Skip to content
State of GPU Compute 2026

Instant Clusters

Up to 64 GPUs, one fabric.

Dedicated 8-GPU servers on a 400 Gb/s per-GPU network, with a shared volume and Slurm ready at boot. Stop it when the run ends.Join dedicated 8-GPU servers on a 400 Gb/s per-GPU network, with a shared volume and Slurm ready when the cluster boots. Create it in the console, stop it when the run ends.
Create a clusterPlan a reserved cluster
A Nexpod Instant Cluster: four 8-GPU H200 servers in a two by two layout, each GPU board linked to an InfiniBand switch, 32 GPUs in total.

The fabric

Every GPU gets its own 400G port.

Each server hangs off a leaf switch with one port per GPU, and every leaf connects to both spines. All-reduce traffic never queues behind storage or another tenant.
cluster · h200-x4 · eu-fra-1 fabric up
Spine 01
Spine 02
Leaf 01
Leaf 02
node-01
node-02
+2 nodes on leaf 03 and 04
Nodes
4
GPUs
32× H200
Fabric
400 Gb/s per GPU
Rate
€128.00/hr

Plans

Self-serve today, or held for months.

Instant Clusters start from the console with no contract. Reserved Clusters lock capacity and a rate for a long training run.
  • Self-serveconsole · nexctl · API

    Instant Clusters

    Create one, train, stop it. Metered per second on every node.

    Servers
    2 to 8
    GPUs
    16 to 64
    Fabric
    400 Gb/s per GPU, RDMA
    Ready with
    Shared volume · Slurm
    Term
    None
    FamilyPer node · hr8 nodes · hr
    • B300€48.00€384.00
    • B200€40.00€320.00
    • H200€32.00€256.00
    • H100€24.00€192.00
    Create a cluster
  • Planned with our teamquote in one business day

    Reserved Clusters

    Capacity held from a start date you pick, at a fixed euro rate.

    Servers
    More than 8
    Fabric
    Designed around your job
    Spares
    Standby nodes for swaps
    Interconnect
    Private link to your cloud
    Term
    From 3 months, quoted
    Family12-month rate · node·hrvs on demand
    • B300€40.00−€8.00
    • B200€32.00−€8.00
    • H200€24.00−€8.00
    • H100€16.00−€8.00
    12-month rate shown. Shorter terms, from 3 months, are quoted.
    Talk to an engineer

In plain words

What is an Instant Cluster?

Two or more Nexpod servers on one private, high-bandwidth network, booted as a single unit for distributed training, large-model inference and jobs that need more than eight GPUs at once.

Self-serve

A cluster without a purchase order.

Console or CLI, your container, Slurm on top, and a meter that stops with the cluster.
Create a clusterTalk to an engineer
nexctl — clusters
$ nexctl clusters create h200-x4 \
--family h200 --nodes 4 \
--region eu-fra-1
→ placing 4 servers on leaf pair 03/04
→ fabric test · all-reduce 32 GPUs
→ volume ckpt-eu mounted on 4/4
✓ cluster ready in 3m 05s
ssh head@h200-x4.fra.nexpod.cloud
$
nodes 4/4 · fabric test passed · 3m 05s
request to Slurm
3m 05s
meter, every node
1 s
servers, one fabric
2 to 8
  • Bring-up · h200-x43m 05s
    1. 00:00Request · 4× H200 in eu-fra-1
    2. 00:38Nodes placed on one leaf pair
    3. 01:52Fabric test · all-reduce passed
    4. 02:40Volume ckpt-eu mounted on 4/4
    5. 03:05Slurm up · 32 GPUs idle

    Minutes from request to Slurm

    Placement, fabric test and volume mounts run before you log in.

  • h200-x4 running
    00:45:12
    4 nodes · 8× H200
    €128.00/hr
    Per second
    €0.03556
    Billed so far
    €96.43

    Metered per second, per node

    Four servers, one meter, stopped in one command.

  • tokens/s · ×10001 → 8 nodes
    118
    231
    342
    455
    561
    668
    774
    879
    1n··4n···8n

    Add nodes, add throughput

    Collectives run on a dedicated fabric, so training throughput tracks node count from one server to eight.

Reserved Clusters

Held for your dates. Priced for the term.

For runs that last months: capacity booked in advance, a fabric planned around your job and a rate fixed for the whole term.
Plan with an engineer
  • Reserved · 20273 blocks
    JFMAMJJASOND
    32× H200
    64× B300
    16× H100
    16× H100
    Jan to Apr
    32× H200
    Mar to Aug
    64× B300
    Jun to Dec

    Capacity on your calendar

    Book blocks of servers from a start date, months ahead.

  • Fabric partitionyours · 8 nodes
    your partition not routable

    A partition of your own

    Your fabric is fenced off. Nothing routes in from other tenants.

  • One term · Jan to Dec 2027Sample
    152 GPUs reserved
    • B30064 GPUs · 8 servers
    • B20040 GPUs · 5 servers
    • H20032 GPUs · 4 servers
    • H10016 GPUs · 2 servers

    Many GPUs, one term

    Mix families under one reservation and one invoice line, each at its own reserved rate.

FAQ

Clusters, answered.

Size, network, storage and billing, without the sales deck.
How is a cluster different from one Cloud Server?
A Cloud Server is one machine with eight GPUs on NVLink. A cluster joins several of those machines over a fast network so one job can use all of their GPUs, which is what large training runs and big-model serving need.
How small and how large can a cluster be?
Self-serve clusters run from 2 to 8 servers, so 16 to 64 GPUs. Larger fabrics are planned with the team as a Reserved Cluster.
How are clusters billed?
Each server in the cluster is metered per second at its normal hourly rate. Stop the cluster and every node stops billing at the same moment.
What bandwidth do nodes get between them?
Every GPU has its own 400 Gb/s port into a non-blocking leaf and spine fabric with RDMA, so collective operations do not compete with storage traffic.
Is my cluster network isolated from other customers?
Yes. Each cluster gets its own fabric partition. There is no route from your partition to another tenant's, and your inter-node bandwidth is not shared.
Where do large models and checkpoints live?
On a network volume in the same region, mounted on every node at boot. Local NVMe on each server stays available for scratch data.
Can I connect a cluster to my own cloud account?
Yes, over mutual TLS at the application layer or through a private interconnect on a Reserved Cluster. Data stays in the region you picked.
Do clusters support Kubernetes?
Clusters ship with Slurm and plain SSH today. Bring your own orchestrator on top if you need it; Kubernetes support is on the roadmap.
Is there a minimum term?
Not for Instant Clusters: create one, train, stop it. Reserved Clusters start at three months.

Customers

For runs that need more than one node.

Pre-training, large fine-tunes and big-model serving all outgrow eight GPUs at some point. Clusters are where they go next.
Placeholder

Placeholder quote. An AI app team says, in one or two plain sentences, what changed after moving production traffic to Nexpod.

Customer nameEngineering lead, AI app
  • Customer 01
  • Customer 02
  • Customer 03
  • Customer 04
  • Customer 05
  • Customer 06
  • Customer 07
  • Customer 08
Customer logosPlaceholder

Start now

Your next run, on 64 GPUs.

Two H200 servers start at €64.00 per hour together. Add nodes when the run needs them and stop the whole cluster in one command.

Create a clusterBook a demo
On-demand ratesEUR · excl. VAT
  • B300€6.00/GPU·hr
  • B200€5.00/GPU·hr
  • H200€4.00/GPU·hr
  • H100€3.00/GPU·hr