From a single GPU to a whole cluster

For teams that run their own models: launch a virtual machine with root access, train without managing machines, or reserve a cluster with our engineers.

Which option fits your workload

Reserve capacity for the long run, or launch a GPU when you need one.

  • Reserved

    Capacity held for you

    GPUs or a cluster, priced per contract.

    Best for: Long training runsSteady inferenceMulti-node jobs

    Show available capacity now
  • On demand

    Start and stop when you like

    Root access. Billed per second, no subscription.

    Best for: ExperimentsShort jobsCustom stacks

    See GPUs and prices
  • Spot

    Same GPUs, lower rate

    About 50% less. Can be reclaimed at any time.

    Best for: Checkpointed jobsBatch evaluationsReruns

    Compare spot prices
  • Serverless training

    No machine to look after

    Your container, on the GPU you choose.

    Best for: Fine-tuningBatch jobsRepeatable runs

    See how it works

Reserved GPUs and clusters

Plan a reservation

Your answers pre-fill the booking form.

GPU
Backup GPU
GPUs in total
Start
For how long
Region

Current capacity

GPUGPUsRegionEarliest startMinimum termQuote
RTX PRO 600040 GPUsWestern EuropeNow1 monthGet a quote
B200256 GPUsCentral EuropeDecember 202636 monthsGet a quote
B300512 GPUsNorth AmericaDecember 202636 monthsGet a quote

GPU virtual machines

Launch a virtual machine with root access in our European data centres and install what you need. You pay per second while it runs.

GPU prices and capacity

GPU and memoryPrice per GPU hour, spot and on demandLaunch
NVIDIA B300288 GBOn demand $12.79 per GPU hour, spot $6.39, 50% less.Launch
NVIDIA B200192 GBOn demand $9.99 per GPU hour, spot $4.99, 50% less.Launch
NVIDIA H200141 GBOn demand $7.09 per GPU hour, spot $3.59, 49% less.Launch
NVIDIA H10080 GBOn demand $5.39 per GPU hour, spot $2.69, 50% less.Launch
NVIDIA A10080 GBOn demand $2.59 per GPU hour, spot $1.29, 50% less.Launch
NVIDIA L40S48 GBOn demand $2.19 per GPU hour, spot $1.09, 50% less.Launch

Size a model

See the memory a Hugging Face model needs and how many requests it serves.

Try

Context length

Tokens per request

Requests at once

Users or streams at the same time

Serverless training

Train models without managing machines: package your training code in a container, choose a GPU and submit the job.

  1. You

    Package your code

    Put your training code and its dependencies in a plain Docker container.

  2. You

    Choose a GPU and submit

    Pick the GPU your model needs, check its price, then submit the workload.

  3. We

    We run it on that GPU

    The job runs on the hardware you chose. When it finishes, you inspect its output.

train.sh
# 1  Package your code
$ docker build -t your-team/train .
$ docker push your-team/train# 2  Choose a GPU and submit
$ lyceum docker run your-team/train \
    -m gpu.h100 -c "python train.py"# 3  We run it on that GPU
#    The output streams here until
#    the run ends. Add -d to detach.
Example image and command · Lyceum CLI