Skip to content
State of GPU Compute 2026
B300 live in EU-AMS-1

Whole 8-GPU servers, by the hour.

NVIDIA B300, B200, H200 and H100 nodes with no other tenant on board. From €3.00 per GPU-hour, billed by the second.NVIDIA B300, B200, H200 and H100 nodes with no other tenant on board. Run them on demand, behind a Serverless endpoint or as a cluster, from €3.00 per GPU per hour, billed by the second.

Deploy a serverSee pricing
Exploded view of a Nexpod node: an 8U server with eight NVIDIA B300 GPUs on an HGX baseboard, an NVLink switch tray above it, and data paths to the EU-AMS-1, EU-FRA-1 and US-NYC-1 regions.
SXM GPUs
8× B300
NVLink
1.8 TB/s
per GPU-hr
€6.00
console.nexpod.ai/servers
Servers2 running
Deploy
Name€/hr
  • llm-train-01
    NVIDIA DGX B300 · EU-AMS-1
    €48.00
  • inference-eu
    ASUS ESC NB8-E11 · EU-FRA-1
    €40.00
  • finetune-h200
    NVIDIA DGX H200 · EU-AMS-1
    €32.00
  • batch-render
    ASUS ESC N8A-E12 · US-NYC-1
    €24.00
Billed per second · EUR€88.00/hr running
families
4families
B300 · B200 · H200 · H100
servers
20servers
in the catalog, specs public
regions
3regions
Amsterdam, Frankfurt, New YorkEU-AMS-1 · EU-FRA-1 · US-NYC-1
second
1second
billing granularity

From the founders

Why we only rent out whole servers.

Shared GPUs make benchmarks lie. A full node gives you the same throughput at 3 a.m. and at peak, and a bill you can predict.

Read our story
RACK AMS-1 · R07 1 of 4 yours
U42–35
GIGABYTE G894-ZD3-AAX7
rented
U33–24
NVIDIA DGX B300
yours
U22–13
NVIDIA DGX B200
rented
U11–4
GIGABYTE G894-AD1-AAX5
free

What's new

ChangelogOct 24, 2026

nexctl 2.0 puts every product in one CLI

One grammar for servers, endpoints and clusters, with JSON output.Servers, endpoints and clusters now share one grammar: nexctl <resource> <verb>, with JSON output and a --watch flag that follows a deploy to ready.

ReleaseOct 22, 2026

B300 servers open in Amsterdam

DGX B300 and GIGABYTE B300 in EU-AMS-1, on demand or reserved.NVIDIA DGX B300 and GIGABYTE B300 chassis can now be deployed in EU-AMS-1, on demand or on a 12-month reserved term.

Nexpod research2026
State of
GPU Compute
Rental patterns, utilisation and
the B300 effect on training
2026 editionnexpod.ai/reports
2026 edition

Report

Free PDF

7 chapters · 38 pages

State of GPU Compute 2026

What teams rent, how long they keep it, and where B300 changes the maths for training and inference.

  1. 01Where the GPU-hours wentp.4
  2. 02The B300 yearp.9
  3. 03Inference takes the leadp.14
  4. 04How big clusters really arep.19

7 chapters · 38 pages

Get the report

Platform

One account from first job to fleet.

Servers, Serverless and Clusters on one console, API and invoice.Cloud Servers, Serverless and Instant Clusters share one console, one API and one invoice. Move a workload between them without rewriting it.

Boot a whole server in about a minute.

Pick B300, B200, H200 or H100, choose an image and connect over SSH. All eight GPUs, the CPUs and the NVMe are yours.

Explore Cloud Servers
nexctl — eu-ams-1
$ nexctl servers create --family b300
→ reserving 8× B300 in eu-ams-1
→ image nexpod/pytorch:2.9-cuda13
→ health checks 8/8 passed
✓ ready in 47s · billed per second
$ ssh root@ams-04.nexpod.cloud
$ nvidia-smi -L | wc -l
8
$

Workflow

Train and serve on the same hardware.

One image, one API and one invoice, from launch to scale.Launch, train, serve and scale with one image, one API and one invoice. Nothing to port between stages.

Deploy a serverStart building
  • 01
    $ nexctl servers create --family h200
    → reserving 8× H200
    ✓ ready in 52s

    Launch

    ~1 min

    A dedicated 8-GPU server in the region you pick, reachable over SSH.

  • 02
    train loss0.60

    Train

    hours to weeks

    Fine-tune or pretrain with your own container, framework and dataset.

  • 03
    POST/v1/endpoints/llm-eu/run
    status
    200
    p50
    182 ms
    workers
    3

    Serve

    one push

    Wrap the model in a handler and publish it as a Serverless endpoint.

  • 04
    workers2 → 14

    Scale

    per request

    Workers follow the queue, from zero up to what the traffic needs.

  • Health-checked hardware

    3/3 passed

    GPU, NVLink and disk checks before you get the keys.

  • Queueing built in

    0 dropped

    Requests wait for a free worker. No broker to run.

