What does using GPU servers abroad to train or run AI models mean under Turkey's KVKK? We explain Article 9 as amended by Law No. 7499, the transfer mechanisms and practical techniques such as anonymisation, synthetic data and encryption that reduce your risk.
AI projects need GPU power, and much of that power sits in data centres in other countries. Using a GPU server abroad to train a model, fine-tune it or run inference takes only minutes technically. The legal side needs a little more care.
If the data you process contains personal data, sending it to a server abroad may count as a cross-border transfer under Turkey's Personal Data Protection Law No. 6698 (KVKK). In this article we outline the general framework and the practical techniques you can use to reduce risk.
Important note: this article is for general information only and is not legal advice. We recommend working with legal counsel for an assessment of your specific situation.
Where Personal Data Shows Up in AI Projects
Many teams believe they are training their model on "just text" or "just images". In reality, this data often contains personal data:
- Customer support logs: names, phone numbers, addresses, order details.
- Audio recordings: call centre conversations, voice assistant data.
- Images: faces, licence plates, scanned identity documents.
- Documents: contracts, invoices, CVs, emails.
- Health and financial data: sensitive or special category data such as patient reports and loan applications.
Once this data enters a training set, the location of the GPU server matters. During inference, the prompts users send to the model may also contain personal data.
That is why it pays to build a simple data inventory before the project starts. Gather in one table the data sources you will use, the personal data fields each contains, which system the data goes to at each stage and how long it will be kept. This table speeds up your conversation with legal counsel and makes clear to the technical team which data needs to be anonymised.
Pay special attention to special category personal data while building the inventory. Categories such as health, biometric data or criminal convictions are subject to stricter rules under KVKK. For projects involving this kind of data, prioritising an architecture in which the data never leaves Turkey is often the soundest path.
KVKK Article 9 and the Change Introduced by Law No. 7499
Article 9 of the KVKK, which governs cross-border data transfers, was significantly amended by Law No. 7499 in March 2024. Under the old regime, explicit consent was the route most often relied on in practice. The new regime introduces a tiered structure of transfer mechanisms.
Broadly, the new system works like this. First, one of the processing conditions in the law must apply. Next, an adequacy decision is sought for the destination country or sector. If there is no adequacy decision, one of the appropriate safeguards must be in place. If none of these is possible, occasional transfers may be relied on in limited cases.
Cross-Border Transfer Mechanisms
| Mechanism | What it means in short | What to watch |
|---|---|---|
| Adequacy decision | The Board decides that a country, a sector within a country or an international organisation offers adequate protection | Check whether a current decision exists for the destination |
| Standard contract | The parties sign the standard contract text announced by the Board | It must be notified to the Board within 5 business days of signing |
| Binding corporate rules | Board-approved rules for transfers within a group of companies | Requires an approval process, usually suited to large groups |
| Written undertaking | An undertaking with provisions ensuring adequate protection, used with Board permission | Requires Board permission |
| Occasional transfer | A non-recurring transfer based on exceptional cases listed in the law | Not suitable for ongoing workloads |
For recurring work such as AI training or regular inference, occasional transfers are generally not a suitable basis. In these scenarios the route most often considered is the standard contract. Which mechanism fits you should be confirmed with legal counsel based on your data types and workflow.
Practical Techniques That Reduce Risk
Legal mechanisms matter, but the most effective approach is often to technically reduce the amount of personal data that leaves the country. Data that is not transferred carries no transfer risk.
Anonymisation vs Pseudonymisation
These two concepts are often confused, but their consequences under KVKK are very different:
- Anonymisation: making data impossible to link to a specific person in any way, even when combined with other data. Truly anonymised data is no longer personal data.
- Pseudonymisation: replacing direct identifiers such as names with a code or random ID. As long as a mapping table exists somewhere, the data can be linked back to the person. Pseudonymised data is therefore still personal data.
Pseudonymisation is a valuable security measure, but on its own it does not remove transfer obligations. Anonymisation needs care too: free text, rare attributes or a combination of several fields can lead to re-identification.
Synthetic Data and Training on Non-Personal Data
In some projects it is possible to move forward without needing real personal data at all:
- Synthetic data: generating artificial data that mimics the statistical structure of real data but does not belong to real people. This is especially useful for prototypes and early training stages.
- Non-personal data: training on sources that contain no personal data, such as product catalogues, technical documentation, machine sensor data or publicly available text with a suitable licence.
When generating synthetic data, make sure the generator model does not memorise and reproduce real records verbatim.
Keeping Raw Data in Turkey
An architecture that works well in practice: raw personal data stays on your systems in Turkey. Anonymisation and cleaning happen there. Only the anonymised training set is sent to the GPU server abroad.
You can use CloudRUX S3 object storage to keep raw data and anonymised copies in Turkey. Your data preparation pipeline stays in the country, while compute-heavy training runs on GPU servers.
| Stage | Where? | Data type |
|---|---|---|
| Raw data collection and storage | Turkey | May contain personal data |
| Anonymisation and cleaning | Turkey | Personal data to anonymous data |
| Model training | GPU server abroad | Anonymous or non-personal data |
| Retrieving model weights | Turkey | Trained model |
Encryption, Deletion and Logging
Whatever data you transfer, basic security measures are essential:
- Encryption in transit: send data to the server over encrypted channels such as SSH, SFTP or HTTPS.
- Encryption at rest: keep datasets on the server on encrypted disks or in encrypted archives.
- Delete after the job: when training is complete, delete datasets, intermediate files and checkpoints from the server. Retrieve only the model outputs you need.
- Logging: record which data was sent to which server, when and by whom, and when it was deleted. These records demonstrate accountability in case of an audit.
- Access restriction: keep the number of people with server access to a minimum and manage SSH keys per person.
Remember that a GPU can only process data that has been decrypted in memory. Encryption therefore improves security, but on its own it does not change the fact that personal data has been transferred abroad.
Where Do CloudRUX GPU Servers Run?
We want to be transparent: CloudRUX GPU servers run in our partner's data centers abroad. For projects that will process personal data, we therefore recommend carrying out the assessment described above.
The service is in early access, with hourly starting prices excluding VAT such as 0.74 USD for the L4 24 GB and 4.99 USD for the H100 80 GB. Billing is done in Turkey with an e-invoice, and support is provided by our Turkish-speaking engineers. We also help you with the standard contract process.
Frequently Asked Questions
Is processing anonymised data on GPUs abroad within the scope of KVKK?
If the data has truly been anonymised, meaning it cannot be linked to a person in any way, it is not personal data and falls outside the cross-border transfer rules. However, whether the anonymisation is sufficient needs careful assessment.
Is signing a standard contract enough?
A standard contract is a transfer mechanism, but it does not solve everything on its own. A processing condition must also apply, the duty to inform must be fulfilled and the contract must be notified to the Board within 5 business days of signing. Consult your legal counsel for details.
Can model weights contain personal data?
That depends on the type of model and its training data. Some models can memorise fragments of their training data. Not training on personal data, or anonymising it beforehand, is therefore the safest approach.
Can explicit consent still be used?
Since Law No. 7499, explicit consent is no longer the first route for transfers and plays a limited role among the occasional transfer cases. Regular workloads call for a different mechanism.
CloudRUX GPU Cloud
CloudRUX offers GPUs from the L4 to the B200 for your AI projects at hourly prices, with SSH and Jupyter access and ready PyTorch and CUDA templates. We state clearly from the outset that the servers run in our partner's data centers abroad, and we help you design a KVKK-compliant data flow and work through the standard contract process. See cards and prices on our GPU Cloud page, or get in touch to review your project with us.