Buy a server with 4 or 8 NVIDIA L40S GPUs in Europe

Who still buys an L40S server in 2026

You can buy a server with four or eight NVIDIA L40S cards from us, built to order and delivered anywhere in the EU. Our team sizes the configuration to your workload and budget and replies within one business day. The L40S is an Ada Lovelace card from 2023 with 48 GB of GDDR6 at 864 GB/s and 350 W; in 2026 it suits companies whose work fits in 48 GB per card and mixes graphics, video and AI, or who extend servers they already run with it.

As of October 2026 we found no NVIDIA end-of-life notice for the L40S, and NVIDIA's vGPU lifecycle page, updated on 2 October 2026, lists it as fully supported. NVIDIA's page for the RTX PRO 6000 Blackwell Server Edition calls it the previous generation. Four kinds of buyer still choose it.

• Graphics and AI on the same hosts. The card has 142 third-generation RT cores and 91.6 TFLOPS of FP32 next to its FP8 Tensor Cores, and NVIDIA names rendering, 3D graphics and Omniverse among its workloads. An engineering firm can run virtual workstations by day and inference or rendering jobs at night on the same servers.
• VDI on the long-term vGPU branch. The L40S is supported by vGPU 19 on the R580 branch, which NVIDIA supports until July 2028, and by vGPU 20. The RTX PRO 4500 Server Edition needs vGPU 20, whose branch NVIDIA supports until March 2027, so an estate that standardises on the long-term branch keeps the L40S.
• Video. Three NVENC and three NVDEC engines per card handle AV1 encode and decode, and NVIDIA's DeepStream 9.2 tables give one L40S 1,114 frames per second on PeopleNet 2.6.3 with a tracker.
• Existing fleets. A company with L40S hosts keeps one vGPU profile set, one power budget and one spare pool when it adds cards of the same type.

The card has no MIG, no NVLink and no FP4 arithmetic. A team that plans a 70B model for many users, hardware isolation between tenants or NVFP4 models is better served by the RTX PRO 6000 Server Edition, as our comparison of the L40S and the RTX PRO 6000 Server Edition explains.

What the four- and eight-card servers contain

The table shows the two configurations we build most often with this card and what we change per project.

Our AI servers page; NVIDIA-Certified Systems configuration guide (30 September 2026) for sockets, cores per GPU, memory, Redfish and TPM; Lenovo Press LP2128 (updated 20 October 2026), Dell PowerEdge XE7745 technical guide (Rev. A09, September 2026) and Cisco UCS C845A M8 data sheet (24 August 2026) for chassis and power supplies.

NVIDIA's configuration guide for NVIDIA-Certified Systems asks for at least two sockets, six physical cores per GPU and system memory of at least twice the total GPU memory. For four L40S that means 24 cores and 384 GB, for eight cards 48 cores and 768 GB. Our guide to how much RAM a GPU server needs explains when more memory pays off, for example for a large vector index next to the models.

The L40S is a dual-slot, full-length passive board, 4.4 by 10.5 inches, with one 16-pin power connector and a PCIe 4.0 x16 link. In a PCIe 5.0 server it runs at Gen4 speed. Its four DisplayPort 1.4a outputs are off by default, so a rendering or VDI host is reached over the network like any other server.

Chassis that list four or eight L40S

The card runs only in a chassis whose maker lists it with the matching risers, power cables and fans. These are listings we read.

Lenovo Press LP2128 ("4x NVIDIA L40S 350W GPUs"), Dell XE7745 technical guide table 23, Cisco data sheet table 2, Puget Systems configurator, read on 24 October 2026.

Dell's guide sets the inlet limit for the XE7745 with L40S cards at 35 °C, the general ASHRAE A2 limit, against 30 °C for the same server with RTX PRO 6000 Server Edition cards. For a server room that runs warm, that margin is a reason to stay with the 350 W card. We quote a chassis only after we have checked its maker's listing for the card count, the processors and the power supplies.

What four or eight L40S run

We size with our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card for the runtime, less the weights, leaves the KV cache for conversations. NVIDIA's CUDA Best Practices Guide states that ECC on GDDR memory takes 6.25 per cent of the capacity, so with ECC on one L40S keeps about 37.3 GiB for weights and cache and with ECC off about 40.0 GiB. The table uses the figure with ECC on.

Our estimates by the users-per-card rule, not measurements: concurrent conversations of 8,192 tokens with an FP8 KV cache, ECC on; desktops and CAD seats are NVIDIA's maximum vGPUs per card from its vGPU types reference (29 September 2026); engines from NVIDIA's L40S product page.

The conversations are requests in flight at the same moment. With the example values of our guide to a private ChatGPT server by company size, 500 staff produce about 20 of them at the busiest moment, 1,000 staff about 40 and 2,000 staff about 80. Four cards with four copies of a 14B model therefore cover 2,000 staff with room to spare, and one copy of Llama 3.3 70B split over the four cards covers 1,000 staff with headroom. A split copy exchanges data over PCIe 4.0 at every layer, so one answer runs slower than the sum of the four cards suggests. Our page on what an 8× L40S server runs has the full model list, published MLPerf® results and the video figures.

