A dedicated card in our own rack in Falkenberg. The efficiency pick for inference at scale, RAG, LoRA fine-tuning and GPU rendering — single-tenant, never shared, never partitioned.
Each unit is a whole physical card with full memory and full bandwidth. Two cards are free to book right now — four are sold out, so if you want more we add you to the waitlist. If you need NVLink, the HGX node with 8× H100 is what you want.
Sold outl40s-01Falkenberg · SWE1
Sold outl40s-02Falkenberg · SWE1
Sold outl40s-03Falkenberg · SWE1
Sold outl40s-04Falkenberg · SWE1
Freel40s-05Falkenberg · SWE1
Freel40s-06Falkenberg · SWE1
Specification
What you get per card.
GPU
NVIDIA L40S · Ada Lovelace
GPU memory
48 GB GDDR6 with ECC
Memory bandwidth
864 GB/s
Interface
PCIe Gen4 ×16, dual-slot
Host CPU
16 dedicated vCPU — configurable
System memory
128 GB DDR5 — configurable
Local scratch
2 TB NVMe, locally attached
Networking
10 Gbit/s · no egress fee
Software
Ubuntu 24.04 · CUDA 12.6 · cuDNN · PyTorch · Docker — or your own image
Access
Root SSH, bare metal or container · REST API + Terraform
Location
SWE1 · Falkenberg · Kepler's own hardware
Host CPU, RAM and disk are the standard build. If your workload needs a different ratio — say so in the booking and we rebuild.
Prices excl. VAT. Monthly cap = price per hour × 720, capped. The commit ladder (3/6/12/24/36 mo) and the Node + L40S reference package are on the pricing page.
Book
Reserve an L40S.
Two cards are free and bookable right now — an engineer confirms within four hours on weekdays and you usually have SSH the same day. Need more than two? Those are sold out; pick the waitlist below and we get back to you in order.
2 of 6 free · 4 sold out
Two cards left — then waitlist
l40s-05 and l40s-06 are free and bookable right now, first come first served. The rest are sold out — pick the waitlist in the form and we get back to you in order when a card frees up or we expand the fleet.
FROM
4 990 SEK/mo
Monthly cap per card, excl. VAT. Down to 2,994 SEK/mo on a 36-month commitment.
Unsure which card?
Describe the workload in the form. If the L40S is not a fit we say so — and point you to the H100 instead.
L40S questions
What we answer most often.
Is the card shared with anyone else?
No. Single-tenant, the whole card — no vGPU partitions, no neighbor stealing bandwidth. Book l40s-05 and it is yours until you release it.
Can I run multi-card jobs?
Yes, up to six cards over PCIe. The L40S has no NVLink, so for training that needs high GPU-to-GPU bandwidth we recommend the HGX node with 8× H100 and its NVSwitch fabric.
What does it cost to move data out?
Nothing. Inbound traffic is never metered, and GPU nodes have no egress fee. You can pull your checkpoints without a surprise invoice.
What software is preloaded?
Ubuntu 24.04 with CUDA 12.6, cuDNN, NCCL, PyTorch and Docker. Want your own image? Boot it instead — the same REST API and Terraform provider as regular instances.
Where does the hardware sit?
In our own rack in Falkenberg, region SWE1, on 100% renewable Nordic power. Swedish company, Swedish jurisdiction, no hyperscaler under the layer — see Trust.
One card. Yours alone.
48 GB on Swedish soil, bookable today. Describe the workload and we confirm the same day.