GPU Servers
Give your AI and graphics workloads exactly the GPU they need, with no cards to buy
Add just the GPU memory you need to a virtual server in minutes, while driver compatibility and GPU licensing stay with Nesnet. Your model and data never leave Türkiye, and the invoice shows the profile you chose plus the server's resources.
Why now
- You want to add an AI feature to your product, but buying a GPU card, a server to carry it and the cooling to match takes months and a big upfront budget, and nobody knows yet how much it'll be used. With GPU Servers you add the GPU memory you need to a virtual server and remove it when the job is done.
- Your model is small, but the only option on the table is a full card. An inference service or a light compute job doesn't need the whole card; a profile worth a half, a quarter or an eighth of it is enough, and that's what you pay for.
- Legal won't allow customer documents to be sent to an AI service abroad. Run the model on a GPU server in your own network, in the Equinix Istanbul (Dudullu) and Türk Telekom Ankara data centres, and the data stays inside your boundary.
- Every designer is asking for an expensive graphics card, and large model files keep getting copied between laptops. A Windows server with a graphics profile runs CAD and 3D software centrally, and the files live in one place.
- Your team loses days to GPU drivers, licence portals and mismatch errors. Nesnet runs the licence server and the Windows template ships with the driver and licence in place, so your team works on its model instead of the setup.
What changes
| Today | With Nesnet |
|---|---|
| A GPU card, a server and cooling mean a big upfront investment and months of waiting. | The GPU profile you choose is added to your virtual server, and your first GPU server is up within minutes. |
| Even a small model means buying a whole card that sits idle most of the time. | Pick the whole card, or a half, a quarter or an eighth of it, and pay only for that profile. |
| A mismatch between driver version and licence leaves the GPU unusable. | Keeping the driver in line with the platform and handing out licences is Nesnet's job, and the Windows template comes ready. |
| Documents and customer data are sent to an AI service abroad. | The model and the data run inside your network, on infrastructure in Türkiye. |
| Designers are tied to powerful workstations at their desks. | CAD and 3D software run on a central graphics server, and the files stay in one place. |
From today to live: how it works
- 1
Describe your workload
Inference, graphics or fine-tuning; which operating system and how much GPU memory: you talk it through with Nesnet's team. You leave with a recommended profile and a clear view of what drives the price.
- 2
A profile is chosen
A graphics or compute profile and a GPU memory size are picked for the workload. Starting small and moving up once the need is clear is part of the plan.
- 3
The GPU policy joins your allocation
The chosen profile is published to your organisation's allocation as a GPU policy. The policy brings memory reservation and VM preparation with it.
- 4
Launch the server and load your model
On Windows, a server from the GPU template comes up with driver and licence ready; on Linux the driver and licence are configured to bring the GPU into service. Then you install your own libraries, frameworks and model.
- 5
Measure and grow
Watch GPU memory and CPU usage, and add more GPU servers to the same allocation as load grows. If the profile you need changes, you move to a new one with Nesnet's team.
What you get
A whole card or a slice of one
Choose the whole GPU card, or a half, a quarter, an eighth or a sixteenth of it. A small inference service runs on a small slice and the charge follows the profile you pick, so you never pay for GPU memory you don't use.
GPU memory sized to the job
Graphics profiles come with 3, 6, 12, 24 and 48 GB of GPU memory; compute profiles range from 6 GB to 48 GB. Pick the size your model and data fit into, and move up a profile as the need grows.
A GPU server in minutes
Once the GPU policy is published to your allocation, a GPU server is created in the portal with the same steps as any other server. In Nesnet's own measurement, organisation, network and GPU server were ready end to end within minutes.
A Windows template with driver and licence ready
The Windows Server 2022 GPU template has the driver and licence token built in. A server from that template uses its GPU without any extra install: no driver download, no version matching, no licence file to place.
GPU compute on Linux
Nesnet has verified sliced GPUs on Ubuntu 24.04 with the Linux GPU driver and Nesnet's licence server. The licence service is set to restart itself if it stops, so a GPU never slips quietly into an unlicensed state.
Model inference and AI services
Run services such as a language model, speech recognition or image classification on a compute-profile server and connect your application to it from inside the same network. User requests and answers never leave your organisation's network, and a second GPU server joins when load grows.
