The obstacle in front of AI projects is often not the idea but access to hardware. With portvMind GPU / AI Compute you reach NVIDIA H100 and H200 cards in minutes and pay for the hours you use. You don’t wait for card supply, you don’t tie up budget in hardware that will age in three years — and most importantly, your training data never leaves the country.
GPUs are processors designed to run a large number of simple operations in parallel. The matrix calculations used to train and run AI models are exactly this kind of operation, so a training job on a GPU finishes many times faster than the same job on a CPU.
The problem is that GPUs are expensive and take a long time to procure. If training a model takes a few weeks, buying that hardware means idle capacity for the rest of the year. Cloud GPUs change the equation: you switch them on when you need them and off when the job is done.
portvMind GPU / AI Compute offers GPU-backed instances as self-service from the console. You can use them as standalone servers or define them as a separate node pool in your Kubernetes cluster. You keep your training data in Object Storage and read it directly from the GPU nodes; since it is in the same region, this traffic is not charged.
Hosting in Turkey carries extra weight for AI workloads. The data you use to train a model is often your organization’s most sensitive data: customer records, transaction history, document archives. Being able to process it without it leaving the country decides, in many sectors, whether a project is feasible at all.
Advantages
No Hardware Investment
Start without buying GPUs or waiting for procurement; resources are ready in minutes.
H100 and H200 Options
From large-model training to long-context inference workloads, choose the card family that fits your scenario.
High-Speed Interconnect
NVLink inside a node and InfiniBand between nodes keep the network from becoming a bottleneck in multi-card training jobs.
Kubernetes Integration
Define GPUs as a separate node pool and route your workloads to it with labels.
Close to Your Data
Your training dataset sits in Object Storage; access within the same region is free and low-latency.
Data Sovereignty
Your sensitive training data stays in the Istanbul and Ankara data centers and never leaves the country.
Configuration options, prebuilt image contents and regional coverage are published in the documentation.
For detailed technical specifications and release notes, see Documentation →
Use cases
Training models on your own data and fine-tuning existing models
Model endpoints running in production
Image processing and rendering workloads
Document classification, information extraction and natural language processing applications
Short-term prototyping to test feasibility before an investment decision
Calculate the Price of This Service
Enter the card family and how long you will use it, and see the total cost instantly.
Metering: Hourly — unlike other services, metered by the hour
As soon as a GPU resource is allocated, the whole card is reserved for you and cannot be used by any other workload during that time. That is why the metering unit is the hour. For continuously running inference workloads you can lower the unit cost with reserved commitment options.
Your resources are metered for as long as they are running — you never pay for time you don’t use. In the console you can define a monthly or period-specific budget and get alerts when your spending reaches the 50%, 80% and 100% thresholds. Turn on overrun forecasting to see in advance whether your current consumption rate will take you over budget by the end of the period.
Alerts arrive by e-mail; with webhook support they can also be sent to Slack, Microsoft Teams or your own application.
GPU resources are metered hourly and cost more than other services. We especially recommend defining a budget threshold for your training jobs and turning on overrun forecasting.
Budget alerts
%50%80%100
Budget alerts send notifications; they do not stop resources automatically. You decide which resource to scale down.
NVIDIA H100 and H200 cards are targeted. Current configuration options are listed in the documentation and the calculator.
After you complete the configuration in the console, your GPU resource is up and running within minutes. You don’t wait for card supply or run a purchasing process.
As soon as a GPU resource is allocated, the whole card is reserved for you and cannot be used by any other workload during that time. That is why the metering unit for GPU is the hour. All other services are metered per minute.
Yes. Because metering is hourly, when planning your trials it is more efficient to run your tests back to back than to start the resource and leave it idle for a long time. You can use your first $45 of free usage for this kind of trial.
You start the resource from the console and configure the environment to your own needs. Available image options are listed in the documentation.
On Object Storage or Block Storage, in the Turkish region you choose. The data never leaves the country.
