Introducing GPU VMs on NVIDIA AI infrastructure in Sines (EU): flexible, cost-efficient compute for AI workloads

Some AI jobs require the full power and predictability of dedicated bare metal clusters. Others need something more agile: compute that can be sized up or down quickly, used for a burst of experimentation, powered down when idle, and spun back up when the next training run begins.

To support agile AI workloads, we’re introducing a new addition to Gcore’s AI infrastructure: GPU Virtual Machines on NVIDIA Hopper H100, launching first in Sines-3, our sovereign AI region in Portugal. This brings a new level of flexibility to the same infrastructure with NVIDIA Quantum InfiniBand fabric that is already available as Bare Metal on GPU Cloud.

Like with GPU Cloud bare metal and Gcore Everywhere Inference, you can deploy GPU VMs in just three clicks.

A more adaptable way to use NVIDIA Hopper GPUs

GPU VMs give teams a way to tap into Hopper performance without committing to long-running hardware. Instead of provisioning a full server, you can start with a single NVIDIA H100 GPU, scale to two or four, or jump to an eight-GPU VM when your workload requires it.

This makes it easier to match GPU capacity to the stage of your project, whether you're testing a training script, running a few fine-tuning cycles, or pushing through a heavier training window.

Everything runs on the same AI-optimized infrastructure as Gcore bare metal GPU instances, including high-bandwidth NVIDIA Quantum InfiniBand networking. The result is a VM environment that still feels like working on serious AI hardware, just with more elasticity built in.

Cutting idle costs without complexity

One of the biggest advantages of GPU VMs is how they behave when they’re not in use. Many AI workloads come in waves: bursts of experimentation, rapid iteration, then a quiet period while teams review results or prepare the next dataset.

With GPU VMs, when you power off the instance, GPU billing pauses automatically. Your volumes, IPs, and configuration remain intact, but the GPU meter stops running. You only pay for optional additional volume storage and IPs while paused. When you’re ready to start work again, simply restart the VMs as required without needing to set up or reconfigure.

For teams experimenting, fine-tuning, or running training jobs on a schedule, this can significantly reduce operational overhead. You don’t need to redesign your workflows to fit spot markets or preemptible machines; you simply power the VM off when it’s not needed and resume when you’re ready, subject to available capacity.

When GPU VMs make the most sense

Teams already using our bare metal GPU Cloud often tell us they handle two distinct types of workloads. On one side are long, uninterrupted training cycles that need maximum consistency and dedicated hardware. On the other are the day-to-day pieces of AI development: testing a hypothesis, trying out a new architecture, running a few epochs to validate an idea, or preparing a model for deployment.

GPU VMs are designed for the latter. They’re ideal when:

• You want to iterate quickly without reserving hardware for days at a time
• Your training or fine-tuning runs aren’t continuous
• You’re validating training code before scaling it to dedicated nodes
• You need short-term GPU capacity that fits within an existing workflow
• You want to serve or test inference traffic before moving it to Everywhere Inference

Whether you’re an early/growth-phase AI startup looking for performant GPUs without the high fixed costs, an EU R&D lab needing sovereign infrastructure for burst PoCs and experiments, or a research institution seeking to run short-term, high-intensity fine-tuning runs on a budget, Gcore GPU VMs deliver the flexibility and cost efficiency you need.

Many teams use both Bare Metal and VM options side by side: VMs for experimentation and agile development, and bare metal for the heavy lifting once the model is ready for large-scale training. Gcore makes this easy with 3-click BM and VM deployments in a single platform.

Part of a larger AI infrastructure roadmap

Alongside this new VM solution, we are also now offering spot bare metal GPUs. Spot instances are ideal for interruptible, high-throughput jobs that require uncompromising performance at a cost-effective price. With Gcore, you can combine GPU solutions.

If your team is exploring how to mix VMs and bare metal, needs help calibrating the right GPU shapes, or wants to test workloads ahead of a migration, we're here to help.

Deploy a workload on Gcore GPU VMs