Ecodesign Regulation (EU) 2019/424 for servers and GPU servers: what applies, what to ask

What Regulation (EU) 2019/424 requires of a GPU server

Commission Regulation (EU) 2019/424 is the EU ecodesign regulation for servers and data storage products. For servers with one to four processor sockets it sets minimum power supply efficiency, limits for idle power and active-state efficiency, firmware and security updates for at least eight years, a secure data deletion function, construction that allows repair, and a product information sheet the maker publishes. A GPU server with PCIe cards falls under it, but its idle and efficiency figures are measured with the GPU cards removed, so the regulation says almost nothing about the power the GPUs draw.

For a managing director or a purchasing department this has two practical effects. The duties sit with the companies that place servers on the EU market or put them into service, and the information they must publish gives a buyer concrete lines for a tender: power supply efficiency, operating temperature class, firmware support and data deletion. The energy the server uses under AI load has to be stated and checked separately, because the regulation does not cover it.

This guide explains the regulation as it applies to GPU servers, as of October 2026. It is general information on EU law, read from the official texts on EUR-Lex, and not a legal assessment; the regulation is directly applicable in every member state, and whether a duty falls on your own organisation, for example when it imports servers itself, is for your legal counsel to confirm.

Which servers it covers and which it exempts

Article 1(1) covers "the placing on the market and putting into service of servers and online data storage products". Article 2 defines a server as "a computing product that provides services and manages networked resources for client devices", accessed mainly over the network rather than through a keyboard and screen, built for server operating systems or hypervisors, with error-correcting or buffered memory. A deskside workstation used by one engineer at its own keyboard does not meet that definition; a rack GPU server does.

Article 1(2) exempts nine kinds of product. Five of them can matter for a GPU purchase.

Regulation (EU) 2019/424, Article 1(2) and Article 2 definitions, EUR-Lex, read on 25 October 2026. The other exemptions are embedded servers, network servers and small and large data storage products.

A two-socket server with two, four or eight GPU cards, of the kind on our AI servers page, therefore sits inside the scope. Three kinds of server are exempt only from the idle and active-state limits, not from the whole regulation: resilient servers, servers with an integrated APA and HPC servers. The regulation defines an HPC server as one "designed and optimized to execute highly parallel applications, for higher performance computing or deep learning artificial intelligence applications", which must consist of multiple computing nodes clustered to increase computational capability and include high-speed interconnections between the nodes. A single GPU server is one node; whether a particular GPU cluster meets that definition is for the maker and your legal counsel to confirm.

GPU cards count as expansion APAs and are removed for testing

The regulation calls a GPU an auxiliary processing accelerator, APA: "a specialized processor and associated subsystem that provide an increase in computing capacity", which cannot run without a CPU. It distinguishes two kinds. An expansion APA is "an APA that is on an add-in card installed in an add-in expansion slot". An integrated APA is one "integrated into the motherboard or CPU package". The NVIDIA RTX PRO 6000 Blackwell Server Edition, the H200 NVL, the L40S and the L4 are PCIe add-in cards, so in a server they are expansion APAs.

Servers with an integrated APA are exempt from the idle-power and active-efficiency limits. Servers with expansion APAs stay subject to the limits; Annex III states that "the unit under test shall be tested with the expansion APA removed" when idle power, active-state efficiency and active-state performance are measured. If the cards connect through a separate PCI Express switch board or riser, that board is removed for those tests too.

The consequence for a buyer is plain. The idle power and the efficiency score in a GPU server's information sheet describe the server without its GPUs. By our estimate, a server with four RTX PRO 6000 Server Edition cards at 600 W and two EPYC 9355 draws about 3.4 kW at the wall, as our four-card server guide works out, and the cards account for 2.4 kW of it. None of that appears in the regulation's limits. Our article on GPU energy consumption per token shows how to measure the energy against the work done.

Energy efficiency of the base server at idle and in the active state

For servers with one or two sockets, the idle limit is a base allowance plus allowances for the installed components. The base is 25 W for one socket, 38 W for two and 40 W for a blade or multi-node server. On top come 10 W per redundant power supply, 5.0 W per drive, 0.18 W per GB of memory above 4 GB, 4.0 W per buffered DDR channel above eight, an allowance for processor performance, and, for network devices with more than two ports of 1 Gb/s or more, an allowance per active port that rises to 15 W at 10 to under 25 Gb/s, 20 W at 25 to under 50 Gb/s and 26 W at 50 Gb/s or more.

The memory allowance shows how large a GPU server's base becomes. For 768 GB of system memory, the amount we quote with four 96 GB cards, the allowance alone is 0.18 × 764 = 137.5 W by our arithmetic. Active-state efficiency must reach at least 9.0 for one socket, 9.5 for two and 8.0 for blade or multi-node servers. Annex IIIa, added by Regulation (EU) 2021/341, lists ETSI EN 303470:2019 among the transitional measurement methods and asks for testing at an EU voltage and frequency, for example 230 V and 50 Hz.

Regulation (EU) 2021/341 also replaced Article 6 on circumvention: products may not be "designed to be able to detect they are being tested" and change their behaviour during the test to reach a better result.

Power supply efficiency and power factor

The power supply requirement applies to every server in scope, GPU server or not, and it is the line where a GPU server's large supplies come under the regulation directly.

