EU AI Act and Hosting: What You Need Before August 2026

The EU AI Act started biting in 2025. The general-purpose AI obligations kicked in on August 2, 2025. The high-risk AI system rules land in full force on August 2, 2026. That deadline is now eleven weeks away, and a lot of teams who built AI features into their products are realizing their hosting setup is part of the compliance equation, not separate from it.

This guide cuts through the legal text. It covers what the AI Act actually requires from a hosting and infrastructure perspective, which provider categories make compliance easier versus harder, and the specific data flows that get you fined if you get them wrong.

Disclaimer up front: this is not legal advice. Talk to an actual EU AI lawyer for your specific situation. What follows is the practical infrastructure layer of compliance that engineers and founders need to understand before they call that lawyer.

What the AI Act actually classifies

The AI Act sorts AI systems into four buckets:

• Prohibited: social scoring, real-time biometric ID in public spaces, manipulation of vulnerable groups. Banned outright since February 2025.
• High-risk: AI in hiring, education, credit scoring, law enforcement, medical devices, critical infrastructure, biometric ID. Heavy obligations.
• Limited risk: chatbots, deepfakes, emotion recognition. Transparency obligations (you must tell users they are interacting with AI).
• Minimal risk: AI-powered spam filters, video game NPCs. Mostly unregulated.

Most SaaS companies building AI features land in "limited risk" with a disclosure obligation. If you sell AI to HR teams, schools, banks, or hospitals, you are in "high-risk" and the hosting decisions matter a lot more.

The hosting-relevant obligations

Five parts of the AI Act care about where and how your AI runs:

1. Data governance (Article 10)

High-risk AI providers must document training, validation, and test data. They must show data quality, representativeness, and where the data came from. This means your training pipeline needs auditable storage. Object storage with versioning and access logs becomes a compliance requirement, not a nice-to-have.

2. Record-keeping and logging (Article 12)

High-risk systems must log automatically. You need to capture inputs, outputs, the model version that produced them, and timing. Logs must be retained "appropriate to the intended purpose," which most lawyers read as six months minimum, often years.

This is a hosting question. Where do those logs live? Are they in the EU? Are they encrypted? Can a regulator pull them in a compliant format?

3. Human oversight (Article 14)

Architectural requirement: your system must let a human review, intervene, or override AI decisions. Hosting plays a role because the override path needs to be available, fast, and itself logged.

4. Accuracy, robustness, cybersecurity (Article 15)

High-risk AI must be "resilient against errors, faults, inconsistencies." It must include cybersecurity measures appropriate to the risk. In practice this means hardening your hosting stack: managed firewalls, DDoS protection, regular patching, intrusion detection. Your provider's security posture becomes your security posture.

5. GPAI model obligations (Articles 53-55)

If you train or significantly fine-tune a general-purpose AI model that meets the threshold (10^25 FLOPs of training compute), you have systemic-risk obligations including incident reporting and red-teaming. Most companies will not hit this threshold, but if you fine-tune Llama or DeepSeek on serious infrastructure, double-check the math.

Data residency: not strictly required, but practically essential

Common misconception: the AI Act does not require EU data residency. It is technology-neutral about where compute happens. What it does require is GDPR compliance for any personal data processed, plus the ability to produce documentation and logs to EU regulators on demand.

Combine that with the post-Schrems II reality (US providers cannot guarantee CLOUD Act exemption) and EU data residency goes from "not required" to "the only path that actually works in practice."

If your AI processes data on US-headquartered cloud (AWS, Azure, GCP, Vercel, DigitalOcean), you have three options:

1. Use their EU regions and rely on the EU-US Data Privacy Framework. Legal but fragile, given that previous frameworks were struck down.
2. Add Standard Contractual Clauses plus supplementary measures. Expensive and complex.
3. Move to a EU-owned provider.

For high-risk AI systems serving EU users, option 3 is increasingly the default. Auditors find it easier to approve. Customers find it easier to accept. Regulators find it easier to inspect.

EU-owned hosting that satisfies AI Act practical requirements

The providers below have EU ownership, EU data centers, EU-only data flows for their core services, and the security certifications regulators expect.

Provider	HQ	EU AI Act fit	Cheapest VPS	Notes
Hetzner	Germany	Excellent	3.49 EUR (CX23)	ISO 27001, SOC 2, German data law applies. Sovereign by default.
OVHcloud	France	Excellent	~3.50 EUR	SecNumCloud certified for tier-restricted EU sovereign cloud. Used by French government.
Netcup	Germany	Very good	4.87 EUR (VPS 500 G12)	German owned, ISO 27001, value pricing on ARM and AMD.
IONOS	Germany	Very good	Varies	Listed on Frankfurt exchange, EU sovereign cloud product, ISO 27001.
Contabo	Germany	Good	3.60 EUR (Cloud VPS 10)	Cheap and EU-owned but fewer certifications. OK for non-high-risk workloads.

Hetzner and OVHcloud are the two that show up in serious EU compliance conversations most often. OVHcloud has the most aggressive sovereignty positioning thanks to their SecNumCloud product. Hetzner has the best raw price-to-performance for non-classified workloads.

See our deeper comparison in European GDPR Hosting Alternatives for the post-Schrems II legal landscape these providers operate in.

What about AWS, Azure, GCP EU regions?

Honest answer: legally usable if your DPA is tight and you have done a transfer impact assessment. Practically uncomfortable for high-risk AI systems.

