Skip to content
State of GPU Compute 2026

Use case · Fine-tuning

Fine-tune on a whole node.

Rent a full 8-GPU server for a LoRA pass or a week of full fine-tuning. The GPUs, CPUs and disks are yours alone, checkpoints land on network volumes, and billing ends when the server stops.
Get startedTalk to sales

8× H200 · €32.00 /hrroot over SSHstop any time

Exploded view of a Nexpod node: an 8U server with eight NVIDIA H200 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.
runs / ft-mistral-07
training
step 6,400 / 10,000ETA 1h 52m
Train loss
0.61
96
94
97
91
95
93
98
92
8× H200 · EU-AMS-1ckpt-6000 saved

Teams training on Nexpod

Placeholder
  • Customer
  • CUSTOMER
  • customer
  • Customer Co
  • customer.ai
  • CUSTOMER

Customer stories

Runs that shipped a model.

Domain LLMH200
Customer
PlaceholderCloud Servers

Placeholder quote. A customer explains, in two plain sentences, what changed once their domain LLM workload moved to Nexpod.

Customer nameRole, Company
Read case study
Vision modelB200
Customer
PlaceholderCloud Servers

Placeholder quote. A customer explains, in two plain sentences, what changed once their vision model workload moved to Nexpod.

Customer nameRole, Company
Read case study
Speech modelB300
Customer
PlaceholderInstant Clusters

Placeholder quote. A customer explains, in two plain sentences, what changed once their speech model workload moved to Nexpod.

Customer nameRole, Company
Read case study
RecommenderH100
Customer
PlaceholderCloud Servers

Placeholder quote. A customer explains, in two plain sentences, what changed once their recommender workload moved to Nexpod.

Customer nameRole, Company
Read case study

Instant Clusters

Outgrow one node? Add seven more.

Instant Clusters join up to eight 8-GPU servers on one fabric. Your training script sees 64 GPUs and runs from the same image it used on one.

Explore Instant Clusters
Instant Cluster · H200
64GPUs
8 nodes × 8 GPUs
n0 = your first node
n0
n1
n2
n3
n4
n5
n6
n7
one fabric · torchrun across all nodes

01Hardware

The whole server, nothing shared.

Eight GPUs on NVLink, the host CPUs, the RAM and the local NVMe are reserved for your run from boot until you stop it.

Pick the family
B300, B200, H200 or H100, with every server's specs listed in the catalogue.
Root over SSH
Your keys, your shell, your CUDA stack. Install what the run needs.
NODENVIDIA DGX H200
8/8 online
0
H200
1
H200
2
H200
3
H200
4
H200
5
H200
6
H200
7
H200
NVLink
all-to-all

02Data

Checkpoints that outlive the node.

Write checkpoints to a network volume in the same region. Stop the server, start another one, and resume from the last saved step.

Network volumes
Mount the same volume on any server in the region, one at a time or many.
No egress fees
Pulling data out of Nexpod network storage is not billed.
run ft-mistral-07 · two servers, one volume
step
02k4k6k8k10k
h200-ams-01
stopped · Tue 21:14
billing stops
h200-ams-02
running · Wed 08:02
ft-checkpoints
network volume
ckpt-4000 · resumed here ckpt-6000 · latest ckpt-6400 · writing0 steps lost · egress €0.00

03Scale

From one node to a cluster.

When a run needs more than eight GPUs, the same container runs on an Instant Cluster and torchrun spreads it across the nodes.

Fast interconnect
Nodes in a cluster talk over a non-blocking fabric built for all-reduce.
Same image
Move from one node to eight without rebuilding your container.
Throughput · tokens/s
879k at 64 GPUs
8 nodes
816243240485664
GPUs · one more 8-GPU node per bar

Developer tools

Train from the scripts you have.

Keep your PyTorch, DeepSpeed or Axolotl config. Launch it through the API or nexctl, stream the logs, and pull the weights when the run ends.

Get startedRead the docs
import nexpod
 
client = nexpod.Client()
 
server = client.servers.create(
    family="H200",
    region="eu-ams-1",
    image="nexpod/pytorch:2.9-cuda13",
    volumes=["ft-checkpoints:/ckpt"],
)
 
server.run("torchrun --nproc-per-node 8 train.py --save /ckpt")
server.stop_when_idle(minutes=10)

SSH and Jupyter

Connect over SSH or open a notebook on the node in the browser.

$ nexctl ssh ft-h200

Volumes API

Create, mount and snapshot checkpoint volumes from code.

$ nexctl volumes snapshot ft-checkpoints

Runs from CI

Start a run on merge and stop the server when the job exits.

$ nexctl servers stop --on-exit

More use cases

Same servers, other workloads.

llm-eu · req/s1.24k

Inference

Serve models with steady latency under load

run_8f21c · trace6.9 s

Agents

Bursty, long-running agent workloads

render-batch-11475%

Compute-heavy tasks

Rendering, simulation and batch jobs

Start with fine-tuning

Start the run tonight. Stop it when it converges.

Rent a full 8-GPU server, mount your checkpoint volume and launch. When the loss flattens, stop the server and the bill stops too.

Get startedTalk to sales
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