New, refurbished or used GPU servers: what differs, and why we supply only new

New, refurbished or used GPU servers and what we supply

We supply only new equipment: GPU servers built from new components with manufacturer warranty, assembled and tested before delivery. Refurbished and used systems are a different kind of purchase, and the differences can be read in the makers' own documents. NVIDIA's warranty for RTX PRO workstation cards does not extend to a buyer who acquires a card on a used basis, the oldest GPU generations on the second-hand market have a fixed last driver branch and a last CUDA version, and the EU's export control list covers new and used goods alike.

Three cases have to be kept apart: a new system, a system refurbished in its own maker's programme, and a used system from a third party. This comparison states what the documents say about each, as read on 12 November 2026, and where a used system remains a reasonable choice. It judges no seller and no programme.

NVIDIA "RTX PRO Workstation Graphics Cards Warranty Information", EU page (last updated September 2019); NVIDIA "Services for DGX Products – Terms and Conditions" (24 October 2023); NVIDIA Data Center Drivers guide, architecture matrix (last updated 9 September 2026); HPE "Overview HPE Renew" page (no date); Dell Outlet pages for Ireland and the UK; Dell KB 000125103 (last modified 10 June 2026); Supermicro "Limited Warranty" page; Regulation (EU) 2021/821, Annex I. All read 12 November 2026.

What NVIDIA writes about cards and systems that change hands

NVIDIA's warranty for RTX PRO workstation graphics cards runs three years and is tied to the first sale. The EU page, last updated in September 2019, says it applies only to original purchasers who bought "from an NVIDIA-authorized third party or www.nvidia.com", and continues: "this warranty will not extend to any person that acquires a Warranted Product on a used basis." The word non-transferable does not appear in the card warranty. It stands in NVIDIA's terms for DGX support of 24 October 2023, and there it concerns the service contract: "Services are non-transferable, non-assignable and your Services are terminated when the Product is transferred to another party."

For data-centre cards NVIDIA's warranty pages carry no statement of their own, and its warranty index tells buyers of most products to contact the distributor or reseller first. NVIDIA's Enterprise Support guide adds: "Excluding the limited standard warranty on NVIDIA hardware, customers are not entitled to support unless they have purchased a support contract and have an active support entitlement." What a second owner of an accelerator server can claim therefore follows from the terms of the first sale and of the server maker. The periods the makers publish for new systems are in our guide to GPU server warranty in the EU.

What the server makers write about second owners

The makers differ, and one of them documents a way for a used system to keep its cover.

• Supermicro. "This Limited Warranty applies only to the original purchaser of the Product". The warranty is void when the original identification markings are removed, altered or destroyed. The page has no sentence about used goods.
• ASUS. The general warranty card (December 2022) says: "This Warranty applies only if the Product was newly manufactured on the Date of Purchase and not sold as used, refurbished or manufacturing seconds." The RMA guidelines for servers send a buyer who is not the original purchaser to "the original distributor or reseller from whom the product was purchased".
• Gigabyte. The warranty applies to standard products that were sold through the channels Gigabyte has approved and have not been structurally modified without approval. The page says nothing about a transfer or a second-hand sale.
• Dell. Dell's knowledge-base article 000125103, last modified on 10 June 2026, describes an ownership transfer. To use "any remaining warranty" on a Dell system "purchased through an auction site or through a third party", the article says, "you must complete the transfer of ownership form". The details of the previous owner are required, and Dell may ask for proof of purchase. A service contract is valid "only in the country or region in which the computer was purchased" and is moved on request where the service exists at the destination.
• HPE. We could not open HPE's limited warranty document, so this article says nothing about a transfer at HPE.

On-site service contracts follow the same logic and are tied to a registered system and its location, as our guide to on-site service for GPU servers shows.

HPE Renew and Dell Outlet, the makers' own refurbished programmes

A unit that its own maker has taken back and reworked is a separate case, and here the maker's warranty comes with it. HPE describes its programme in these words: "The HPE Renew Program offers you a comprehensive portfolio of completely remanufactured HPE HIT products". The units are "returned, loaned or trial units", and HPE states: "Just like new Hewlett Packard Enterprise products, the warranty period starts from the date of purchase, and refreshes the full original warranty." The page carries no date and, as far as we read it, no statement on who sells the units.

Dell's Outlet page for Ireland describes "Certified Refurbished laptops, desktop gaming computers, monitors, docks & servers" under the heading "We Offer Same-as-New Warranty". The UK page adds: "Outlet systems are refurbished, previously ordered new or scratch and dent." PowerEdge tower and rack servers appear in the Outlet navigation. The pages we read state no warranty length and contain no warranty statement specific to servers.

An offer that says refurbished should therefore say by whom. A unit from the maker's own programme and a unit reworked by a third party are covered differently, and only the first is described on the makers' pages.

We supply new GPU servers and NVIDIA professional cards across the EU; send us the system you are considering and the work it should do and we reply with a configuration and quote within one business day.

