Two obstacles stand out in AI projects: access to GPUs and where the data goes. The first is a matter of budget and supply; the second often determines whether the project is feasible at all.
portvMind solves both. You reach NVIDIA H100 and H200 cards in minutes and pay for the hours you use. You don’t send your training data out of the country — both the dataset and the model artifacts stay in the Istanbul and Ankara data centers.
GPU hardware is expensive, takes a long time to procure and investing before a project delivers clear results is risky. A purchased card becomes idle capacity outside training periods.
On portvMind: With hourly GPU usage you validate the idea first. For continuously running inference workloads you lower the unit cost with the reserved option.
The data used to train models is often an organization’s most critical data: customer records, transaction history, document archives. Many organizations do not accept this data leaving the country.
On portvMind: Your training data stays in Object Storage in Turkey, and your GPU nodes read the data from the same region. The data does not leave the country and in-region transfer is not charged.
Training is intensive and periodic; inference is continuous and lighter. Running both with the same resource model is either expensive or insufficient.
On portvMind: By using hourly GPU for training and reserved resources for inference, you optimize the two workloads separately. On Kubernetes you run them in separate node pools.
The path from raw data to a model includes many processing steps, and each step needs a different resource profile.
On portvMind: You combine an Object Storage data lake, Compute preprocessing, Kubernetes orchestration and the GPU training layer on one platform, in one console.
Making a trained model servable is a separate engineering job; most projects get stuck at this stage.
On portvMind: You run your model endpoints on Kubernetes, publish them with Load Balancer and scale the replica count with load.
The checkpoint files of every experiment pile up; if they are not cleaned up, storage cost quietly grows.
On portvMind: By using versioning and lifecycle rules in Object Storage, you build a structure that manages the archive by itself.
| Service | Role |
|---|---|
| GPU / AI Compute | GPU resources for model training and inference |
| Object Storage | Training dataset, checkpoints and model artifacts |
| Kubernetes | Job orchestration and model serving layer |
| Compute | Data preprocessing and transformation workloads |
| Block Storage | High-speed working space during training |
| Load Balancer | Publishing model endpoints |
| Network Data Transfer | Free in-region data flow |
This combination is a starting point; we adapt it together to your data pipeline.
Your raw data sits in Object Storage as a data lake. You run preprocessing and transformation jobs on Compute and write the processed dataset back to Object Storage.
Training jobs run on GPU / AI Compute. If you use Kubernetes, you define the GPUs as a separate node pool and route training jobs to that pool with labels. When training starts only the GPU nodes come into play, and when the job ends the pool shrinks — you keep the expensive resource running only for as long as it is needed.
During training you read the dataset many times. Since the data and the GPUs are in the same region, these reads create no data transfer cost.
When the model is ready for production you publish the inference service in the same cluster behind a Load Balancer. Since it will run continuously, using reserved resources in this layer lowers your unit cost.
Most teams start with a validation experiment: they upload the dataset to Object Storage, open a single GPU resource and see whether the model gives meaningful results on their own data. This stage usually takes a few hours and can be covered by the first $45 of free usage.
If the result is promising you scale training: you move to a multi-card configuration and automate your data pipeline on Kubernetes. When the model goes to production you fix the continuous cost by moving the inference layer to reserved resources.
If you would rather work with an end-to-end team than do the setup yourself, you can consider vMind’s managed AI services.
NVIDIA H100 and H200. You can see the current configuration options in the calculator and the documentation.
On Object Storage or Block Storage, in the Turkish region you choose. The data does not leave the country and this can be documented with audit records.
No. Your GPU nodes access Object Storage in the same region directly and in-region data transfer is not charged.
GPU resources are metered hourly; other services are metered per minute. The unit is the hour because the card is reserved solely for you for as long as it is allocated.
Yes. You start the resource, run your experiment and shut it down. You can use your first $45 of free usage for this kind of experiment.
Open your account in minutes and try the first $45 of usage for free. If you want to plan your data pipeline, write to our team.