Colocation for GPU servers in the EU: what to ask a data centre before placing AI servers

GPU colocation in Europe: the questions in one table

Colocation for GPU servers, and so for most AI servers, is decided by power per rack. A PCIe server with eight 600 W cards draws about 6.3 kW by our estimate, Dell's datasheet gives about 8 kW for its eight-card nodes, and one NVIDIA DGX B300 is rated at 15.1 kW, while in Uptime Institute's 2026 survey the average of the facilities' most typical rack densities is above 11 kW. Ask the operator how many kW per rack it delivers and cools, over which feeds, at which supply air temperature and how the power is billed. We do not offer colocation; this checklist is for companies that place the GPU servers we supply in a third-party data centre.

our checklist; the figures behind the third column are given with their sources in the sections below.

Colocation power density: kW per rack for GPU servers and in the market

The server side first. These are the figures our articles use, with their basis.

our articles on GPU rack power and on air-cooled AI servers in a 10 kW rack (estimate: board power of the cards and the processors plus 15 per cent for the rest); Dell's figure and NVIDIA's DGX B300 figures as cited there and on our DGX B300 page. A measurement on the delivered system replaces the estimate.

How the estimate is built is in our guide to GPU rack power and cooling, and which servers fit an existing rack is in air-cooled AI servers in a 10 kW rack. The second row uses the NVIDIA RTX PRO 4500 Blackwell Server Edition (specs, buy page): a server with eight of these cards stays near 2.2 kW by our estimate, inside the "4-5 kW average rack design density" that Uptime's survey names. We build servers to order and supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project; the DGX B300 is supplied per project as well.

The market side comes from Uptime Institute's Global Data Center Survey 2026 (Keynote Report 209, July 2026; more than 800 owners and operators, surveyed in April and May 2026). It states: "For the first time, the average of modal (most typical) rack densities surpassed 11 kW in our sample", after 9 kW the year before. The figure needs care. It is the most typical rack of each facility, averaged over facilities, and not an average over all racks. "Without these high-density facilities, the average modal density sits at 7.8 kW, up from 7.5 kW in 2025." Racks between 8 and 15 kW are the modal band in a quarter of the facilities. The survey also says that "The majority of facilities still do not have any racks at 30 kW or above in deployment"; 24 per cent of respondents have some, up from 19 per cent in 2025.

Higher densities are on offer. One large operator's product page for high-density colocation, undated and read on 12 November 2026, starts "at just 30 kW per cabinet" and goes "up to 150 kW per cabinet" with air-assisted and direct liquid cooling. That is one product line of one operator and an example of what can be bought, not a market norm.

We supply GPU servers built to order across the EU and check rack, power and airflow before we quote; send us the rack specification from your colocation offer and we size the configuration to it.

Feeds, redundancy and circuits at the rack

One large operator's product documentation, read on 12 November 2026, describes its arrangement: "Power is delivered to cabinets through redundant pairs of AC circuits." The contract sets a limit on the load drawn across all circuits of the cabinet, and circuits may be installed with usable power of up to twice that limit; in our reading, either circuit of a pair can then carry the whole load. In its EMEA region this operator counts 100 per cent of a circuit as usable: 3.68 kVA on a 230 V 16 A single-phase circuit and 22.17 kVA on a 400 V 32 A three-phase circuit. Our own articles plan at 80 per cent, which gives 2.94 kW and about 17.7 kW. Ask which rule your operator applies, and whether the contract counts in kVA or in kW: the two differ by the power factor.

On the server side, the question is whether the power supplies can be split over two feeds. That depends on their redundancy design, which our article on PSU redundancy in GPU servers explains, and the circuit each server size needs is in our guide to 16 A, 32 A and three-phase feeds. The AC version of the DGX B300 has twelve C19/C20 inlets, so one system takes twelve C19 outlets.

For the building behind the rack, operators quote redundancy as N+1 or 2N and often an Uptime Tier. What a Tier III rating covers, and why Uptime attaches no availability percentage to it, is in our article on the Tier 3 data centre.

How colocation power is billed: flat rate or metered

Operators' pages describe two models. With unmetered power the customer pays "a flat rate for the circuit(s)", whatever is drawn. With metered power the customer pays for measured consumption, which one operator's page (dated 2017) calls "a pass through cost from the data center to the customer"; its meters show cumulative kWh and real-time kW, and a demand option records kW peaks at 15-minute intervals. Questions that follow for a GPU rack:

• Is the contracted power a commitment that is paid in full, or a ceiling?
• What happens when the measured load passes the contracted limit?
• What does the energy charge cover besides the kWh the servers draw?
• Can the contracted power be raised during the term, and is the capacity reserved?

Cooling method and inlet temperature for AI servers

Every watt a server draws leaves it as heat. Our rack guide plans about 160 CFM of air per kW at an 11 °C rise, so the operator's answer has to cover air volume as well as kW. For temperature, ASHRAE's recommended range for classes A1 to A4 is 18 to 27 °C at the inlet (Table 2.1 of its Thermal Guidelines, fourth edition, errata of 25 July 2016); the allowable ranges are wider. The edition of 2021 added a class for high-density servers, whose values we could not confirm on an ASHRAE page and do not print.

