GPU & AI compute · live in Falkenberg

GPU on Swedish soil. H100 and L40S.

Our own hardware in our own rack — 6× NVIDIA L40S, 8× H100 80 GB as individual cards, and an HGX baseboard with 8× NVIDIA Delta-Next H100 80 GB. Single-tenant, the whole card is yours, and your model weights never leave Sweden.

L40S
48 GB GDDR6 · Ada Lovelace
H100
80 GB · individual cards
HGX
8× Delta-Next H100 80 GB, NVLink
EU soil
Falkenberg · 100% renewable power
The fleet · available now

Two cards. Three ways to book them.

Everything below sits physically in the rack in Falkenberg. No waitlist, no overselling, no shared capacity — book a card and it is yours alone.

6× available

NVIDIA L40S

48 GB GDDR6Ada Lovelace864 GB/sPCIe Gen4

The efficiency pick. Inference at scale, RAG, LoRA fine-tuning, vision and GPU rendering — at a fraction of the H100 hourly price.

6 cards · own hardware · live now To the L40S page →
8 cards + 1 HGX node

NVIDIA H100 80 GB

80 GB HBMHopperNVLink on HGX640 GB HBM3 total

The training card. Eight individual cards for single-GPU jobs, plus an HGX baseboard with 8× Delta-Next H100 on an NVSwitch fabric for real training runs.

8 cards + HGX node · own hardware · live now To the H100 page →

Per-card and per-node prices are on each page, or together under GPU pricing. The commit discount of up to 40% applies to GPU too.

Choosing right

Which card do you actually need?

Most people buy too big. If the model fits in 48 GB and you serve requests rather than train from scratch, the L40S wins on cost per token. Take the H100 when memory bandwidth or multi-GPU scale is the bottleneck.

Workload
We recommend
Why
Inference & API serving
L40S
48 GB is enough for most models — best value per token.
RAG & vector search
L40S
Embedding and retrieval are rarely memory-bandwidth bound.
LoRA / QLoRA fine-tuning
L40S
Adapter training fits comfortably and saves you the H100 hourly rate.
Rendering, vision, video
L40S
The Ada generation's encoders do the job far more cheaply.
Full fine-tuning, 7–70B
H100 PCIe
80 GB HBM and ~2 TB/s bandwidth — shorter epochs.
Multi-GPU training
HGX · 8× H100
NVLink via NVSwitch scales almost linearly across eight cards.
Long context window, large batch
H100
It's the bandwidth, not the flops, that sets the ceiling.
Data sovereignty

Your models and your data stay in Sweden.

Training a model means condensing your company's know-how into a few hundred gigabytes. Where those files sit is a trade secret — not a triviality.

Kepler is a Swedish company with its own hardware in Sweden. No US Cloud Act, no foreign jurisdiction, no hyperscaler under the layer. The GPU fleet sits on the same EU soil as the rest of the platform.

Read the full Trust page →

01

Per-tenant isolation

Your own project, your own credentials via OpenBao.

02

The whole card, never shared

No vGPU partitioning, no neighbor stealing bandwidth.

03

No egress fee

Move datasets in and checkpoints out without a surprise invoice.

Included with every GPU

Ready to run, not ready to be configured.

Stack

CUDA 12.6 & PyTorch Now

Ubuntu 24.04 with CUDA, cuDNN, NCCL, PyTorch and Docker preinstalled — or boot your own image.

Access

Root SSH & API

Bare metal or container. The same REST API and Terraform provider as the rest of the platform.

Storage

Local NVMe scratch

Fast disk right next to the card so your dataloader isn't the bottleneck. Persistent volumes for datasets you keep.

Networking

Unmetered traffic

Inbound always free, outbound with no egress fee on GPU nodes. InfiniBand on HGX when needed.

Support

Engineers who run GPUs

Driver trouble, NCCL tuning, thermals — you reach someone who has been inside the rack.

Price

Commit & save −40%

Per hour with a monthly cap, or lock 3–36 months and lower the price predictably without negotiation.

Honestly: the numbers on these pages are our current physical fleet in Falkenberg. If you need more capacity than what is in the rack, we say so plainly — and when we expand, whoever booked first hears first.

15 GPU units free right now.

6× L40S, 8× H100 and an HGX node with 8× Delta-Next H100 80 GB. Describe the workload — we tell you which card fits.

Own hardware · Falkenberg · live now

GPUs that sit in Sweden.

Tell us what you train or serve, and we come back with card, price and start date — usually the same day.