A single API call to a US-hosted language model containing customer data can trigger a GDPR breach notification obligation. Article 44 restricts transfers of personal data outside the EEA unless adequate safeguards exist, and the Schrems II ruling invalidated Privacy Shield. For UK organisations processing personal data through AI models, the safest architecture is one where data never leaves UK-hosted GPU infrastructure. This guide covers the practical steps to achieve GDPR-compliant AI inference from lawful basis through technical implementation.
Establishing Lawful Basis for AI Processing
Before deploying any AI system that processes personal data, you must identify a lawful basis under Article 6. Legitimate interest (Article 6(1)(f)) is the most common basis for AI inference, but requires a documented Legitimate Interest Assessment. Consent (Article 6(1)(a)) applies when users explicitly opt in to AI-powered features. Contract performance (Article 6(1)(b)) covers AI processing necessary to deliver a contracted service.
Special category data — health records, biometric data, political opinions — requires an additional condition under Article 9. Healthcare AI projects should review our GDPR-compliant hosting guide for sector-specific considerations. Document your chosen basis before writing a single line of inference code.
Data Minimisation in AI Pipelines
Article 5(1)(c) requires that personal data be adequate, relevant, and limited to what is necessary. For AI inference, this means stripping unnecessary personal identifiers before they reach the model. Practical techniques include:
| Technique | Implementation | Impact on Accuracy |
|---|---|---|
| Named entity redaction | spaCy NER before inference | Minimal for classification tasks |
| Pseudonymisation | Replace names with tokens, reverse after inference | None — full fidelity preserved |
| Aggregation | Process summaries rather than individual records | Moderate — loses individual context |
| Prompt truncation | Send only relevant text segments | Minimal with good chunking |
On a private AI server, you control the entire preprocessing pipeline. No third-party processor sees the original data. Pseudonymisation is particularly effective: replace personal identifiers with random tokens before inference, then map results back to real identities in your application layer.
Data Protection Impact Assessment
Article 35 requires a DPIA when processing is likely to result in a high risk to individuals. AI inference on personal data almost always qualifies. Your DPIA should document the nature and scope of processing, necessity and proportionality, risks to individuals, and mitigation measures. The ICO provides a DPIA template, but you must customise it for your AI use case.
Key risk areas for AI inference include model hallucination producing incorrect personal data, unintended memorisation of training data, inference results being used for automated decision-making under Article 22, and security of GPU infrastructure holding personal data in VRAM during processing.
Technical Controls for Compliance
Self-hosting on dedicated GPU servers provides the technical foundation for compliance. Essential controls include:
- Encryption in transit: TLS 1.3 between application and inference endpoint — mandatory even on localhost if containers are involved
- Encryption at rest: LUKS full-disk encryption for model weights and any cached inference data
- Access logging: Every inference request logged with timestamp, user context, and data categories processed
- Retention limits: Automated deletion of inference logs after the documented retention period
- Network isolation: Inference server on a private VLAN with no direct internet access
Deploy models through vLLM with the OpenAI-compatible API, then enforce access controls at the reverse proxy layer. See our infrastructure guides for deployment patterns.
UK Adequacy and Data Sovereignty
Post-Brexit, the UK operates under the UK GDPR and Data Protection Act 2018. The EU granted the UK an adequacy decision in June 2021, but this is subject to review. Hosting AI inference within the UK ensures compliance regardless of the adequacy decision’s future. Data processed on a UK GPU server never triggers international transfer provisions.
For organisations serving both UK and EU customers, UK hosting satisfies both jurisdictions while the adequacy decision holds. Cross-sector use cases demonstrate how regulated industries implement this architecture. The open-source LLM hosting approach means no vendor lock-in and full transparency over model behaviour.
Implementation Checklist
Start with the legal groundwork: document your lawful basis, complete a DPIA, and establish a data processing record under Article 30. Then build the technical layer: provision a UK GPU server, deploy your model with vLLM, implement preprocessing for data minimisation, and configure audit logging. Test the full pipeline with synthetic data before processing real personal data. Appoint a DPO if required under Article 37, and schedule annual reviews of your DPIA as model capabilities evolve.
Run GDPR-Compliant AI on UK GPU Servers
Personal data stays on UK soil. Dedicated GPUs, full-disk encryption, no third-party data processors in the inference path.
Browse GPU Servers