Regulation (EU) 2019/424, Annex II points 1.1.1 and 1.1.2, EUR-Lex, read on 25 October 2026. PF is the power factor; 2020 rows apply from 1 March 2020, 2023 rows from 1 January 2023.

Servers placed on the market since 1 January 2023 must meet the lower rows. In a GPU server, the 10 per cent point matters more than it seems: with redundant supplies on A and B feeds, each supply often carries half the load or less, so its efficiency at low load decides part of the losses. The information sheet states the efficiency at each load point to one decimal place, which lets a buyer compare offers.

The information sheet a maker must publish

From 1 March 2020 the maker publishes product information for each server model in the manual and on a free-access website, until at least eight years after the last unit of the model is placed on the market. Point 3.1 exempts "custom made servers, made on a one-off basis". The sheet lists the following items.

Regulation (EU) 2019/424, Annex II point 3.1, EUR-Lex, read on 25 October 2026. The sheet also gives product type, maker, model numbers, year of manufacture and the components claimed for idle allowances.

The operating condition classes follow Table 6 of Annex II: allowable inlet ranges of 15 to 32 °C for A1, 10 to 35 °C for A2, 5 to 40 °C for A3 and 5 to 45 °C for A4, with 18 to 27 °C recommended for all four. A declared class is a property of the server model; the maker's limit for a full GPU configuration can be tighter. Dell limits its eight-card PowerEdge XE7740 and XE7745 to 30 °C ambient with RTX PRO 6000 Server Edition or H200 NVL cards, as our guide to GPU server inlet temperature shows, so a tender asks for both values.

We build GPU servers to order on server platforms whose makers list the cards for that chassis. Send us the ecodesign and power lines of your specification, and our team replies with a configuration that answers them.

Firmware, secure data deletion and repair

Three material-efficiency points matter over the life of a GPU server.

Firmware, point 1.2.3, from 1 March 2021: the latest available firmware must be made available from two years after the first unit of a model is placed on the market, for at least eight years after the last unit, free of charge or at a fair, transparent and non-discriminatory cost. The latest security update to the firmware must be available from the time the model is placed on the market until at least eight years after the last unit, free of charge. Our guide to GPU server firmware updates explains which parts of a GPU server carry firmware and who publishes it.

Secure data deletion, point 1.2.2, from 1 March 2020: "a functionality for secure data deletion shall be made available" for all data storage devices in the server. The regulation defines it as erasure that makes access to the original data infeasible for a given level of effort, and the information sheet explains how to use it and which standards it supports. When a server leaves service, our five-year GPU server lifecycle guide covers sanitisation records and the WEEE directive.

Repair, point 1.2.1: joining, fastening or sealing techniques must not prevent disassembly for repair or reuse of drives, memory, processor, motherboard, "expansion or graphics cards", power supplies, chassis and batteries. Under point 3.3, third parties dealing with maintenance, repair, reuse, recycling and upgrading of servers who register on the maker's website receive the disassembly instructions free of charge, together with indicative weight ranges for cobalt in the batteries and neodymium in the hard drives.

Lines for a tender or purchase specification

These lines turn the regulation into wording a bidder can answer. They work for a public tender, where our guide to a GPU server tender specification covers the procurement rules, and for a company's own request for proposal.

Our wording, based on Regulation (EU) 2019/424 as read on EUR-Lex on 25 October 2026. The last line is outside the regulation.

The last line matters most for running costs. The regulation measures the server without its cards, so the tender asks for the maximum input power with all GPUs under load and checks it during the acceptance run, as our guide to GPU server acceptance testing describes. Our GPU rack power and cooling guide converts that figure into feeds and airflow.

The review: status in October 2026

Article 8 required the Commission to review the regulation and present the results, with a draft revision if appropriate, to the Consultation Forum by March 2022. The review was to consider, among other things, active-state efficiency and idle power, scope and definitions, material efficiency and critical raw materials, and whether servers with an integrated APA should stay exempt from the idle and active-state limits.

The Commission's review page describes a review study by ICF that "started in January 2023 and is expected to conclude in June 2024". Compliance & Risks reported on 24 February 2025 that a Commission working document with a draft regulation to replace 2019/424 had been circulated ahead of the Consultation Forum of 10 March 2025, and that a possible energy label for servers would be discussed there. On 25 October 2026 the Commission's product page still lists Regulation (EU) 2019/424 as the regulation for servers and states that it "is under review"; we found no adopted successor on the pages we read.

The legal frame has changed meanwhile. The Ecodesign for Sustainable Products Regulation, Regulation (EU) 2024/1781, replaces Directive 2009/125/EC, under which 2019/424 was adopted. Its recital 121 states that existing implementing measures remain in force until a delegated act under the new regulation repeals them. Until a successor is published in the Official Journal, a server placed on the EU market follows 2019/424 as amended by 2021/341.

Data-centre energy reporting is a separate obligation for operators of data centres with an installed IT power demand of at least 500 kW, under Delegated Regulation (EU) 2024/1364, which our article on energy per token summarises.

What we supply

You can buy a GPU server from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two to eight NVIDIA cards, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component; quantities and lead time are confirmed in writing with the quote. If your specification carries ecodesign lines, tell us which ones and how many servers it covers, and the quote answers them line by line.

FAQ