GPU server lifespan: planning five years of warranty, drivers, licences and refresh

How long a GPU server lasts: the five limits

The lifespan of a GPU server bought in 2026 lasts as long as five things remain covered: the hardware warranty and its extensions, a supported NVIDIA driver branch, the NVIDIA AI Enterprise term where the software needs one, CUDA support for the GPU architecture, and enough GPU memory for the models the company wants to run. The documents cited here give support periods and terms, not a service life in years for the cards. NVIDIA’s architecture matrix of 9 September 2026 lists driver and CUDA support for Ada (L4, L40S), Hopper (H200 NVL) and Blackwell (RTX PRO 6000) as “Ongoing”, so the earliest fixed dates in a plan come from the driver branch, the warranty and the licence terms.

A five-year plan for a server commissioned in late 2026 therefore contains at least one driver branch change, a warranty decision around the third year, licence renewals for cards without an included subscription, and the retirement steps at the end. No document sets a date for the fifth limit, the growth of the models in use.

Lifecycle dates for GPU servers as of October 2026

NVIDIA driver lifecycle and supported-drivers pages, AI Enterprise licensing guide and infrastructure lifecycle page, CUDA architecture matrix and RTX PRO warranty page; Lenovo Press product guide for the SR675 V3; NIST and the Official Journal of the EU, all read on 10 October 2026. The five-year total for the Lenovo server is our addition of base term and extension.

Driver branches and CUDA support over five years

NVIDIA’s driver lifecycle page, updated on 9 September 2026, describes three kinds of branch. A new feature branch arrives about every quarter for early adopters. A production branch gets fixes for a year: “Bug fixes and security updates are provided for up to 1 year.” A long term support branch is a production branch supported for three years, released “at least once per hardware architecture”. In the current table R580 is the long term support branch, supported to June 2028, and R595 the production branch, supported to March 2027.

A server commissioned in October 2026 on R595 changes branch within six months, and one on R580 does so by June 2028, in its second year. NVIDIA does not publish when the next long term support branch arrives, so the change goes into the plan as a maintenance window with a rollback plan, tested on one server before the others. Our guide to driver branches and CUDA versions covers the choice between R580 and R595 and container compatibility.

Volta shows how support for an architecture ends. CUDA 13.0 removed Maxwell, Pascal and Volta from libraries such as cuFFT, whose release note reads “Removed support for Maxwell, Pascal, and Volta GPUs, corresponding to compute capabilities earlier than Turing.” The architecture matrix now lists CUDA 12.x as the last toolkit for Volta and R580 as its last driver, so a V100 server keeps a supported driver until June 2028 and stays on 12.x containers. For the cards in a 2026 purchase, each CUDA release note has a section titled “Deprecated or Dropped Features”, the place to check at every toolkit update.

For virtual desktop and vGPU hosts, NVIDIA’s list of GPUs supported by vGPU, updated on 2 October 2026, gives each card a support status. The L4, the L40S and the RTX PRO 6000 Blackwell Server Edition are in “Full Support”, the T4 and the A40 in “Extended Support”, and the Volta, Pascal and Maxwell cards in “Maintenance”. The page does not define the phases.

NVIDIA AI Enterprise terms over the life of a server

NVIDIA’s licensing guide, updated on 2 September 2026, states: “Each NVIDIA H200 NVL Tensor Core GPU includes a five-year NVIDIA AI Enterprise subscription. Software activation required.” The start date “is based on the ship date of the GPU board to the OEM partner, plus 90 days” and cannot be changed, because it is tied to the card. The end date therefore follows each card, not the day the server goes into service, and the term may already be running when the server arrives. Record the serial numbers and activate the subscriptions on delivery, since cards added in a later year end on later dates.

For the L4, the L40S and the RTX PRO 6000 Server Edition, which include no subscription, the licence is bought per GPU where the software needs it. A subscription runs for its stated term, “after which it must be renewed to remain active”. A perpetual licence includes five years of Business Standard support, which can then be renewed each year. The licence term is separate from the software branch. NVIDIA’s infrastructure lifecycle page states that drivers are supported “for 1 year unless the branch is designated as an LTSB Infrastructure Branch, in which case it is supported for 3 years”, and lists infrastructure 7.x, on R580, as the long term branch. It gives no end date for 7.x, so the plan works with the R580 date of June 2028. Our AI Enterprise licensing guide explains which deployments need the licence at all.