  • Logs and metrics, live

    streaming

    stdout, GPU load and memory per device, via nexctl.

Serverless

Inference billed only for busy seconds.

Endpoints start workers when requests arrive and bill per second while they run. When the queue is empty, the count drops to zero.

Explore Serverless
  • Requests / s · llm-eu
    24 h
    +6 workers
    −5 workers
    Active workers
    max 13
    Autoscaling

    Workers follow the queue

    Set a target wait time; the endpoint adds workers as requests line up and removes them as it drains.

    How autoscaling works
  • p50p95
    SLO 300 ms
    p50 179 ms
    p50 now
    179 ms
    p95 now
    290 ms
    target
    300 ms
    Start latency

    Weights wait on local NVMe

    New workers load the model from the node's own disk, so p50 stays flat through a burst.

    Inference on Nexpod
  • Workers per hour running zero
    00:0012:0024:00
    00:00 – 06:00 billed€0.00
    Scale to zero

    No requests, no bill

    Minimum workers can be zero. Busy time is billed by the second, from €0.00083 per GPU.

    Serverless pricing
  • Volume mountsEU-AMS-1 · 5 workers
    w-01B200
    w-02B200
    w-03B200
    vol-weights
    2 TB · NVMe
    train-018× B300
    w-04H200
    Storage

    Network volumes in every region

    Mount one volume on many workers. Weights and datasets stay in the region you chose between runs.

    About volumes

Servers

4 GPU families. 20 servers.

Every server ships with 8 GPUs. Rates are per GPU per hour on demand; a whole server costs 8 times that.

Browse all serversCompare two
8× B300 · 288 GB each
GPU familyOn demand
Browse all 20 servers

EUR, excl. VAT. Reserved 12-month rates from €2.00/GPU·hr.

Case studies

What runs on a dedicated node.

Pretraining, inference and batch work, one write-up each. Sample cards until customers approve their stories.

All case studies
PretrainingEU-AMS-1
8× B300Instant Clusters
Placeholder

Placeholder. The studio explains how starting every new model on an 8-node cluster changed their release plan.

Video generation studioRead case study
Inference2 EU regions
8× H100Serverless
Placeholder

Placeholder. Two sentences on traffic that follows office hours and workers that scale to zero at night.

Speech AI labRead case study
Batch jobsEU-FRA-1
job-11100%
job-12100%
job-1386%
job-1454%
job-1521%
job-160%
8× H100Cloud Servers
Placeholder

Placeholder. One line on clearing the morning job queue with four reserved servers.

Protein design labRead case study
3 sample studies
  • Pretraining8× B300
  • Inference8× H100
  • Batch jobs8× H100
All case studies
Swipe or use the arrows

Choosing hardware

Match the server to the workload.

Start with the job, then compare memory, interconnect and rate.Compare memory, interconnect and hourly rate across every server before you commit to one. Start with the job you need to run.

Compare GPUs

Enterprise

Ready for your security review.

Dedicated hardware, EU regions and reserved terms.Dedicated hardware, EU regions and reserved terms, with one engineer who knows your setup by name.

Explore EnterpriseTalk to sales
  • region pinning2 live · 1 planned
    EU-AMS-1Amsterdam
    EU-FRA-1Frankfurt
    EU-HEL-1Helsinkisoon

    EU data residency

    Pin a workload to Amsterdam or Frankfurt. Weights, volumes and logs stay there.

  • dgx-b300-04account: ml-research
    other tenants0

    Single tenant

    The server, its GPUs and its disks belong to one account. Nothing to share, nothing to leak.

  • €/GPU·hr on demand 12 mo
    B300€5.00
    B200€4.00
    H200€3.00
    H100€2.00

    Reserved capacity

    Fix a rate for 12 months: €5.00 per GPU-hour for B300, in blocks of whole nodes.

Explore Enterprise

FAQ

Short answers first.

Billing, regions and what dedicated means here. The rest is in the docs.

Read the docs
Do I really get the whole server?
Yes. A Cloud Server is one physical machine with 8 GPUs, plus its CPUs, RAM and local NVMe. No other account runs on it while you hold it.
How is usage billed?
Per second, from boot until you stop the server, at the hourly GPU rate. On-demand B300 is €6.00 per GPU-hour, so a full server is €48.00 per hour. Invoices are in euros.
Which regions are live?
EU-AMS-1 (Amsterdam), EU-FRA-1 (Frankfurt), US-NYC-1 (New York). EU-HEL-1 (Helsinki) is next.
When should I pick Serverless over a Cloud Server?
Pick a Cloud Server when a job keeps GPUs busy for hours. Pick Serverless when traffic comes in bursts: workers start on request and stop when the queue is empty.
Can I reserve capacity for a year?
Yes. A 12-month term costs €5.00 per GPU-hour on B300 and €2.00 on H100. Sales can hold blocks of whole nodes for a cluster.

Your server is one command away.

No contract and no minimum term. Stop the server and the meter stops with it.

Deploy a serverGet startedTalk to salesBook a demo
$nexctl servers create --family b300