Unit 6.6: AI Sovereignty and the EU AI Act
Introduction
Unit 1.4 set three filters for every service: location, credential scope and trust model. The AI tier adds a fourth question that regulators now ask: under the EU AI Act, who is responsible for what? For FinCorp the answer decides which controls its assistant needs before it goes live.
This unit applies the filters to AI Model Hub and maps the roles. It is a design unit: the build of Unit 6.5 does not change, but the obligations around it do.
1. The Data-Sovereignty Boundary
1.1 What Happens to a Request
All inference is performed in IONOS CLOUD data centers in Germany, and prompts and generated outputs are processed there and nowhere else. The content of prompts and outputs is never logged, never accessed by IONOS CLOUD personnel and never written to persistent storage. It is not used to train, fine-tune or otherwise improve any model, and it is not disclosed to any third party, including the developers of the models.
The inference API is stateless per request (Unit 6.5). During processing, the prompt is held in plaintext in host memory and GPU memory, and it is not retained after the request completes. What is recorded is metadata: the timestamp, the model invoked and the number of tokens processed, for billing and operation. Access to operational data is restricted to IONOS CLOUD personnel located in the European Union. No third-party sub-processors are engaged, and model developers supply only the weights, so they have no access to your data.
Transport uses TLS for all external API traffic, persistent storage uses AES-class encryption, and API access is authenticated with tokens that you create and revoke.
1.2 What Stays With You
The boundary has an inside that you own. IONOS CLOUD does not pseudonymize request content, because the model needs the actual text. If FinCorp needs pseudonymization, it applies it in its application before data leaves its environment, for example by replacing names with tokens and restoring them in the response. FinCorp remains the controller of what its requests contain. For special-category data under Article 9 GDPR, no restrictions apply beyond the measures above, and FinCorp is responsible for its own legal basis and, where required, its data protection impact assessment. The data processing agreement applies as part of the contract and needs no separate signature.
1.3 Credentials and the Training Boundary
Two credentials enter the compliance mapping. IONOS CLOUD data centers are certified to ISO 27001, and AI Model Hub falls within the scope of BSI C5, the C5:2020 Type 1 attestation of 28 May 2026 (Unit 1.4). Cite each with its scope, not as a platform claim.
The boundary above describes inference only. AI Model Hub serves pre-trained models and keeps customer data out of training. If FinCorp ever fine-tunes a model, the work runs on Cloud GPU VMs that FinCorp operates (Unit 6.5). The data then sits in FinCorp's own VMs and volumes, and the statements of Section 1.1 no longer describe it: retention, access and evidence become FinCorp's design.
2. Roles Under the EU AI Act
2.1 Three Parties, Three Roles at Once
In AI Model Hub, three parties each hold a role, and all three apply at the same time:
| Party | Role under the AI Act | What it carries |
|---|---|---|
| Model developers | Providers of the general-purpose AI models they publish | The model obligations of Article 53: technical documentation, information for downstream providers, a copyright policy and a public summary of training content |
| IONOS CLOUD | Provider of the AI system (Article 3(3)), for every model in the catalogue, whether modified or not | The system obligations, including machine-readable marking of synthetic output under Article 50(2) |
| The customer | Deployer, and possibly the provider of its own AI system | The risk classification of its application and the deployer obligations of Article 50 |
As shown above, the roles stack rather than replace each other. IONOS CLOUD links each model's documentation page to the developer's published model card and license, but it does not verify or warrant the completeness of the developer's Article 53 documentation; FinCorp assesses that from the developer's sources.
2.2 When IONOS CLOUD Modifies a Model
Modification does not change who the system provider is: IONOS CLOUD is the provider of the system for every model. What modification adds is documentation. Where IONOS CLOUD modifies a model, for example by quantization, it identifies the original base model and the nature of the modification, and it publishes its own technical documentation for the modified model, including the internal benchmarks that check accuracy and robustness against the original. For FinCorp's model selection, that documentation belongs in the evidence file next to the developer's model card.
2.3 Article 50 and the Deployer's Work
Article 50(2) puts the marking of synthetic output on the provider of the AI system. IONOS CLOUD is implementing these marking and detection measures per media type, within the transitional period for AI systems placed on the market before 2 August 2026. The other Article 50 obligations fall on the deployer of the application:
- Disclose to natural persons that they are interacting with an AI system (Article 50(1)).
- Inform persons exposed to emotion recognition or biometric categorization (Article 50(3)).
- Label deep fakes and AI-generated text published to inform the public on matters of public interest (Article 50(4)).
Beyond transparency, the risk class is the customer's assessment. A limited-risk system, such as a general-purpose chatbot, mainly needs the transparency notice. A high-risk use, such as employment, credit or education, needs data governance, event logs and human oversight. IONOS CLOUD does not provide a pre-configured AI Act logging template, so the audit trail of requests and responses is built in your application.
3. FinCorp's Position
FinCorp's advisor assistant passes the filters of Unit 1.4: processing in Germany, a service within C5 scope, and no training on its data. FinCorp records itself as deployer. Its own assessment classes the assistant as limited-risk, so the advisor interface states that answers come from an AI system. Customer names and account numbers are pseudonymized before each request.
A planned credit pre-assessment would be a high-risk use. Before it goes live, FinCorp adds three controls in fc-k8s-ai: a request and response log written to its own storage with a defined retention, a human approval step before any output reaches a credit decision, and documentation of the data used for retrieval. These are FinCorp's controls, and the evidence pack cites them separately from the IONOS CLOUD credentials.
Decision Summary
| Question | Answer for AI Model Hub | FinCorp's design action |
|---|---|---|
| Where is data processed? | IONOS CLOUD data centers in Germany | Passes the location filter |
| Is content kept or trained on? | Not logged, not stored, not used for training; metadata only | Record the metadata fields in the processing register |
| Who can access operational data? | IONOS CLOUD personnel in the EU; no sub-processors | Cite in the transfer assessment |
| Who pseudonymizes? | The customer | Pseudonymize in the application |
| Which credentials? | ISO 27001 data centers; AI Model Hub in BSI C5 scope | Cite each with its scope |
| Who is the system provider? | IONOS CLOUD, for modified and unmodified models | Keep model cards and modification documentation |
| Who logs for high-risk use? | The customer | Build the log and human oversight before high-risk use |
Summary
AI Model Hub draws a clear sovereignty boundary: inference in Germany, no content logging or storage, no training on customer data, EU-only operational access and no sub-processors. Inside that boundary, FinCorp still owns pseudonymization, its legal basis and the risk classification of each use. Under the EU AI Act, the model developers, IONOS CLOUD as system provider and FinCorp as deployer hold their roles at the same time, and a high-risk use adds logging and oversight that FinCorp builds itself.
Key Points:
- Prompts and outputs are processed only in Germany and are never logged, stored or used for training.
- Pseudonymization, the legal basis and any DPIA stay with the customer as controller.
- AI Model Hub is within BSI C5 scope; cite it per service, as in Unit 1.4.
- IONOS CLOUD is the AI system provider for every model; modifications such as quantization are documented, not a change of role.
- The deployer handles Article 50(1), (3) and (4), and builds the logs and human oversight a high-risk use requires.