AMS
Available
Amsterdam
Netherlands
Ada Lovelace · PCIe card
Rent a dedicated L40S GPU server.
48 GB of GDDR6 ECC at 864 GB/s, in a machine that is yours alone. Monthly, paid in crypto, opened without an identity check and delivered in under 5 minutes.
- $714/moone card, everything included
- 48 GBGDDR6 ECC
- 864 GB/smemory bandwidth
- 1×, 2×, 4×, 8×node sizes available
Price, by node size and term
One number per line, and it is the whole number: no setup fee, no reinstall fee, no bandwidth charge, no support plan. Longer terms are paid up front.
| Node | CPU and memory | Monthly | 3 months | 12 months | |
|---|---|---|---|---|---|
| 1 × L40S | Intel Xeon Silver 4410Y | $714/mo | $678/mo | $628/mo | Configure |
| 2 × L40S | Intel Xeon Gold 5416S | $1,427/mo | $1,356/mo | $1,256/mo | Configure |
| 4 × L40S | 2 × Intel Xeon Gold 6438Y+ | $2,754/mo | $2,616/mo | $2,424/mo | Configure |
| 8 × L40S | 2 × Intel Xeon Platinum 8462Y+ | $5,251/mo | $4,988/mo | $4,621/mo | Configure |
Every configuration above is available in all 6 data centres, racked and burned in, with an image ready to write — which is why delivery is under 5 minutes rather than a working day.
Full specification
GPUNVIDIA L40SAda Lovelace architecture, PCIe add-in card.
Memory48 GB GDDR6 ECC864 GB/s of bandwidth — the number that predicts generation speed, far better than any teraflop figure.
InterconnectPCIe 5.0 ×16Peer-to-peer over PCIe. Fine for one model per card and for pipeline parallelism; slower than NVLink for tensor parallelism. Which one your job needs.
Node sizes1 × card, 2 × card, 4 × card, 8 × cardPCIe cards go up to eight in a standard chassis, and ten only on low-power cards.
Storage2 × 2 TB NVMeHanded over raw beyond the system volume — we impose no RAID level. More NVMe and archive HDD are monthly options.
Network1 Gbit/s, unmeteredBoth directions, no included volume and no overage at any traffic level. One IPv4, a routed IPv6 /64, DDoS filtering on from delivery.
What runs on an L40S
Every model below fits on a single card at 8k context — no sharding, no interconnect to think about. The precision shown is the best one that fits; the arithmetic is in the VRAM sizing guide, so you can check it.
| Model | Best precision that fits | Memory needed | Estimated tokens/s |
|---|---|---|---|
| Llama 3.1 8B | BF16 / FP16 | 20 GB | ~24 |
| Llama 3.3 70B | 4-bit (AWQ, GPTQ) | 43 GB | ~11 |
| Qwen 3 32B | FP8 | 38 GB | ~12 |
| Mistral Small 24B | FP8 | 28 GB | ~16 |
| Gemma 3 27B | FP8 | 33 GB | ~14 |
| Phi-4 14B | BF16 / FP16 | 34 GB | ~14 |
On the 8-card node the same models are held across 384 GB in total, which changes what is possible entirely — up to Llama 3.1 405B. Sharding costs synchronisation at every layer, so a model that fits on one card should stay on one card.
Throughput to expect
Generation is bounded by memory bandwidth: each token requires re-reading the active weights. At 864 GB/s, that sets a ceiling no amount of tuning gets past.
Our figures are a floor, not a promise.
The estimates on this page are calibrated against published measurements and are deliberately conservative — they model single-stream generation. With continuous batching under vLLM the aggregate across concurrent requests is several times higher. If a number here looks low against a benchmark you have seen, that is usually the difference.
Compared with nearby cards
The three cards closest to this one in price. Memory decides what loads; bandwidth decides how fast it answers.
| Card | Memory | Bandwidth | Interconnect | From |
|---|---|---|---|---|
| L40S This page | 48 GB | 864 GB/s | PCIe | $714/mo |
| A100 PCIe | 40 GB | 1,555 GB/s | PCIe | $392/mo |
| A100 PCIe | 80 GB | 1,935 GB/s | PCIe | $1,091/mo |
| RTX 5090 | 32 GB | 1,792 GB/s | PCIe | $335/mo |
See all 12 cards side by side, or let the configurator pick from your model and context length instead of from a price.
