × 24 staffing
Whose books it sits onYour fixed asset; the 3-year warranty follows the serial numberOur asset. Normal wear is on us; loss or damage while at your site is insured by youOwnership does not change — what you buy is the operations layer
How you exitThe asset is yours. We can assess and take it back at the prevailing market price14 days' written notice, 14-day return window30 days' written notice; we de-rack and pack, you pay the return freight
Credit capPart swap or depot repair; if the model is discontinued, refunded at residual valueDowntime deducted hourly from the rent, capped at the rentCredited on monthly uptime, capped at that month's hosting fee for that rack
Who should not pick thisTeams with under 12 months of demand, or still undecided on a GPU generation — renting costs lessTeams running flat out beyond 9 months — by then the rent exceeds the purchase price; buyTeams needing physical access at will, or whose data cannot leave their own facility

The three routes mix: rent for a month to validate, then buy for the long run — usually without moving the hardware out of the rack. Whether the same units carry over depends on what is racked and their condition at the time, and the plan says so. Rent is a usage fee and does not offset a purchase price. Prices in the table are indicative ranges and are not an offer.

When to come to us, and when not to

Who this is for

✓ In these cases our route costs less

  • The workload runs for more than three months and the on-demand bill no longer adds up
  • You want compute on the balance sheet as an asset, not as a floating monthly cloud line
  • Nobody on the team is willing or free to handle dropped GPUs, fans and PDUs at night
  • Data has to stay inside the EU, or you need Europe and Asia-Pacific covered at once
  • You already bought the machines and what you lack is space, power and on-call staffing
  • You need an asset register and uptime report you can hand to finance and audit

✕ In these cases, do not come to us

  • A few days of experiments — on-demand cloud is simpler; our rental minimum is 24 hours and hosting is 3 months
  • Hands on the hardware whenever you like — site access needs 3 business days' notice and an engineer escorts you
  • Data that must never leave your own facility — then buy the machines and keep them there; we only supply
  • You want us to own the OS, drivers, CUDA and cluster scheduling — that is your boundary
  • You want per-second elasticity — we deliver by the term and do not run a burst pool
  • You want us to underwrite business losses from an interrupted training run — credits are capped at the period fee and exclude indirect loss
  • The end user or end use would not pass export-control review — this is not something we help anyone route around

Setting the boundary before signing is much cheaper than setting it after something breaks.

From one sentence of demand to a monthly report, in four steps

Delivery · day 0 → monthly

  1. 01 · Day 0

    Describe the workload, get a plan

    You tell us what you need to run, for how long, and where. We come back with scale, configuration, unit price and a rack-ready date. Valid 7 days; scale is locked against that day's free capacity.

  2. 02 · Days 1 – 3

    72 hours at full load

    Temperature curves, memory ECC error counts, measured inter-GPU link bandwidth — each one logged. A machine that misses any of them is not racked and does not ship.

  3. 03 · Day 4

    Racked here, or shipped to your site

    In our racks: cabling, asset labels, unboxing and inspection photographed and archived. Shipped to you: install manual in the box, power on and go. The cross-border leg runs at the pace of customs and export licensing and is not counted in the lead time.

  4. 04 · Every month

    One operations report

    Uptime, minutes of power loss, intake temperature curve, ticket response times. Miss the contracted figure and the credit is automatic, deducted from the following month's invoice.

Signature → racked · last 12 deliveries

7.0 days average

Scheduling limit 10 days128400112
One deliverySlowest this periodSelf-imposed scheduling limit

"Racked in 5 – 10 days" — here are the last 12, drawn out.

Each bar is one completed delivery, counted from the day the contract was signed and the hardware arrived to the day the machine was running in the rack: 5 days at the fastest, 10 at the slowest, held up at customs. The dashed line is the scheduling limit we set ourselves; going over it is explained in the report. It is not a separate credit clause.

Basis: hosting deliveries into our own racks only; customs, export licensing and waiting on your facility are excluded. Past deliveries do not predict future scheduling.

Read how credits are written

The racks are in three places

Availability zones · AMS · FRA · SIN

We hold the rack contracts and set the on-call rota ourselves; nothing is subcontracted. Data staying in the EU goes to AMS or FRA, Asia-Pacific coverage goes to SIN, and one contract can be split across zones. The monthly uptime below is a measured 12-month average; the contracted figure is ≥ 99.9% throughout.

AMS-02Running

Amsterdam · Netherlands

Power / rack15 kW
CoolingContained cold aisle · air-cooled
Uplink2 × 100 G uplink · AMS-IX peering
Monthly uptime · measured99.98%
FRA-01Running

Frankfurt · Germany

Power / rack15 kW
CoolingContained cold aisle · air-cooled
Uplink2 × 100 G uplink · DE-CIX peering
Monthly uptime · measured99.98%
SIN-01Running

Singapore · Singapore

Power / rack12 kW
CoolingContained cold aisle · air-cooled
Uplink2 × 100 G uplink · SGIX peering
Monthly uptime · measured99.95%

Every machine runs 72 hours at full load before it is racked

Burn-in · 72 h · serial-tracked

A machine short of spec does not enter a rack. Reports are archived against the serial number for 5 years and can be re-issued any time.

