GPU · NVIDIA H100 80 GB · 8 cards + HGX node

NVIDIA H100. A single card — or eight on one fabric.

Eight individual H100 80 GB for single-GPU jobs, plus an HGX baseboard with 8× NVIDIA Delta-Next H100 80 GB and NVLink for real training runs. All in our own rack in Falkenberg.

8 cards
H100 80 GB, individually bookable
HGX
8× Delta-Next H100 · 640 GB HBM3
20,70 SEK/h
Monthly cap 14,900 SEK per card
NVLink
NVSwitch fabric on the HGX node
← GPU & AI compute
Two configurations

One card, or all eight on the same fabric.

The difference matters. Individual PCIe cards suit you when one card is enough. The HGX node's NVLink fabric is what makes training across eight cards scale almost linearly.

8× available

H100 PCIe 80 GB

80 GB HBM2e~2 TB/sPCIe Gen5Hopper

Eight individual cards, bookable one at a time. Full memory per card, dedicated host CPU and local NVMe scratch. An NVLink bridge can pair two adjacent cards.

8 cards · h100-01 … h100-08 · live now
1 node available

HGX · 8× H100 80 GB

640 GB HBM3NVLink 4NVSwitchDelta-Next

An NVIDIA HGX baseboard with 8× Delta-Next H100 80 GB on a single NVSwitch fabric. The whole node is yours — nothing is shared, nothing is partitioned.

hgx-delta-01 · 8× SXM5 · live now
Freeh100-01PCIe · SWE1
Freeh100-02PCIe · SWE1
Freeh100-03PCIe · SWE1
Freeh100-04PCIe · SWE1
Freeh100-05PCIe · SWE1
Freeh100-06PCIe · SWE1
Freeh100-07PCIe · SWE1
Freeh100-08PCIe · SWE1
Freehgx-delta-01 · HGX baseboard8× NVIDIA Delta-Next H100 80 GB · NVLink 4 via NVSwitch · 640 GB HBM3 · SWE1 Falkenberg

Our current physical fleet in Falkenberg. Booked capacity is reserved for you alone — we never resell idle hours on a card you pay for.

Specification

Both configurations, side by side.

Component
H100 PCIe (per card)
HGX 8× H100 (whole node)
GPU
H100 PCIe · Hopper
8× Delta-Next H100 SXM5
GPU memory
80 GB HBM2e
640 GB HBM3 total
Memory bandwidth
~2 TB/s
~3.35 TB/s per GPU
Interconnect
PCIe Gen5 ×16 · NVLink bridge possible
NVLink 4 + NVSwitch · 900 GB/s per GPU
Host CPU
24 dedicated vCPU — configurable
Dual-socket, 112 cores
System memory
256 GB DDR5
2 TB DDR5
Local scratch
4 TB NVMe
30 TB NVMe RAID
Networking
10 Gbit/s · no egress fee
25 Gbit/s · InfiniBand on request
Software
Ubuntu 24.04 · CUDA 12.6 · cuDNN · NCCL · PyTorch · Docker — or your own image. Slurm on request.
Access
Root SSH, bare metal or container · REST API + Terraform provider
Location
SWE1 · Falkenberg · Kepler's own hardware on 100% renewable Nordic power

Host CPU, RAM and disk are our standard build per configuration. If you need a different ratio — say so in the booking.

Price

Dedicated hardware, flat prices.

Configuration
Commitment
Price
Note
H100 PCIe · hour
none
20,70 SEK/h
billed per hour
H100 PCIe · monthly cap
none
14 900 SEK/mo
per card, capped at 720 h
H100 PCIe · 12 mo
12 mo
11 622 SEK/mo · −22%
invoiced monthly
HGX · 8× H100
from 1 mo
112 000 SEK/mo · whole node
NVLink · 640 GB HBM3
HGX · 8× H100 · 36 mo
36 mo
67 200 SEK/mo · −40%
best value per GPU-hour

Prices excl. VAT. Monthly cap = price per hour × 720, capped. The commit ladder (3/6/12/24/36 mo) applies to H100 and HGX too — see the pricing page.

Book

Reserve H100 capacity.

An engineer confirms within four hours on weekdays. Individual cards are usually up the same day; the HGX node is provisioned within one business day.

Book NVIDIA H100

No card needed to book. We confirm availability before anything is charged.

By submitting you agree to our privacy policy. We do not spam.

8 cards + 1 node free

No waitlist

The whole H100 fleet and the HGX node are unbooked right now. We confirm per unit, not per queue.

FROM
14 900 SEK/mo

Monthly cap per PCIe card, excl. VAT. The whole HGX node from 112,000 SEK/mo.

Need NVLink?

If your training scales across more than two cards, the HGX node is nearly always the right call. Describe the job — we cost out both.

H100 questions

What we answer most often.

What's the difference between the eight cards and the HGX node?
The eight PCIe cards are individual cards in separate hosts — perfect when one card is enough. The HGX node is a baseboard where eight Delta-Next H100 sit on the same NVSwitch fabric with NVLink 4, giving 900 GB/s per GPU between the cards. It is the difference between eight machines and one training machine.
Can I book half the HGX node?
No. The node is booked whole — that is the point of the NVLink fabric, and we never partition a fabric between customers. If you need fewer cards, individual PCIe cards are both cheaper and faster to spin up.
Do I get Slurm or InfiniBand?
Yes, both on request. Slurm is set up on the HGX node if you want to queue jobs, and InfiniBand can be brought in for multi-node training as we expand the fleet.
What does it cost to move checkpoints out?
Nothing. Inbound traffic is never metered and GPU nodes have no egress fee — your weights are yours, with no ransom to retrieve them.
Do my model weights stay in Sweden?
Yes. Own hardware in Falkenberg, Swedish company, Swedish jurisdiction. No US Cloud Act, no hyperscaler under the layer — see Trust.

640 GB HBM3 on one fabric — in Falkenberg.

The HGX node with 8× NVIDIA Delta-Next H100 80 GB is free. The whole node, NVLink included, single-tenant.

hgx-delta-01 · free now

Train where your weights stay.

Eight individual H100 or the whole HGX node. Own hardware, Swedish soil, no egress fees.