Driver and CUDA support for used V100, P40, A100 and H100 servers

The support window of a GPU depends on its generation and is the same for a new and a used card. It matters for used systems because the oldest generations on the second-hand market have reached their last branch.

NVIDIA Data Center Drivers guide, chapter "CUDA Toolkit, Driver, and Architecture Matrix" (last updated 9 September 2026); Maxwell, Pascal and Volta product names from the release notes of driver 580.196.07 (21 October 2026), A100 and H100 NVL from those of driver 620.70.06 (4 November 2026). Read 12 November 2026.

R580 is a long-term-support branch with an end of life in June 2028; R595 is a production branch to March 2027, and R535 ended in June 2026. The release notes of 595.112.07 (22 October 2026) and of 620.70.06 (4 November 2026) list no Maxwell, Pascal or Volta product. NVIDIA's release list marks branch 620 as a long-term-support branch as well, and its notes list the A100 and the H100 NVL.

On the CUDA side the cut has already happened. The CUDA Toolkit 13.0 release notes say about Maxwell, Pascal and Volta: "Offline compilation and library support for these architectures have been removed in CUDA Toolkit 13.0 major version release." The same notes state: "CUDA 13.0 supports all NVIDIA architectures from Turing through Grace Blackwell." A used V100 or Tesla P40 server therefore has a supported driver branch until June 2028 and a software stack that stays at CUDA 12. For virtualised use NVIDIA's vGPU list of 2 October 2026 gives every V100 variant branch 19 as its last, with maintenance to July 2028; for the Tesla P40, P100, P4, P6 and M60 the last branch is 16, with maintenance to July 2026.

For Ampere and Hopper NVIDIA's matrix names no end, and we read no end date for these generations elsewhere. For a used A100 or H100 server the matrix shows the same entry as for a current card; what differs is the warranty position and the unknown history. What replaces a V100 job by job is in replacing the V100 after CUDA 13, the branch types in NVIDIA driver branches and CUDA versions, and the A100 case in replacing the A100.

Export control applies to used goods as well

The EU's dual-use rules make no difference between new and used. Note 3 of the General Notes to Annex I of Regulation (EU) 2021/821 reads: "Goods specified in this Annex include both new and used goods." The same note stands in the Annex I text of Commission Delegated Regulation (EU) 2025/2003. Article 3(1) of the Regulation says: "An authorisation shall be required for the export of dual-use items listed in Annex I." Under Article 12(4) the exporter gives the authority complete information about the end-user, the country of destination and the end-use, and "Individual export authorisations shall be subject to an end-use statement." Article 27(3) requires records to be kept for at least five years from the end of the calendar year of the export.

Whether a particular accelerator or server falls under an entry of Annex I is a classification question that this article does not answer, and we found no official EU text that deals specifically with second-hand data-centre GPUs or with provenance papers for used hardware. This section is general information on the EU text and no legal assessment. By our reasoning, whoever may later send a used system out of the EU needs the same facts about it as about a new one, beginning with the exact model and its origin.

History and accessories of a used system

Two further points come from our own reasoning; no maker document read for this article states them.

The first is the past load. Cards in a server often run under sustained load for years, and an offer seldom documents hours, temperatures or the reason the system was retired. The card itself reports its serial number and its memory error counters, and our guide to GPU server acceptance testing describes those checks. They show the state on the day of the test and say little about the years before it.

The second is what came with the system when it was new. Power cables for the cards, rails and brackets belong to a chassis and are easily separated from it on the way to a second owner. A card that moves into another chassis needs that chassis maker's cable and that maker's listing of the card.

Where a used or refurbished system is a reasonable choice

There are cases in which a used system does its job. In our assessment these are the usual ones:

1. A test bench or a training rig where a failed card costs time and no service is affected.
2. Spare parts for a fleet of the same generation that will stay on its present software until it is retired, for example a V100 cluster on CUDA 12.
3. A unit from the maker's own refurbished programme with the maker's warranty, where the model fits the task.
4. A Dell system for which the ownership transfer is complete and warranty remains.

For a production service that has to run for several years, the missing card warranty, the support window of older generations and the open questions about history weigh more. Users of older workstation cards find the current counterparts in our upgrade map for RTX A-series cards.

What new means in our deliveries

Our pages for AI servers and GPUs state it. Every server is assembled and tested before delivery, with a power-on and basic load test. Every component carries its manufacturer warranty, which is handled through us; DOA units are replaced, and the assembly itself is covered by us. Commissioning on site is available on request. The written quote names price, lead time and warranty terms. We do not supply refurbished or used equipment.

If you are comparing a used offer with a new system, describe the workload and the offer you have, and we answer with a configuration of new components for the same job. What a request should contain is in our guide on what to specify when buying an AI server; the buying channels are compared in where to buy an AI server in Europe, and the checks on a seller are in our GPU supplier checklist.

FAQ