LLM Hosting with EU Data Residency: GDPR Compliance Guide

Hosting an LLM under GDPR requires running inference on servers inside the European Economic Area, with strict data residency controls, end-to-end encryption, and binding processor agreements. Remove personally identifiable information before any prompt or training data enters the model pipeline.

Defensible Architecture for GDPR LLM Hosting

Keep your model and inference logs inside EU-controlled infrastructure. This satisfies residency rules and eliminates cross-border transfer risks. Run managed inference on dedicated EU GPUs or deploy vLLM on EU cloud instances, keeping audit logs in storage you control.

The DCXV EU hosting guide recommends vLLM for production serving. Benchmarks show 25 to 40 tokens per second for single requests and 100 to 180 tokens per second when batching four concurrent requests. Target 300 to 600 milliseconds time to first token in production; for lower-throughput workloads, llama.cpp offers a CPU-based alternative.

Use private VPCs and private endpoints so inference traffic never traverses the public internet. IP addresses, user IDs, and session tokens in prompts count as personal data under GDPR, so routing must be fully isolated. GPU memory isolation is also required: model weights and intermediate activations must not leak to shared memory pools or swap files.

Implement crypto-shredding or secure deletion protocols that permanently erase prompts and logs after the defined retention window. Apply the same segmentation to development and staging environments. Never use production data for testing without explicit consent and rigorous anonymization.

Route every prompt and output through EU endpoints. This keeps personal data inside your operational control and simplifies compliance with Article 5 purpose limitation and Article 25 data protection by design.

GDPR Article 6 Lawful Bases for Prompts and Logs

Document a lawful basis under GDPR Article 6↗ for processing prompts, outputs, and logs. Legitimate interest or contractual necessity usually covers live inference, while retaining data for debugging requires a separate purpose analysis.

Relevant Article 6 bases

• Art. 6(1)(a) consent: Required where processing is not strictly necessary.
• Art. 6(1)(b) contract: Fits where inference delivers a requested service, limited to necessary parts.
• Art. 6(1)(f) legitimate interests: Plausible for enterprise inference or analytics, but requires a balancing test.

Store LLM requests in structured audit logs within your own infrastructure to meet Article 5 purpose limits. TrueFoundry’s LLM deployment playbook↗ notes that regulated workloads need different lawful bases for inference versus debugging retention. Training or fine-tuning on customer prompts generally requires a separately analyzed basis.

Map every data category to its specific lawful basis. Live inference relies on contractual necessity; anonymized telemetry falls under legitimate interests. Third-party monitoring tools processing prompts act as independent processors and require separate contractual coverage.

GDPR mandates a 30-day window to respond to data subject requests. Your logging architecture must support rapid retrieval and deletion of specific user data across inference caches and audit trails without disrupting live services.

Transparency obligations under Articles 13 and 14 require clear communication. Update your privacy policy to list LLM use cases, retention periods, and subprocessors, and provide straightforward opt-out mechanisms.

GDPR Article 28 Processor Obligations

GDPR Article 28 requires that inference, governance logic, and audit logs stay inside the controller’s infrastructure when a processor handles personal data on its behalf. The controller must use only processors that provide sufficient guarantees.

A binding Article 28(3) contract must cover:

• Processing only on documented instructions
• Confidentiality commitments
• Appropriate security measures
• Conditions for engaging sub-processors
• Assistance with data subject rights and DPIAs
• Deletion or return of data at end of services

Prediction Guard’s regulated industries overview↗ notes that managed LLM inference must keep prompts, outputs, and traces inside your organization. Model routing, logging, observability, and support access must be covered in the Data Processing Agreement. Sub-processors, including cloud or GPU infrastructure providers, must be disclosed and controlled.

GDPR Article 32 Security Controls

Article 32 mandates encryption in transit and at rest, pseudonymisation, and regular testing of security measures. Set up least-privilege role-based access and require multi-factor authentication for any admin access.

Encryption must cover:

• In transit: TLS for client-to-gateway, gateway-to-model, admin, and logging paths.
• At rest: Encrypted disks, object storage, backups, and log stores.

Customer-Managed Keys (CMK) or Bring Your Own Key (BYOK) models ensure that only your organization holds the decryption keys, neutralizing risks from cloud provider access or insider threats. Your Data Processing Agreement must explicitly specify model versioning, prompt retention periods, sub-processor lists for GPU providers, and audit rights.

Under Article 33, you must notify the relevant supervisory authority of a personal data breach within 72 hours of becoming aware of it, unless the breach is unlikely to result in a risk to individuals. Breach detection and monitoring are required under Article 32’s resilience obligations, covering unauthorized access monitoring, logging and alerting, incident response procedures, and backup recovery testing.

Incident response playbooks must account for AI-specific threats including prompt injection, model extraction, and training data poisoning. Conduct tabletop exercises quarterly to test response times and decision-making chains. Maintain an immutable audit trail of all security events, access logs, and configuration changes to support forensic analysis if a breach occurs.