The libraries and frameworks you know
Workloads have run on this infrastructure with widely used GPU compute, deep-learning and linear-algebra libraries. Install your frameworks such as PyTorch or TensorFlow, your own Python environment and your models on the virtual server as you always would.
A graphics workstation in the cloud
Graphics profiles are designed for CAD, engineering and 3D visualisation software and, depending on the profile, support several high-resolution virtual displays. Designers run the software on a central GPU-backed Windows server, and only pixels travel to their device.
Graphics profiles that compute too
Graphics profiles run compute applications alongside display output. Teams that visualise by day and render or batch-process by night don't need to keep two separate servers.
Same network, same firewall
A GPU server comes up on your organisation's network behind the gateway firewall and talks to your other servers over internal addresses. If your model service has to be published, you do it for just the port you want with a single rule.
Mix GPU and regular servers
GPU and non-GPU servers run together in the same allocation: your database on an ordinary virtual server, your model on a GPU server. Expensive GPU capacity goes only to the server that genuinely needs it.
Skip the licence portal
Compute profiles use a compute licence and graphics profiles a graphics-workstation licence, and the right edition is pulled automatically from Nesnet's licence server. Licence counts, renewal dates and licence files never make it onto your team's calendar.
VM preparation handled by policy
The current virtual hardware a sliced GPU needs, the setting for addressing large GPU memory and full memory reservation are applied through the portal when a GPU policy is chosen. No ticket and no hand-edits in the virtualisation layer.
Object storage for model files
Keep training data, model weights and output files in S3-compatible Object Storage ordered from the same portal. The GPU server reads from the same infrastructure, so large files move between server and store without leaving the country.
Spread GPU servers across hardware
Use an anti-affinity rule to place two copies of a model service on different physical hosts. Placement takes GPU capacity into account, and a single hardware fault can't stop both copies at once.
Your disk is backed up like any server's
A GPU server's disk is backed up through the Backup product on the same path as other virtual servers. After a bad update you roll model weights, environment settings and code back to that moment.
Profile names you can read
Profiles are named in the portal by GPU memory and use (graphics or compute). You see exactly what you're choosing without memorising vendor codes, and you can tell your team which profile is for which job at a glance.
Why Nesnet
The licence server is Nesnet's
Nesnet runs its own licence server for GPU licences. A GPU server pulls its licence from it automatically when it starts, so you never open an account on the vendor's licensing portal or move licence files around.
Nesnet keeps drivers compatible
On a sliced GPU, the driver inside the virtual server has to match the software branch on the physical server the card sits in. Nesnet picks the template's driver for that branch and refreshes the template at every major upgrade, so a device error from a mismatched driver is never yours to chase.
Your data stays in Türkiye
GPU servers run on Nesnet's infrastructure in the Equinix Istanbul (Dudullu) and Türk Telekom Ankara data centres. Your model, your training data and what your users send in never pass through a service abroad.
GPUs on the same network as your servers
A GPU server comes up on your organisation's network, behind the same gateway firewall as your other servers. Traffic between the application and the model never touches the internet, and there's no extra provider or separate security setup.
Settings by policy, not by hand
Full memory reservation, the setting for addressing large GPU memory and placement in the GPU server group are applied through the portal as soon as a GPU policy is chosen. Your GPU server never turns into an undocumented, hand-tuned exception.
Nesnet runs its own AI work here too
Nesnet runs its own AI workloads, such as speech recognition and language-model fine-tuning, on the same GPU infrastructure. The profiles, drivers and licensing you get are the setup Nesnet uses every day.
Example scenarios
A law firm querying its documents inside its own network
Situation
A law firm wants to search and summarise thousands of contracts with a language model, but sending client documents to a service abroad is not allowed.
Outcome
The model runs on a Linux server with a 24 GB compute profile inside the firm's own network, and the documents sit in storage on the same infrastructure. Queries and answers never leave the firm's boundary.
A software team running an inference service
Situation
A software team wants to serve a model behind a new AI feature in its product. The model doesn't need a full card, but the office workstation isn't reliable enough for production.
Outcome
The model runs on a server with a 12 GB compute profile, on the same network as the application servers. When load rises a second GPU server joins the same allocation, while Nesnet keeps licences and drivers in line.