Warranty, extensions and spare parts

For a server, the warranty is set per machine type. Lenovo’s product guide for the SR675 V3, updated on 8 October 2026, lists a “Three-year or one-year (model dependent)” limited warranty, three years for machine type 7D9R and one year for 7D9Q. Its optional service upgrades include a “1-year or 2-year warranty extension”. Added to the three-year base, the longer extension gives five years, but server makers word these terms differently, so have the quote state the total term. NVIDIA warrants its RTX PRO Blackwell workstation cards for “Three (3) years from the date of purchase”, with claims through the distributor or reseller; its warranty page lists the Workstation and Max-Q editions of the RTX PRO 6000, not the Server Edition.

Have the quote state whether the GPUs fall under the server’s warranty and its extensions, the response time, and what happens to a failed NVMe drive. A plan that runs past the base warranty also needs spares for the replaceable parts, such as fans, power supplies and drives. Same-model GPUs are easiest to add while the server maker still lists the card for the machine type, and our guide on adding GPUs to a server later covers what has to be ordered on day one for that.

We supply AI servers with manufacturer warranty on every component, and the warranty terms come in writing with the quote. Tell us how many servers and cards you plan over five years, and we quote the servers and the licences together.

A five-year plan for three GPU servers

As an example, take a company of 1,500 staff that commissions three servers in October 2026. Servers A and B each hold eight H200 NVL for the production assistant and RAG, so that one keeps the service running when the other is in maintenance. Server C holds eight RTX PRO 6000 Server Edition for embeddings, document models, development and virtual workstations. The 16 H200 NVL include their AI Enterprise subscriptions; for server C, licences are bought per GPU where vGPU compute profiles or NIM in production need them.

Example only. Dates from the NVIDIA driver table of 9 September 2026 and the licensing guide of 2 September 2026; later branch dates are not yet published and are left open.

Unlike the driver and warranty dates, the refresh decision of year 4 depends on how the models in use have grown against 141 GB per H200 NVL and 96 GB per RTX PRO 6000, which our per-model sizing guides cover.

For a mid-life addition or a refresh, we check the rack, power and airflow before we quote. Send us the servers you run today and the year each warranty ends through the form below.

Second life: from production inference to development

When newer servers take over production inference, the older ones can serve where a slower response or a maintenance stop costs less. Development and test environments, evaluation runs, batch jobs such as document indexing, embedding models and fine-tuning experiments on smaller models fit there. Before the move, check that the driver branch you need still supports the cards, that the AI Enterprise term covers any software that needs it, and whether the server runs past its warranty, where spare parts take the place of the service contract.

Our article on replacing A100 PCIe cards sets out what changes in power, cooling and software when a generation is replaced, and the H200 NVL against H100 NVL comparison shows when more memory per card pays off and when the cards already in use are enough.

Retirement: NVMe sanitisation and WEEE take-back

The drives of an AI server hold RAG indexes, vector stores, prompt and response logs and fine-tuned weights. NIST SP 800-88 Rev. 2, published in September 2025 and superseding Rev. 1 of 2014, sorts sanitisation into clear, purge and destroy. Clear is typically applied through standard read and write commands, while purge techniques “make the recovery of target data infeasible”. NIST states that “Except for cryptographic erase (see Sec. 3.2), technology-specific sanitization techniques are out of scope”, refers to IEEE 2883 for the techniques, and asks that “a certificate of sanitization (see Appendix C) should be completed for each ISM”, that is, for each storage medium.

On NVMe drives that support it, the Sanitize command starts a block erase, an overwrite or a crypto erase. The nvme-cli documentation lists crypto erase as action 0x04 and takes the controller’s character device, for example nvme sanitize /dev/nvme0 --sanact=0x04, with --wait to wait until the operation finishes. Record each drive’s serial number, the method and the result on its certificate.

The equipment itself falls under Directive 2012/19/EU on waste electrical and electronic equipment (WEEE). Article 13 requires Member States to ensure that the costs for “collection, treatment, recovery and environmentally sound disposal of WEEE from users other than” private households are financed by producers for products placed on the market after 13 August 2005. Article 13(2) adds that producers and business users “may, without prejudice to this Directive, conclude agreements stipulating other financing methods”. How these rules apply to a specific contract is a legal assessment for the company’s legal department.

What we supply

We build AI servers to order with the H200 NVL and NVLink bridges or the RTX PRO 6000 Server Edition, burn-in tested and with manufacturer warranty on every component, on one EU contract and invoice. The GPUs themselves, from the L4 and L40S to the RTX PRO range, come with the same warranty, and NVIDIA AI Enterprise and vGPU licences come on the same invoice as the hardware. We install the operating system, drivers, CUDA and a container runtime on request, so the server arrives ready for your team.

FAQ