DPIA Triggers and Automated Decision-Making

Run a Data Protection Impact Assessment under GDPR Article 35 if your LLM processes personal data at scale or drives automated decisions. EDPB Opinion 28/2024↗ makes clear that AI models fall under GDPR the moment personal data enters the pipeline.

DPIA triggers for LLMs

A DPIA is commonly triggered when an LLM:

• Processes personal data at scale
• Profiles individuals or makes automated decisions
• Combines large datasets with opaque inference
• Evaluates behavior, performance, preferences, location, or health
• Processes sensitive data in ways that create heightened risk

Keeping prompts for debugging or model improvement usually needs a separate lawful basis and raises your DPIA risk. Memorization and prompt retention are key triggers under Article 35(3)(a). If your system profiles people or handles credit, hiring, or healthcare decisions, strict DPIA documentation is required.

Automated decisions based on LLM outputs trigger Article 22 rights. You must guarantee a right to human intervention, the ability to contest decisions, and clear explanations of the logic involved. Document these safeguards in your DPIA.

A common oversight is failing to account for model memorization, where the LLM might regurgitate training data containing personal information. Implement output filtering and post-processing checks to detect and redact PII before responses reach the end user.

Cross-Border Transfers: EU, LATAM, and US Compliance

Cross-border transfers require distinct strategies for the EU, LATAM, and the US. US providers need an active EU-U.S. Data Privacy Framework certification or Standard Contractual Clauses paired with a Transfer Impact Assessment.

US Compliance

The EU-U.S. Data Privacy Framework remains in force as of 2026. US organizations must self-certify annually to the Department of Commerce and appear on the live DPF list. The General Court upheld the adequacy decision in Latombe v. Commission, though litigation risk persists. For SCC-based transfers, exporters must conduct a Transfer Impact Assessment and implement supplementary measures such as strong encryption with exporter-controlled keys.

LATAM Transfer Rules

Latin American jurisdictions enforce distinct requirements for LLM inference handling personal data. Brazil (LGPD) requires transfers to follow Article 33 mechanisms, typically relying on adequacy or standard contractual clauses. Mexico (LFPDPPP) updated its framework in 2025 following the dissolution of INAI, requiring notice, consent, and recipient commitments for cross-border moves. Colombia (Ley 1581/2012) centers on habeas data rights under the SIC, restricting international transfers without adequate protection.

Chile is transitioning to a modernized regime with stricter transfer conditions. Argentina (Ley 25.326) holds long-standing EU adequacy status, though exporters must still comply with local transfer conditions.

Practical Strategy

Dedicated EU and LATAM infrastructure keeps data within regional boundaries, reducing transfer friction. DCXV recommends self-hosting LLMs on EU cloud infrastructure to bypass SCCs and TIAs entirely. US providers should use the DPF only if certified and maintain SCCs as a fallback.

Managed Inference with Flat Monthly Pricing

GDPR-compliant managed inference runs on dedicated GPU clusters in the EU and LATAM, removing cross-border transfer friction and keeping personal data under your control.

Tessera AI runs these regions and serves Qwen3.6-35B-A3B through an OpenAI-compatible chat completions API. Flat monthly pricing scales with compute usage, avoiding unpredictable token-based billing. Estimate your costs using the pricing calculator.

Dedicated tenancy ensures inference workloads run on isolated hardware, preventing cross-tenant data leakage. Every component from the API gateway to the underlying GPU drivers operates within a single, verifiable compliance boundary. Configure rate limits and tiered access to enforce least privilege across different user roles and environments.

Regulated industries still need documented lawful bases under GDPR Article 6 and structured audit logs in customer-owned storage. Tessera AI provides a Data Processing Agreement that covers Article 28 requirements and supports your compliance workflow. Rate limits and tiers and limits are configurable to match your throughput needs.

FAQ

Does GDPR apply to LLM inference?

Yes. The EDPB makes clear that AI models do not sit outside GDPR. If your prompts, outputs, or logs contain personal data, Article 6 lawful bases and Article 28 contracts apply.

What are the GDPR Article 32 security requirements for LLMs?

Encryption in transit and at rest, pseudonymisation, and regular security testing. Structured audit logs must stay in customer-owned storage. Access controls must include least-privilege role-based access, MFA for admin access, and segregation of duties.

Can I use US-hosted LLMs for EU data?

Only if the US provider holds EU-U.S. Data Privacy Framework certification or uses SCCs with a Transfer Impact Assessment. Dedicated EU infrastructure removes that transfer risk entirely.

When is a DPIA required for LLM processing?

Article 35 requires a DPIA for large-scale processing, profiling, or automated decisions. EDPB Opinion 28/2024 flags memorization and prompt retention as key triggers.

How does LATAM compliance affect LLM hosting?

Brazil uses LGPD Article 33 mechanisms. Argentina has EU adequacy status. Mexico and Chile have evolving frameworks requiring notice and consent. Dedicated EU hosting simplifies transfers by keeping data within the EEA.