Cloud Servers

A full 8‑GPU server, live in a minute.

Pick B300, B200, H200 or H100, choose a region and connect over SSH. Nothing on the machine is shared, and the meter runs per second, in euros.

Deploy a server See all 20 servers

Deploy

Four picks, then a root shell.

Family, region, server model and image. The summary shows the hourly rate and the price per second before anything starts billing.

console.nexpod.ai/deploy

ml-research

1GPU family8 GPUs per server

B300

288 GB HBM3e

€48.00/hr

B200

180 GB HBM3e

€40.00/hr

H200

141 GB HBM3e

€32.00/hr

H100

80 GB HBM3

€24.00/hr

2Region

EU-AMS-1EU-FRA-1EU-HEL-1US-NYC-1

3Server model1 of 8

NVIDIA DGX B300

Boot · dgx-b300-0447 s

Reserve node
Attach volume
Pull image
Boot + checks

0s15s30s45s60s

Running in under a minute

Steps overlap, and all eight GPUs pass health checks before SSH opens.

Regions3 live · 1 planned

EU-AMS-1

EU-FRA-1

EU-HEL-1soon

US-NYC-1

3 regions, one rate card

Amsterdam, Frankfurt and New York today, Helsinki next. One price everywhere.

dgx-b300-04 running

00:03:12

8× B300: €48.00/hr
Per second: €0.01333
Billed so far: €2.56

Metered per second

Stop the server and the meter stops too.

4
GPU families: 20
server models: 3
live regions: 1 s
meter granularity

In plain words

What is a Cloud Server?

One physical machine with eight NVIDIA GPUs, rented by the hour and run by you as root. Pick it when a job wants the full node, its local disks and a fixed hourly price: model training, steady serving, rendering or a service that runs for weeks.

Comparing providers?AWS Google Cloud Azure Oracle Cloud

Rates

Four families. 20 servers.

The rate follows the GPU family, not the chassis. Pick a model for its CPU, memory or local storage and pay the same per hour.

Browse servers Talk to sales

EUR · excl. VAT · metered per second

B300288 GB HBM3e per GPU8 models

B200180 GB HBM3e per GPU8 models

H200141 GB HBM3e per GPU2 models

H10080 GB HBM3 per GPU2 models

Use cases

One server type. Four kinds of work.

Whole nodes suit work that wants steady throughput and no neighbours: serving, training, agents and heavy batch jobs.

dgx-b300-04 · per GPUtensor parallel 8

avg util95%HBM per GPU270/288 GBp95241 ms

Inference with steady latency

Serve a large model from eight GPUs on NVLink, with no noisy tenant next to you.

NODEft-run-07 · epoch 2/3

Fine-tuning and training

Multi-day runs on one node, with checkpoints on a volume that outlives it.

KV cache · 16 slots per GPU96 sessions warm

gpu013/16
gpu112/16
gpu212/16
gpu311/16
gpu413/16
gpu512/16
gpu611/16
gpu712/16

Agents with long sessions

Keep models warm in memory for tool-heavy runs that last minutes, not milliseconds.

render-0917 · 8 GPUs2,400 frames

gpu0
gpu1
gpu2
gpu3
gpu4
gpu5
gpu6
gpu7

0 minnow · 24 min38 min

Rendering and batch

Split a job into shards across eight GPUs and stop the server when the queue is empty.

Developer tools

Run the whole fleet from code.

Every button in the console is an API call first. Script it from Python or TypeScript, drive it from your shell with nexctl, or let CI roll out new images.

