Comparing GPU server quotes: the line items behind the total and the five-year cost structure

How to compare GPU server quotes line by line

When you compare GPU server quotes, two offers for the “same” eight-GPU server often describe different machines and different services. Before you compare totals, put both quotes into one list of line items: the exact GPU edition and its power cables, chassis and power supplies, processors, memory, NVMe drives, network cards with cables and transceivers, rails and power cords, warranty term and service level, burn-in, commissioning and software installation, NVIDIA licence lines, delivery terms and take-back. A total becomes comparable once every line is either in both quotes or added to the one that leaves it out.

Our checklist. GPU and licence facts from NVIDIA’s product pages, vGPU support list, vGPU licensing guide and AI Enterprise licensing guide, Lenovo Press guides LP2263 and LP1611, ICC’s Incoterms® 2020 page and Directive 2012/19/EU, read on 10 October 2026.

Where a private AI platform for several departments starts with two servers of eight cards each, every line a quote leaves out appears twice, and a missing licence line is missing sixteen times. The wording for a formal tender is the subject of our guide to writing the technical specification for a GPU server tender; this article reads the quotes that come back.

GPU edition, power supplies and cables

A line reading “RTX PRO 6000 96 GB” does not say which card it is. NVIDIA lists the RTX PRO 6000 Blackwell Server Edition with a maximum power of “Up to 600W (configurable)” and 1,597 GB/s of memory bandwidth. The RTX PRO 6000 Blackwell Max-Q Workstation Edition has the same 96 GB with 1,792 GB/s, a 300 W limit and an “Active” thermal solution, a card with its own fan for a workstation. NVIDIA’s list of GPUs supported by vGPU, updated on 2 October 2026, names the Server Edition and no Workstation or Max-Q edition of this card. If the server will host virtual machines with vGPU, a Max-Q line is a different product.

Lenovo’s product guide for the Server Edition, updated on 28 July 2026, lists a second variant power-capped to 450 W “to allow 4x GPUs to be installed in the SR650a V4”. It also states that the GPU option “does not ship with auxiliary power cables” and that “Cables are server-specific due to length requirements”. In a factory configuration the configurator adds them; for a field upgrade they are separate items. Check that each card has its cable and that the quote says whether the cards run at their full rating or capped.

For the H200 NVL, NVIDIA lists “2- or 4-way NVIDIA NVLink bridge” at 900 GB/s per GPU. The bridges are items of their own, and the quote should say how many there are and which cards they join. Eight 600 W cards draw up to 4.8 kW before processors and fans, so ask each supplier for the power supply count, the rating per supply and whether the server keeps redundancy at full load. Our guide to specifying an AI server before the quote covers feeds, outlets and airflow at the rack position.

CPU, memory, NVMe and network items

Processors, memory modules and drives should appear as items with model, count and capacity, not as a summary such as “1 TB RAM”. Module size and count show whether every memory channel is populated, and two quotes with the same total can differ there. For NVMe, compare the number of drives, the capacity and the endurance class of each, and which drives hold the operating system mirror. If your data rules require it, ask whether a failed drive must be returned under the warranty or may stay on site.

Network cards are often quoted without what connects them. List the ports and their speed, and for each port the cable or transceiver type and length to your switches, plus the out-of-band management port. Rails and power cords complete the item list: the rail kit for your rack depth and one cord per power supply that matches your PDU outlets.

Warranty term and service level

Warranty lines differ in term, response and what is replaced on site. Lenovo’s product guide for the ThinkSystem SR675 V3, updated on 8 October 2026, gives a “Three-year or one-year (model dependent) customer-replaceable unit and onsite limited warranty” with “9x5 next business day (NBD)”. Its optional upgrades are “4-hour or 2-hour response time, 6-hour fix time, 1-year or 2-year warranty extension”. A response time says when someone answers or arrives, and a fix time says when the server runs again, so compare the same measure in both quotes. Customer-replaceable units are parts your own staff exchange, so ask which parts those are on the quoted model.

A five-year plan on a three-year base warranty needs the extension as a line now, not as a renewal later. For the GPUs, settle who receives a failure report: the server maker, the card supplier or the reseller, and whether a replacement card ships before the failed one is returned.

On GPUs and servers we supply, the manufacturer warranty is handled through us and DOA units are replaced. Write to us with the service level your operations need, and we quote the warranty terms in writing with the configuration.

Burn-in, commissioning and software installation

Ask whether the server is assembled and load-tested before shipment, which test runs and whether the report comes with the delivery. Our GPU server acceptance testing guide lists the checks to repeat on site and the pass criteria to agree in advance. Commissioning on site, the operating system, the NVIDIA driver branch, CUDA and a container runtime are separate services. A quote that lists them as optional without saying what they include cannot be compared with one that does.