The CLOUD Act gives US authorities the ability to compel US-headquartered providers to produce data regardless of where it physically sits. That means an AWS EU region is still subject to US legal process if a US court orders it. EU data protection authorities have been increasingly skeptical of arrangements that try to wave this away.

For limited-risk AI (a chatbot with a disclosure label), AWS Frankfurt is fine. For high-risk AI (medical decision support, hiring evaluation, credit scoring), you want a path that does not have a US legal hook in it.

Some teams compromise with a hybrid model: training and model artifacts on AWS in the EU, inference and customer data on a sovereign EU provider. Defensible, but doubles your ops surface area.

Logging architecture for AI Act compliance

The recordkeeping obligations in Article 12 translate to a specific infrastructure pattern.

Minimum log payload per AI decision:

• Timestamp (UTC).
• Model name and version.
• Input data fingerprint (hash, not the data itself if it is sensitive).
• Output produced.
• Confidence or risk score if applicable.
• Whether a human reviewed or overrode.
• Outcome of any override.

That log needs to live in EU object storage with versioning and immutable retention. Hetzner Storage Box, OVHcloud Object Storage, and IONOS Object Storage all work. Hetzner Storage Box at 3.49 EUR per terabyte is the cost floor most teams converge on.

Pair this with an append-only log shipper (Vector, Fluent Bit) that writes structured JSON. Add hash chaining if you want tamper-evident logs, which auditors love.

Retention is the trick. The Act says "appropriate to purpose." Most legal teams settle on 5 to 10 years for high-risk decisions. At 1 KB per decision and a million decisions per year, that is 5 to 10 GB. Trivial cost. The work is in the audit chain, not the storage bill.

Inference hosting: where to actually run the model

The other half of the question. Three patterns work for AI Act compliance:

Pattern 1: hosted LLM API on an EU provider

Mistral AI (France), Aleph Alpha (Germany), and a handful of others offer API access to their models from EU infrastructure. Easy compliance story. You sign a DPA, your data stays in the EU, the provider handles the model.

Pricing is generally higher than OpenAI or Anthropic per token, but improving. For high-risk workloads the premium is easy to justify.

Pattern 2: self-hosted open model on EU compute

Run Llama, Qwen, Mistral, or DeepSeek on your own EU VPS or dedicated server. You own the entire data flow. Highest compliance certainty. Highest ops burden.

For inference on a 7B-class model, Hetzner's GPU offerings (RTX 4000 SFF, RTX 6000 Ada) at around 200 to 500 EUR per month land in the right range. For CPU-only inference on smaller models, see our self-hosted LLM guide.

Pattern 3: hybrid with PII stripping at the edge

Strip personally identifiable data at an EU edge layer, then send anonymized prompts to OpenAI, Anthropic, or whatever frontier model you want to use. The personal data never leaves the EU. The model never sees identifying details.

This is legally cleaner than people think, but operationally fragile. PII detection misses things. A regulator finding even one slip in your logs is a problem. Use this pattern only with mature PII tooling and conservative classification.

Documentation: the part that catches everyone off guard

Article 11 requires technical documentation for high-risk AI systems. The list is long. Specifically the part that surprises engineers:

• Design choices for the system architecture.
• Description of hardware used.
• Computational resources required.
• Capabilities and limitations.
• Monitoring, functioning, and control mechanisms.

That "description of hardware used" is hosting documentation. Which provider, which region, which instance type, which network architecture, which security controls. If you cannot produce this in a regulator-friendly format on demand, you have a gap.

The good news: most EU providers publish this in white papers and audit reports. Hetzner has SOC 2 Type II reports. OVHcloud has their full SecNumCloud documentation. IONOS publishes detailed security architecture papers. Pull these, bundle them into your compliance binder, done.

What to do this quarter if your AI is high-risk and your hosting is US-based

The August 2, 2026 deadline is real. The fines are real (up to 35 million EUR or 7% of global revenue for prohibited AI violations, 15 million EUR or 3% for high-risk violations). The enforcement bodies are spinning up.

If you are on a US-headquartered provider and your AI system processes EU personal data in a high-risk context, here is the minimum path forward:

1. Map your data flows. Where does PII enter the system, where does it go, where does it land.
2. Identify the AI Act classification of your system. Get a legal opinion in writing.
3. If high-risk, plan a migration of personal data processing and logging to an EU sovereign provider. OVHcloud SecNumCloud or Hetzner are the defaults.
4. Implement Article 12 logging immediately. Even if other pieces lag, logging needs to start now to build a defensible record.
5. Document everything per Article 11. Start the binder. Update it monthly.

For limited-risk AI (chatbots, recommendation features), the bar is much lower: a clear AI disclosure to users, GDPR compliance on the data you process, and reasonable documentation. Most teams in this bucket pass without major hosting changes.

The pragmatic take

The EU AI Act is not GDPR. It is more prescriptive, more technical, and the deadlines are tighter. For most AI builders the hosting question reduces to one decision: do you keep AI-related data flows on US-headquartered cloud, or do you move them to EU-owned infrastructure?

For limited-risk systems, you can probably stay where you are with some configuration tightening.

For high-risk systems, you should already be in motion. Hetzner, OVHcloud, IONOS, and Netcup all sit in the sweet spot of price, performance, and compliance-friendly defaults. The migration is unpleasant but tractable. The fines for not migrating are not.

The AI Act is going to do for European AI hosting what GDPR did for European data hosting: shift the center of gravity toward EU-owned providers, faster than anyone expected. The teams that move first land on better infrastructure at better prices. The teams that wait until the fines start landing pay twice.