We supply both servers with the layout planned per model; send us your models, desktops and peak requests, and our team replies with a configuration and the conversations or seats per layout.

Power, cooling and room

By our estimate, the cards plus two 280 W processors, such as two AMD EPYC 9355, plus 15 per cent for memory, drives, fans and conversion losses, give the figures below. A measurement on the delivered system confirms the estimate.

Our estimate: 350 W per card from NVIDIA's L40S product page, 280 W default TDP of the EPYC 9355 from AMD, plus 15 per cent for the rest of the server.

We plan a 230 V 16 A single-phase feed at 80 per cent of its 3.68 kW, so 2.94 kW. The four-card server stays below that, and with redundant power supplies on an A and a B feed each side carries the whole server. The eight-card server needs a 32 A single-phase feed (5.89 kW planned) or a 400 V three-phase position per side. Dell rates the XE7745's 3,200 W supplies at full power only from 220.1 to 240 V and at 2,900 W from 200 to 220 V, so the voltage at the rack position matters. Our guide to GPU server power on 16 A, 32 A and three-phase feeds works through the feed types.

All of the power ends as heat. By the 160 CFM per kW in our guide to GPU rack power and cooling, four cards need roughly 370 CFM of air at an 11 °C rise and eight cards roughly 620 CFM. NVIDIA rates the card for 0 to 50 °C and the server maker sets the inlet limit per configuration. Dell's guide calls the XE7745 "not recommended for deployment in acoustically sensitive environments", so an eight-card server belongs in a server room or data centre.

Software and the vGPU branch

On request the server arrives with the operating system, the NVIDIA driver, CUDA and a container runtime installed. We install a supported branch: R580, NVIDIA's long-term branch with support to June 2028, or R595. NVIDIA's security bulletin of 30 September 2026 names 580.178.04 and 595.91.07 as the first fixed drivers. Our article on NVIDIA driver branches and CUDA versions explains the choice.

For virtual desktops the branch decides how long the host runs without a major upgrade. vGPU 19 on R580 is supported until July 2028 and vGPU 20 on R595 until March 2027, and the L40S runs on both. NVIDIA's release notes for vGPU 20 on vSphere cover the card on VCF 9.0 and 9.1 and on ESXi 8.0 Update 3 P06 and later. Our guide to L40S vGPU profiles lists every profile with its licence. The card does not come with NVIDIA AI Enterprise; compute vGPU profiles and NIM need it per GPU, while open-source serving engines on bare metal need no NVIDIA licence. Licences come on the same invoice as the hardware.

How ordering works with us

The steps are the same for one server or several.

1. You describe the workload: models, desktops or video streams, data, rack and timeline. Sizing is free of charge.
2. We propose the configuration and check that the cards, chassis, power supplies and cooling fit your rack and feed.
3. You receive a written quote within one business day, with price, lead time and warranty terms per component. Quantities and lead time are confirmed in writing with the quote.
4. You order on one EU contract with an EU invoice from Eurokommerz in Vienna; NVIDIA vGPU and AI Enterprise licences are lines on the same invoice.
5. The server is assembled and tested before delivery, with a power-on and basic load test.
6. We deliver anywhere in the EU, with commissioning on site on request; other destinations on request through our logistics partner.
7. Every component carries the manufacturer warranty, handled through us; DOA units are replaced, and the assembly is covered by us.

We also add L40S cards to servers you already run, after a compatibility check of the platform, power and cooling. For your own acceptance test on delivery, use the checks in our GPU server acceptance testing guide, and to compare quotes line by line, our guide to GPU server quote comparison.

What to send us for a quote

A short description is enough to start. These points let us quote the exact configuration in the first reply.

• Models with quantisation and context length, and the peak requests in flight
• Desktops or workstations per user type, the hypervisor and its version, and the vGPU branch you run
• Video work: streams, codecs, resolution and frames per second analysed
• Existing L40S hosts that the new server must match
• Rack position: free height units, feed type and current, A/B feeds, PDU outlets and inlet temperature
• Delivery site, invoicing entity, timeline and any tender requirements, which our GPU server tender specification helps to word

Describe these points in the form below, and our team replies with a configuration and a written quote within one business day.

When another card fits better

Eight RTX PRO 4500 Server Edition cards give 256 GB at 1,320 W of board power, MIG and FP4, against 384 GB at 2,800 W for eight L40S; our page on eight L40S or eight RTX PRO 4500 Server Edition compares the two servers and the card comparison the single cards. For one or two 70B models for many users, four RTX PRO 6000 Server Edition cards hold one copy per card, as our page to buy a 4× RTX PRO 6000 Server Edition server shows. Published measurements of the L40S are collected in our L40S benchmark review.

What we supply

We supply servers with four or eight NVIDIA L40S cards, built to order in chassis whose makers list them, and the L40S as single cards for servers you already run, on one EU contract and invoice with manufacturer warranty. NVIDIA vGPU and NVIDIA AI Enterprise licences come on the same invoice.

FAQ