Compared with other cards
The arbitrations people actually make against an L40S, each worked out on price, memory, bandwidth and the models that fit.
-
L40S vs RTX A6000
Two 48 GB cards, one built for data centres
See the comparison -
L40S vs A100 PCIe
HBM against GDDR, at similar money
See the comparison
What is included
-
No identity check
No document, no selfie, no phone number, no company registration.
Nothing verified · at any spend
-
The whole card
The GPU is passed through to one machine, and that machine is yours.
No MIG · no vGPU · no hypervisor
-
Root and IPMI
Full root from the first minute, plus remote power and virtual media.
Out-of-band on its own VLAN
-
Unmetered bandwidth
No included volume, no overage tier, nothing to watch on a graph.
1 to 25 Gbit/s · both directions
-
No setup or reinstall fee
Reimage as often as you like, to any operating system we offer.
$0 · reinstalls unlimited
-
Hardware swapped, fast
From spares held in the same suite, at any hour, disks left in place.
Four-hour target · 24/7
Where you can have it
Every L40S configuration is available in all 6 data centres. There is no site where a card costs more, and none where it is unavailable.
STO
Available
Stockholm
Sweden
ZRH
Available
Zürich
Switzerland
REY
Available
Reykjavík
Iceland
YUL
Available
Montréal
Canada
SIN
Available
Singapore
Singapore
L40S server questions
How much does an L40S server cost per month?
A single-card L40S server is $714 a month, with everything included: 128 GB of system RAM, 2 × 2 TB NVMe, unmetered 1 Gbit/s, one IPv4 address and a routed IPv6 /64. There is no setup fee and no bandwidth charge. Multi-GPU nodes run from $1,427 for 2 cards up to $5,251 for 8.
How much VRAM does the L40S have, and what fits in it?
48 GB of GDDR6 ECC per card, at 864 GB/s. The largest model it holds on one card at 8k context is Llama 3.3 70B (4-bit (AWQ, GPTQ)). A 8-card node has 384 GB in total, enough for Llama 3.1 405B.
Is the L40S dedicated, or shared with other customers?
Dedicated. The card is passed through to a single physical machine that only you have an account on — no MIG partitioning, no vGPU layer and no other tenant on the same silicon. On a multi-GPU node you get the whole node, including the PCIe 5.0 ×16 fabric between the cards.
Do I need to verify my identity to rent an L40S?
No. There is no document upload, no selfie, no phone number and no company registration, at any level of spend — an eight-card node opens on exactly the same terms as the cheapest single card. Opening an account takes an email address, a password and a billing address, and none of it is verified.
How do I pay, and how quickly is the server delivered?
In cryptocurrency only: Bitcoin, Monero, Ethereum (ERC-20), Tether (ERC-20) and others. Each invoice gets its own address, the amount is fixed at the rate shown when it is issued, and the machine is provisioned on the first confirmation — root access in under 5 minutes, at any hour.
Can I rent an L40S by the hour instead?
Not here. Every price on this site is a monthly price, with no meter and no per-second billing. Against a typical $2.99/hour cloud, a monthly term is cheaper past roughly 239 hours of the machine simply existing — about 10 days out of thirty.
An L40S of your own, from $714 a month.
Racked, burned in and waiting in all 6 data centres. Pay in crypto and it is yours in under 5 minutes.
Related pages
- Hardware policyThe twelve NVIDIA GPUs we operate, the 72-hour burn-in before a node is sold, the four-hour replacement target, and how disks are erased between tenants.
- GuidesEight practical guides: VRAM sizing, NVLink versus PCIe, image and video models, monthly versus hourly, self-hosting versus an API, serving Llama 70B, and QLoRA.
- DocumentationFirst SSH, verifying the hardware, CUDA containers, serving a model with vLLM, RAID, firewalling, IPMI and reinstalls — the commands, on a real machine.
- NetworkUnmetered ports up to 25 Gbit/s, two carriers and an IX per site, always-on DDoS filtering, routed IPv6 — and the four things we do not offer, stated up front.