An engineering firm moving CAD to the centre
Situation
Every designer at an engineering firm has an expensive workstation, and large model files are copied from machine to machine.
Outcome
CAD software runs on Windows Server 2022 servers with graphics profiles, and the driver and licence arrive with the template. Model files live in one place, on the network where the servers are.
A contact centre transcribing its calls
Situation
A customer-service unit wants to transcribe hundreds of call recordings a day and sort them by topic, and the recordings contain personal data.
Outcome
A speech-recognition model runs on a GPU server inside the network where the recordings are kept. Transcripts go to a reporting server on the same infrastructure, and the recordings never leave Türkiye.
A data team trying out fine-tuning
Situation
A data team wants to fine-tune an open-source language model on its own data, but doesn't know how many weeks it will take or how much GPU memory it will need.
Outcome
The team starts with a small profile and moves up once the need is clear. The training code is set up not to use managed memory, and the resulting weights are stored in object storage.
Who does what
| Nesnet | You |
|---|---|
| Runs the GPU licence server and manages licence distribution. | Choose the profile type and GPU memory size that fit your workload. |
| Keeps the GPU software on the physical servers and the template driver in step, refreshing the template at major upgrades. | Install and update the GPU libraries, frameworks and applications inside the virtual server. |
| Defines GPU policies, memory reservation and VM preparation through the portal. | Make sure your code fits the limits of a sliced GPU (no nested virtualisation, no managed memory). |
| Builds the Windows GPU template and sets the Linux licence service to restart itself. | Decide which port your model service is published on and to whom, and write the firewall rule. |
| Runs the physical servers, network and storage, and restarts servers after a hardware failure. | Choose the backup plan for your models, training data and server. |
Who it is for
Software teams adding AI to their product
Teams that want to run a language model, speech recognition or image service on their own infrastructure. Compute profiles are ready for inference, and the model sits right next to the application servers on the same network.
Design and engineering teams
Teams working in CAD, engineering and 3D visualisation software. Graphics workstation profiles run that software on a central Windows server, so heavy model files stop travelling between desktops.
Organisations that must keep data in Türkiye
Organisations that can't send customer data or internal documents to an AI service abroad. The model and the data run on Nesnet's infrastructure in Türkiye, inside the organisation's own network.
Existing Nesnet customers
Teams already on Nesnet virtual servers or Private Cloud. A GPU is added to your existing allocation as a policy, with no new account, no new network and no separate contract.
May not be the right fit
If you need to run Hyper-V, WSL2 or another virtualisation layer inside the GPU server, this isn't the product: GPU slicing doesn't work together with nested virtualisation.
If your training code relies on GPU managed memory, for example paged optimisers, that feature doesn't work on a sliced GPU, and the code has to be adjusted to run without it.
If you need to combine many cards into one large training job, the profiles on this page aren't the right tool; talk to Nesnet's team about a job at that scale separately.
Under the hood
GPU Servers run in the same cloud portal and on the same infrastructure as Nesnet Virtual Servers; the difference is that the VM gets a slice of a GPU. The card isn't passed through to the VM directly (this isn't PCI passthrough): a GPU virtualisation layer divides the card's memory into profiles, and each profile appears in the portal as a GPU policy. One kind of data-centre GPU handles both graphics and compute, the other compute alone. When the VM is prepared, it is moved to current virtual hardware, the setting that lets it address large GPU memory is switched on, and it is placed in the group of GPU-equipped servers. Licensing comes from a licence server Nesnet operates: compute profiles use a compute licence, graphics profiles a graphics workstation licence. On top of that, Nesnet builds the Windows Server 2022 template with the driver and licence token baked in, matches the driver to the GPU software on the physical servers, and configures the Linux licence service to restart itself.
