Skip to main content
Krystal

NVIDIA GPU cloud servers
you can attach to any VM

Krystal Cloud lets you attach an NVIDIA GPU, a T4 or L4, to any virtual machine on demand. Spin up a GPU cloud server for machine learning, inference, image generation or video rendering, then detach the GPU when the job’s done. You’re billed by the second, and only while a GPU is attached.


NVIDIA GPU cloud servers you can attach to any VM

On Krystal Cloud, a GPU attaches to a virtual machine rather than running as a separate instance. You configure a VM with the CPU, RAM and storage you need, add a T4 or L4 when there’s GPU work to do, and detach it afterwards. The VM keeps its configuration the whole time, so you effectively have a GPU virtual server for as long as you need one.

Because GPUs attach and detach on demand, you can hold one for a training run and release it the moment the job finishes. You only pay for the time it’s attached.


T4 or L4: which NVIDIA GPU to attach

NVIDIA T4

16GB of GPU memory. A solid starting point for inference, prototyping, lighter training and video transcoding.

NVIDIA L4

24GB of GPU memory. Room for larger models, higher inference throughput and heavier generative or video work.

If you're prototyping or serving a smaller model, start with a T4. Move to an L4 when your model or batch size needs more than 16GB of GPU memory.

What you can run on a cloud GPU

Model training

Train on a T4 for smaller datasets, or an L4 when you need more memory per pass. Attach an NVIDIA GPU for the run and detach it when training finishes.

Inference and serving

Serve a trained model behind an API. Many inference workloads run well on a T4, with the L4 giving more throughput for busier endpoints.

Image generation

Run diffusion and other generative image models, where GPU memory sets how large a batch you can process at once.

Video rendering and transcoding

Encode, transcode and render on GPU rather than CPU, which clears the queue faster for media-heavy pipelines.

Pay only while a GPU is attached

GPUs are billed by the second, and only for the time they’re attached to a VM. Attach one for a few hours of training and you pay for those hours, nothing more. A monthly cap means a permanently attached NVIDIA GPU never costs more than its listed monthly price. The VM it’s attached to is billed separately on its own per-second rate.


What’s behind a GPU cloud server

A GPU cloud server runs on the same platform as every other Krystal Cloud instance. This includes self-healing Block Storage with three copies of your data, automatic failover if hardware fails, and DDoS mitigation included. Each instance includes 1TB of external transfer, pooled across your organisation, with free inbound transfer.

Learn more about Krystal Cloud Compute

Cloud GPU Pricing

GPUs are billed by the second while attached, on top of your VM’s own price. You only pay for what you use, and you can cancel whenever you want. Full pricing is on the cloud pricing page.

  • Nvidia T4

    Performance
    8.1 TFLOPS (FP32)
    Tensor Cores
    320
    CUDA Cores
    2,560
    Memory
    16GB
    Bandwidth
    320GB/s
    Price
    $201.60/mo$0.2762/hr
  • Nvidia L4

    Performance
    30.3 TFLOPS (FP32)
    Tensor Cores
    300
    CUDA Cores
    7,424
    Memory
    24GB
    Bandwidth
    300GB/s
    Price
    $268.80/mo$0.3682/hr
  • Nvidia A100

    Performance
    312 TFLOPS (TF32)
    Tensor Cores
    432
    CUDA Cores
    6,912
    Memory
    80GB
    Bandwidth
    1,935GB/s
    Price
  • Nvidia H100

    Performance
    989 TFLOPS (TF32)
    Tensor Cores
    528
    CUDA Cores
    16,896
    Memory
    80GB
    Bandwidth
    3.35TB/s
    Price
  • Nvidia H200

    Performance
    989 TFLOPS (TF32)
    Tensor Cores
    528
    CUDA Cores
    16,896
    Memory
    141GB
    Bandwidth
    3.8TB/s
    Price

Why choose Krystal as your GPU cloud provider

Illustrated rainbow iridescent crystal gem with trust badges showing UK Based and 100G Backbone network capacity
  • UK-based engineering support

    Our support team are engineers based in the UK. When you get in touch, you're speaking to someone who understands the platform and can work with you to get it sorted.

  • Pay for what you use

    Compute is billed by the second. You only pay for what you use, and you can cancel at any time.

Illustrated teal crystal gem surrounded by trust badges: ISO 27001 Certification, Certified B Corporation, 1% for the Planet, and over 5 million trees funded
  • Certified security

    Krystal holds ISO 27001 certification for information security and Cyber Essentials Plus, the UK government's cybersecurity certification.

  • Matched with renewable energy

    We match our energy use with renewables, so enough renewable energy is produced to cover what we consume. Krystal is a Certified B Corp and 1% for the Planet member, with over five million trees funded through our reforestation programme.

Start building today

Sign up and you’ll be up and running on Krystal Cloud in less than a minute.

Cloud Enquiry

Take the first step

Marketing preferences

General enquiries

Learn more about our cloud services

Pricing

Transparency is key. We're upfront with our pricing structure meaning you always know how much you'll pay.

View pricing

Documentation

Read our docs and see how you might use Krystal Cloud in your organization.

User documentation

B Corp Certified

Rest easy knowing your cloud runs on 100% renewable energy from a certified B Corporation.

Learn more

FAQs

The most-asked questions about our services and billing. Can’t find the answer? Check our Knowledge Base or drop us a line on .