NVIDIA AI Enterprise and vGPU licence lines

NVIDIA’s licensing guide, updated on 2 September 2026, says the software “is licensed on a per-GPU basis” and that “A software license is required for every GPU installed on the server or workstation”. Two servers with eight cards each therefore need sixteen licences wherever the software is needed: for NIM in production, for compute vGPU profiles and for NVIDIA support, as our article on how NVIDIA AI Enterprise is licensed explains. An open-source engine on bare metal needs none.

The guide states that “Each NVIDIA H200 NVL Tensor Core GPU includes a five-year NVIDIA AI Enterprise subscription. Software activation required.” The term starts from the card’s shipment to the server maker “plus 90 days to account for integration and final delivery to the customer site”, so it is fixed per card. Activation needs the card serial numbers, so ask for them with the delivery. For the included subscriptions, the guide says the software “can be installed on the NVIDIA Certified Systems in which the GPUs are installed”, so check that the quoted server model is on NVIDIA’s certified list with your cards.

For the RTX PRO 6000 Server Edition, Lenovo’s guide lists a “Software Kit (NVAIE, Run:ai, vGPU, Support), per GPU, 3Years” and a kit “per GPU, 5 Years”. Subscriptions must be renewed “to remain active”, so a three-year kit in a five-year plan brings a renewal in year four. Business Standard support is included, and NVIDIA offers Business Critical as an upgrade.

Graphics vGPU editions are counted per concurrent user, not per card. NVIDIA’s vGPU packaging and licensing guide, updated on 24 March 2026, states that “NVIDIA vApps, NVIDIA vPC, and NVIDIA RTX vWS are available on a per Concurrent User (CCU) model.” Compare the number of concurrent users in both quotes.

We supply NVIDIA AI Enterprise and vGPU licences on the same invoice as the hardware. Tell us the cards, the virtual machines and the users per card in the form below, and the licence lines come with the configuration.

Delivery terms and take-back under the WEEE Directive

The International Chamber of Commerce describes the Incoterms® rules as “a set of standards used in international and domestic contracts for the delivery of goods”; Incoterms® 2020 entered into force on 1 January 2020. The rule named in the quote decides who carries the transport risk and up to which point. ICC’s page notes that “under DAP the seller does not unload the goods, under DPU, the seller does unload the goods.” An eight-GPU server still has to be unloaded, carried to the room and racked, so check which quote includes those steps and the delivery rules of the data centre.

For the end of life, Article 13(1) of Directive 2012/19/EU says that the financing of collection, treatment, recovery and environmentally sound disposal of WEEE “from users other than private households” for products placed on the market after 13 August 2005 “is to be provided for by producers”. Article 13(2) allows producers and such users to “conclude agreements stipulating other financing methods”. Have the quote or contract name the take-back terms and who erases the drives before collection. How the directive applies to a given purchase is a legal assessment for the company’s legal department.

Five-year cost structure of a GPU server

A quote total covers the purchase. Over five years, the costs fall into categories that each have a driver and a place where the figure comes from.

Categories are ours; licence terms from NVIDIA’s AI Enterprise licensing guide (2 September 2026) and Lenovo Press LP2263, warranty options from Lenovo Press LP1611, read on 10 October 2026.

Power is the category you can bound before buying. Five years are about 43,800 hours. Eight cards at 600 W draw at most 4.8 kW of board power, so one server’s cards use at most about 210,000 kWh in five years and two servers about 420,000 kWh, before processors, fans, power-supply losses and cooling. Servers draw less when they wait for requests, and our article on GPU energy per token shows how to measure the draw with DCGM and vLLM counters; multiply the measured kWh by your own tariff.

The other categories follow from the quote lines above. A three-year warranty and a three-year licence kit each add a renewal inside the five years unless the quote carries the extension or the five-year kit, while the H200 NVL subscription runs from a date fixed per card. Operating effort depends on who runs the servers after delivery, your staff or a contracted service, and belongs in the comparison even though no quote shows it.

What we supply

We build AI servers to order around the workload, assembled and burn-in tested with a test report on request, with manufacturer warranty on every component and delivery anywhere in the EU, on one EU contract and invoice. The cards include the RTX PRO 6000 Server Edition, the H200 NVL with NVLink bridges, the L40S and the L4, listed on our professional NVIDIA GPU page. NVIDIA AI Enterprise and vGPU licences come on the same invoice as the hardware. Operating system, drivers, CUDA and a container runtime are installed on request, and commissioning on site is available on request. We check the rack, power and airflow before we quote, and the configuration and quote follow within one business day.

FAQ