GPU temperature · 72 h at full load

71 °C peak

Threshold 85 °C10075502500h24h48h72h
GPU temperature in chassisPeakDo-not-rack threshold
Burn-in report · extract (example)Passed
Serial number · exampleSN 4C81-A0937
Test duration72 h 00 m
GPU peak temperature71 °C (threshold 85)
Memory peak temperature78 °C (threshold 95)
ECC correctable errors0
ECC uncorrectable errors0
Measured inter-GPU bandwidth872 GB/s (rated 900)
8-GPU all-reduce184 GB/s
GPU drop events0
Format and thresholds follow this example; the readings are whatever your batch actually measures.
From burn-in bench to dispatch carton, one serial number
The one on the burn-in bench45° side light · nameplate and fins in frame · 5200–5600 K
The last step before dispatchLight ground, overhead · anti-static bag and packing list in frame · asset label legible

One goes on every machine that leaves. Serial number, SKU and burn-in pass date are printed on it — the same number as the line in the delivery list and the archived report.

The seven questions we get asked

All terms
What exactly do you provide, and how is it different from on-demand cloud GPUs?
Three things: machines, space, operations. On-demand cloud sells you a slice of time on someone else's machine — you can stop any time, but past three months of continuous running the unit cost is usually the highest available, and you control neither quota nor neighbours. Here the machine is either an asset you bought or a complete system rented to you by the term, running in a rack in Amsterdam, Frankfurt or Singapore. Power, cooling, uplink, spares and 7 × 24 staffing are ours, with uptime and swap response written into the contract.
What do you commit to, and what do you not?
Three commitments are in the terms: monthly uptime (below 99.9% credits 10% of that month's hosting fee, below 99.5% credits 30%, below 99.0% credits 100%), a part fitted and power-on tested within 4 hours of the spare arriving, and one monthly report carrying minutes of power loss and the intake temperature curve. What is not committed is written just as plainly: how fast your workload runs, how well your model trains, end-to-end network latency (we commit only to the rack-to-uplink-port segment), interruptions caused by upstream carriers or force majeure, and any form of return or yield — whether this pays off depends on your own utilisation. The credit cap and sole remedy are in the next item.
Can we claim further after a credit? What is the cap?
No — the credit is the sole remedy. The amount calculated above is our entire liability for that event, and cumulatively it will not exceed what you actually paid us in the 12 months before it. We do not carry indirect loss: the compute cost of re-running a training job, lost revenue or profit, lost data or model weights, or third-party claims against you. Data backup is outside our scope — keep your own copies. Force majeure — war, strikes, earthquake, flood, regional grid collapse, government order, large-scale upstream carrier outage — is excluded from the uptime measurement and generates no credit; we notify you in writing within 24 hours, and beyond 30 days either side may terminate without liability. This item applies equally to all three lines.
How long from first conversation to running?
The plan comes within one business day. After that: rental ships 2 – 5 days after confirmation; purchase runs 4 – 21 days by configuration, including 72 hours at full load before it leaves; hosting is racked 5 – 10 days after signature. The start points are written into the contract: hosting counts from signature and hardware arrival, purchase and rental from cleared deposit and confirmed configuration. Customs waiting, export licence waiting, and waiting on your facility or network are not counted. You do not need to negotiate a rack contract or wait on a power upgrade first — that part is already done.
How are hosting outages credited?
On monthly uptime: below 99.9% credits 10% of that month's hosting fee, below 99.5% credits 30%, below 99.0% credits 100%. The measurement is actual minutes of power loss at the rack PDU. Four exclusions, stated plainly: your own OS, drivers and application failures; hardware failures in your own equipment; planned maintenance we announced 7 days ahead; and force majeure or large-scale upstream carrier outages (see "Can we claim further after a credit"). A utility power cut is not an exclusion — that is exactly what the UPS and the generator are for, and if they do not hold, we credit. Credits are capped at that month's hosting fee for that rack and are applied automatically to the following month's invoice; you do not file a claim.
Can you host hardware we already own?
Yes. Three requirements: standard 19″ rack-mount, no more than 12 kW per machine, and IPMI or equivalent out-of-band management. We take power, cooling, network, racking and on-site part fitting; we do not take your hardware warranty and are not responsible for its failure rate or lifespan — you supply spares, we fit them within 4 hours of arrival. Before racking we photograph and both sign off the arrival condition, and that record is the baseline: subsequent natural failure and ageing are outside our scope, while physical damage from our handling is credited. Your hardware must also meet export-control and compliance requirements, and you warrant title and provenance.
What happens if a rented machine fails mid-term?
We replace rather than repair. In our racks: swapped from the local spares pool and power-on tested within 4 hours of the ticket. At your facility: a replacement machine dispatched within 4 hours, then arriving at the pace of freight and customs — that leg is outside our control, so we do not commit to arrival hours. We collect the failed unit. Downtime from hardware failure is deducted from the invoice hourly with no claim needed; the deduction is capped at the rent, and the overall cap is in "Can we claim further after a credit".

Request a quote

A plan within one business day: scale, configuration, unit price, rack-ready date.

Not "we have received your enquiry" — an email setting out how this compute is configured, which zone it sits in, what it costs and the day it can run.

Response timeWithin one business day (CET business days)
What comes backScale · configuration · unit price · rack-ready date
Quote validity7 days, locked to that day's free capacity

No form. Email, phone, WhatsApp, Telegram and WeChat are on the page — pick one and reach a person. Six things are enough: what you run, the scale, buy or rent or host, the term, where it goes, and how to reach you. We size the configuration to the workload and the budget.

Get the contact detailsSee how hosting works first