Building blocks
- GPU virtualisation layer
- divides the card into memory profiles and attaches them to VMs
- Graphics and compute GPU
- graphics profiles of 3, 6, 12, 24 and 48 GB; up to four virtual displays per profile
- Compute GPU
- compute profiles of 12 and 24 GB
- GPU licence server
- run by Nesnet; GPU servers pull their licence from it at boot
- Cloud portal and API
- GPU policies and VM preparation
- Virtualisation layer
- placement by GPU capacity and restart after failure
How it compares
| Option | When it is the right choice |
|---|---|
| Buying your own GPU | If you'll keep a GPU close to fully busy for years and have the server, cooling and people to maintain it, owning the card can make sense. If usage is spiky or still unknown, starting on a sliced profile lets you put the investment off. |
| GPU instances from global hyperscalers | For large-scale training that combines many cards in one job, or for regions in several countries, hyperscalers offer a wider choice. If keeping the model and data in Türkiye, local e-invoicing and running on the same network as your other servers matter more, Nesnet is the better fit. |
| A Nesnet Virtual Server without a GPU | If your workload doesn't really benefit from a GPU — say you only run a model now and then on small data — a server without one is cheaper. A GPU can be added to the same server later through a policy. |
| GPUs inside Nesnet Private Cloud | If you want several GPU servers, Kubernetes clusters and your other servers under one quota, the GPU policy is published to your Private Cloud allocation. |
Technical specifications
| Graphics profiles (GPU memory) | 3 GB (1/16), 6 GB, 12 GB, 24 GB, 48 GB (full card) |
|---|---|
| Compute profiles (graphics + compute card) | 6 GB, 12 GB, 24 GB, 48 GB (full card) |
| Compute profiles (compute-only card) | 12 GB (1/2), 24 GB (full card); no graphics profile |
| Slice options | Full card, 1/2, 1/4, 1/8, 1/16 |
| Virtual displays (graphics profile) | Several depending on profile; up to four on the largest |
| Attachment method | Software slicing; not direct card passthrough |
| Licence server | Operated by Nesnet |
|---|---|
| Licence handling | No vendor licensing portal for the customer |
| Licence type | Compute licence for compute profiles, graphics workstation licence for graphics profiles |
| Windows template | Windows Server 2022 + GPU driver + licence token |
| Linux | Ubuntu 24.04 + Linux GPU driver + licence, verified |
| Compute libraries | Run with widely used GPU compute, deep-learning and linear-algebra libraries |
| Memory reservation | 100%, applied automatically by the policy |
|---|---|
| VM preparation | Current virtual hardware, 64-bit MMIO, placement among GPU-equipped servers |
| How you choose | GPU policy in the portal |
| In one allocation | GPU and non-GPU servers side by side |
| Same network | Organisation network and gateway firewall |
| Nested virtualisation | Not supported (no Hyper-V / WSL2) |
| Managed memory | Does not work on a sliced GPU |
| Placement and failure | Placed by GPU capacity; restarted on another GPU server when compatible capacity is free |
| Data centres | Equinix Istanbul (Dudullu), Türk Telekom Ankara |
If something breaks
The driver doesn't match the host
When the driver and the GPU software on the physical server don't match, Windows Device Manager shows the GPU as stopped with an error. Because Nesnet versions the template against that software, a server built from the ready template has that match in place from the start.
The licence can't be obtained
If the GPU can't pull its licence from the licence server, it shows as unlicensed. A defined check of licence-server reachability tells you whether the problem is network access or the licence server itself.
The Linux licence service stops
If the licence service crashes, it restarts itself, so the GPU doesn't drop to unlicensed without anyone noticing.
A GPU host fails
The virtualisation layer's high availability can restart a GPU VM too, provided another GPU host has free capacity for the same profile. For a critical workload, agree spare-capacity needs with Nesnet's team up front.
Protecting the disk
A GPU server's disk can be backed up the same way as any other virtual server; the Backup product covers GPU servers too.
Frequently asked questions
Pricing and first steps
Pricing
GPU Servers are billed monthly. Each GPU profile has a fixed charge, the server's vCPU, memory and disk are added on top, and every item appears on its own line of the invoice.
What the price depends on
- The GPU profile chosen (card type and GPU memory)
- The server's vCPU count
- Memory (GB RAM) — fully reserved on a GPU server
- NVMe storage (GB)
- Operating system licence (Windows Server)
How to get started
- 1.Fill in the form or call 0800 606 1071.
- 2.Your workload is worked out in one call, with a profile recommendation and a price.
- 3.The GPU policy is added to your allocation and you launch your first GPU server from the portal.
Size your GPU to the workload
Buying a GPU card is a big investment before usage is known, and it brings driver and licence work with it. Tell Nesnet's team about your workload; you'll hear back the same business day, and the right profile is chosen together.