Server makers set their own limits. Dell limits the ambient temperature to 30 °C for RTX PRO 6000 Server Edition and H200 NVL cards in its eight-card PowerEdge XE7740 and XE7745, as our article on air-cooled servers cites. NVIDIA's user guide for the DGX B300 gives an inlet air range of 10 to 35 °C, and its data-centre guide plans supply air of at most 30 °C. The user guide lists fans, airflow and inlet air and no liquid connection. Rack-scale systems, which we do not supply, are a different class: NVIDIA describes its DGX GB200 as using "hybrid cooling with both direct liquid cooling and air cooling". What rear-door heat exchangers and direct liquid cooling involve is in our guide to liquid cooling for GPU racks, and how chassis height changes airflow is in 4U vs 8U GPU servers.

EU rules and standards: what an operator reports and what you can ask for

Two EU acts concern the operator. Article 12 of Directive (EU) 2023/1791 has Member States require owners and operators of data centres with an installed IT power demand of at least 500 kW to make information public every year, with an exemption for trade and business secrets. Commission Delegated Regulation (EU) 2024/1364 of 14 March 2024 has operators of data centres from the same threshold report to a European database by 15 May of each year for the previous calendar year. It defines a colocation data centre as one "in which one or more customers install and manage their own network or networks, servers" and storage. The reporting duty lies with the operator. Colocation customers do not report, and the operator may collect data from them.

What is reported helps a buyer: installed IT power demand, total and IT energy consumption, water input, reused waste heat, the average setpoint of the IT intake air temperature, and the electrical and cooling redundancy down to rack level. Annex III gives the indicators, among them "PUE = EDC/EIT", the energy consumption of the data centre over that of its IT equipment. The database publishes data "in an aggregated manner, at Member State and Union level", so there is no public register in which to look up a site: ask the operator for the figures it reports. Neither act, as far as we read them, sets a PUE limit. The Commission proposed a common rating scheme on 21 September 2026, and its page plans a proposal on minimum performance standards for the second quarter of 2027. This is general information; the directive is applied through national law, which can differ.

Among standards, EN 50600 is the European series on data centre facilities and infrastructures: EN 50600-2-2 covers power supply and distribution, EN 50600-2-3 environmental control, EN 50600-2-4 telecommunications cabling, EN 50600-2-1 building construction and EN 50600-2-5 security systems. The CEN/CENELEC/ETSI report on green data centres (11th edition, approved 5 November 2024) describes four availability classes each for power, environmental control and cabling, with the lowest of the three as the overall class. The series is being revised in 2026, so ask which edition a certificate refers to. ISO/IEC 22237 is the international series with the same title.

Delivery, remote hands and cross-connects

For the floor we found no figure: the operator documents we read publish no floor load per rack position, so it has to be asked for, together with the route from the loading dock. NVIDIA's figures for the DGX B300 with power supplies show the scale: 168 kg in the rack and 207.6 kg packed. For other servers, ask us for the weight and depth of the configured system; the checks are set out in our article on the site survey for GPU servers.

Remote hands is the operator's on-site support service. One operator's documentation lists installing equipment, cabling, power cycling, moving shipments from the loading dock to the cage and troubleshooting, on request or scheduled in advance. Ask which tasks are included, at which hours, and who may handle a GPU server under the maker's service terms; our article on on-site service for GPU servers lists what makers require. For the DGX B300, NVIDIA requires installation by NVIDIA partner network personnel or NVIDIA field service engineers, while the colocation provider prepares the rack and the power.

A cross-connect is, in one operator's words, a set of "dedicated, physical cabling links between two parties" inside the data centre. The same operator accepts only single-mode fibre for new orders since 31 March 2024, charges an installation fee and a monthly fee, and requires a Letter of Authorization for a link into a third party's meet-me room. Ask which carriers are present before you plan how data reaches the servers.

Commissioning on site is part of what we do on request; tell us the data centre, the rack position and its delivery rules, and we reply within one business day.

Location and data protection questions

Location decides two practical things: how quickly your own staff or the maker's technician can reach the rack, and which national law applies the EU rules above. The data protection question is separate. Whether a colocation provider is a processor under the GDPR depends on whether it can access personal data. The EDPB's Guidelines 07/2020 (version 2.1, adopted 7 July 2021) name two conditions for a processor, "being a separate entity in relation to the controller" and "processing personal data on the controller's behalf", and say that "not every service provider that processes personal data in the course of delivering a service" is one. They contain no statement on colocation. The Bavarian commissioner for data protection, the supervisory authority for public bodies there, wrote in its guidance on processing (version 2.0, 1 April 2019) that there is no processing on behalf of the controller where access to personal data is not possible, and that the rules apply correspondingly where access during maintenance cannot be ruled out. Remote hands with possible access to disks or consoles are therefore the case to settle in the contract. The legal assessment is yours and your data protection officer's.

If you would rather not own and house servers at all, our EU Cloud hosts IaaS or a private cloud in Tier III data centres in the EU; that is hosted infrastructure, not colocation.

FAQ