Get an API key Read the docs

import nexpod
 
# the key comes from NEXPOD_API_KEY
client = nexpod.Client()
 
server = client.servers.create(
    family="B300",
    region="eu-ams-1",
    image="nexpod/pytorch:2.9-cuda13",
)
 
server.wait_until_ready()
print(server.ssh)

1import nexpod
2 
3# the key comes from NEXPOD_API_KEY
4client = nexpod.Client()
5 
6server = client.servers.create(
7    family="B300",
8    region="eu-ams-1",
9    image="nexpod/pytorch:2.9-cuda13",
10)
11 
12server.wait_until_ready()
13print(server.ssh)

API · SDKs

One REST API, two SDKs

Create, list, resize and stop servers from Python or TypeScript. Keys stay in your environment.

nexctl — zsh

$ nexctl servers list

NAME GPUS STATUS

llm-train-01 8×B300 running

inference-eu 8×B200 running

$ nexctl servers ssh llm-train-01

✓ root@ams-04.nexpod.cloud

$ nexctl servers logs llm-train-01 -f

14:01:58 step 18400 · loss 0.612

14:02:04 step 18410 · loss 0.609

$ nexctl servers stop llm-train-01

✓ stopped · meter off at 14:02:17

$

CLI

nexctl in your shell

List, create and stop servers without leaving the terminal.

deploy · mainv15

git pushmain · 3f9a1c2

buildimage · 2m 04s

v15 livefleet · 4 servers

rollback to v14nexctl rollback

CI

Roll out from your pipeline

Push to main, build the image, swap it onto your servers one by one. One command back if it misbehaves.

Storage

Fast local disks. Volumes that stay.

Each server has its own NVMe for scratch data. Network volumes hold what must survive a stop. Moving data in or out costs nothing.

Local NVMe · as listed per server

ServerLocal storageGPU

Network storage

TypeWhat it isRate

Network volumeMounts on any server in the same region and outlives itConfigurable on request
SnapshotsPoint-in-time copy of a network volumeConfigurable on request
Ingress and egressData moving into or out of NexpodNo fees

Per GPU · hour on demand reserved

B300 · 288 GB HBM3e−17%
€6.00
€5.00
B200 · 180 GB HBM3e−20%
€5.00
€4.00
H200 · 141 GB HBM3e−25%
€4.00
€3.00
H100 · 80 GB HBM3−33%
€3.00
€2.00

Reserved

Commit for a year. Pay less every hour.

Reserve servers for 12 months and each GPU-hour costs €1.00 less, on every family. Same hardware, same per-second meter, one fixed line on the monthly invoice.

Ask for a reservation Compare rates

FAQ

Cloud Servers, answered.

Short answers about hardware, billing and data. Anything else goes to the team behind the console.

Do I really get the whole server?

Yes. A Cloud Server is one physical machine: eight GPUs, its CPUs, memory and local NVMe. While it is yours, no other customer has a process on it.

Which GPUs and server models can I rent?

Four NVIDIA families (B300, B200, H200 and H100) across 20 server models from NVIDIA, ASUS, GIGABYTE and Supermicro. Every model has its own spec page under Servers.

How is a Cloud Server billed?

Rates are listed per hour and metered per second, from the moment the server reports ready until you stop it. You get one invoice in euros with VAT on its own line.

Can I run my own container image?

Yes. Start from a Nexpod base image with CUDA and PyTorch, or pull any image from a public or private registry. You also get root over SSH.

What happens to my files when I stop?

Local NVMe is wiped when the server is released. Keep datasets and checkpoints on a network volume: it stays in the region and mounts on your next server.

Is there a minimum rental period?

No. Ten minutes of work costs ten minutes of server time. There is no setup fee and no notice period.

Where are the servers?

EU-AMS-1 (Amsterdam), EU-FRA-1 (Frankfurt), US-NYC-1 (New York), with EU-HEL-1 (Helsinki) planned. Rates are the same in every region.

Is there a lower rate for long runs?

Yes. A 12-month reservation takes €1.00 off every GPU-hour, on every family. For blocks of servers, talk to sales.

Customers

Teams that need the whole node.

Long training runs, steady inference and heavy batch work all want the same thing: predictable throughput and nothing else on the machine.

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

Start a server. Stop it whenever.

No card needed to look around the console. Once a server runs, you pay per second and stop whenever you like.

Deploy a server Talk to sales

On-demand ratesEUR · excl. VAT

B300· 288 GB HBM3e€6.00/GPU·hr
B200· 180 GB HBM3e€5.00/GPU·hr
H200· 141 GB HBM3e€4.00/GPU·hr
H100· 80 GB HBM3€3.00/GPU·hr