Before a single GPU cycle is spent on training, you must establish a lawful basis for processing personal data under Article 6 of the GDPR. While user consent was once the default, the scale of modern datasets makes explicit consent for every data point functionally impossible. In Opinion 28/2024 on the processing of personal data in the context of AI models, adopted 17 December 2024, the EDPB set out how controllers can rely on Legitimate Interest (Article 6(1)(f)) for model training, while recalling that the GDPR sets no hierarchy between the legal bases [1]. This legal basis requires a delicate balance between the commercial or research interests of the AI developer and the fundamental rights of the data subjects whose information is being processed.

Establishing a Lawful Foundation for Machine Learning

The three-part test for legitimate interest is now more stringent than ever. First, you must identify the specific interest you are pursuing. Second, you must show that the processing is necessary for it, with no less intrusive alternative available. Third, you must run the balancing test and show that your interest is not overridden by the interests and fundamental rights of the data subjects; an unconditional and easily accessible opt-out is one of the mitigating measures that weighs in that balance [1]. The proposed 2025 Digital Omnibus, if adopted, would create a narrow exemption for processing sensitive data, such as health or biometric information, when it is strictly necessary to detect and correct bias in high-risk AI systems. It has not been adopted, so it changes nothing today, and even as drafted it would not be a blanket permission: it would require extensive documentation to prove that no other less-intrusive method could achieve the same result.

Transparency is another pillar of this legal framework. Under the AI Act's general-purpose AI obligations, which have applied since 2 August 2025, providers must publish a public summary of the content used for training. This summary must include the sources of the data and a description of how copyrighted materials were handled. Failing to document this basis can be catastrophic. Since 2018, EU regulators have issued billions of euros in GDPR fines, and the pace of enforcement has risen sharply since 2023. For a startup with 15-100 employees, a single enforcement action regarding training data provenance can be a terminal event. Lyceum helps mitigate these risks by providing the infrastructure needed to maintain data logs and audit trails. The legal assessment of the training process itself stays with you as controller.

Data Residency and the Sovereignty Gap

While the GDPR does not explicitly mandate that data stay within the EU, the legal reality of 2026 makes non-EU hosting a high-risk strategy. The EU-US Data Privacy Framework remains in force and legally valid as of April 2026, the date of this article, having been upheld at first instance with an appeal (C-703/25 P) pending, so teams retain a stable mechanism for transferring personal data to US-based providers. This has created what we call the Sovereignty Gap: the distance between a developer compliance obligations and their infrastructure physical location. When you process data on US-hosted GPUs, that data can be reached by US legal process under the US Cloud Act, although on the DOJ's own account the Act gave US courts no expanded jurisdiction, permits no bulk collection, and whether a provider is subject to US jurisdiction is a highly fact-dependent analysis. This directly conflicts with Article 48 of the GDPR, which restricts the disclosure of personal data to third-country authorities unless there is a specific international agreement in place.

Sovereignty and the Jurisdictional Conflict

For teams in regulated sectors like healthcare or manufacturing, this conflict is a deal-breaker. Relying on US-based hyperscalers means that your training data could be accessed by foreign intelligence services, putting you in direct violation of EU law. This risk remains even if the data is encrypted, as the keys are often held by the service provider or the data is decrypted during the actual training process. To navigate this, many firms are turning to Transfer Impact Assessments (TIAs), but these are complex, expensive, and often fail to provide the legal certainty required for large-scale ML projects.

Lyceum addresses this by running GPU workloads in European data centres in Spain, Paris and the Nordics, operated by a European company. Compute and storage for GPU workloads stay in the EU; in the serverless catalogue, four named models are global-hosted and never receive your traffic unless you explicitly select them. This eliminates the need for complex TIAs and ensures that your training runs are shielded from extra-territorial legal reach. By choosing a provider that processes your data in the EU you can turn regulation from a hurdle into a competitive moat, but compliance cannot be conferred by your host: the provider is typically the processor while you remain the controller, so lawful basis, transparency and data subject rights stay with you. Our platform makes data residency a property of where your jobs actually run: GPU compute and storage sit in European data centres, and the four global-hosted serverless models receive traffic only if you select them. This level of sovereignty is essential for building trust with European enterprise clients who are increasingly wary of the legal liabilities associated with non-EU data processing.

Technical Measures for Regulation-by-Design

Compliance is a technical challenge as much as a legal one. Implementing Regulation-by-Design means embedding privacy controls directly into your ML pipelines. In 2026, the standard for anonymization has become significantly higher, as regulators recognize that LLMs can often regurgitate training data through model inversion attacks. This means that removing names and addresses is no longer sufficient to consider a dataset truly anonymous. Instead, engineering teams must look toward more advanced technical measures to protect individual privacy while maintaining the utility of the training data.

Privacy-Preserving Engineering Workflows

One of the most effective methods is pseudonymization, which involves replacing direct identifiers with cryptographically secure pseudonyms. The proposed 2025 Digital Omnibus would, if adopted, make training on pseudonymized data more legally defensible, provided the mapping keys are stored in a separate, secure environment. Another critical technique is differential privacy, which involves injecting noise into the training process. This ensures that the model learns general patterns without memorizing specific individual data points, making it much harder for an attacker to extract personal information from the final model weights. This is increasingly becoming a requirement for models deployed in the public sector or other high-sensitivity environments.

Data minimization is also a core requirement under Article 5(1)(c) of the GDPR. You must audit your datasets to ensure you are only processing the features necessary for the model objective. Collecting data just in case is a direct violation of this principle. The EDPS orientations on generative AI set out practical expectations for minimization across the model lifecycle [2]. To support these workflows, Lyceum offers on-demand VM provisioning and per-second billing with no base fee, allowing you to spin up isolated environments for data cleaning and pseudonymization without the overhead of long-term commitments. This flexibility is essential for teams that need to run frequent compliance audits on their training sets. By using Lyceum, you can automate the deployment of privacy-preserving pipelines, ensuring that your data is protected from the moment it enters the training environment.

The AI Act Intersection: High-Risk Systems

2 August 2026 is the AI Act's general date of application, with the Chapter III high-risk obligations deferred to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems [3]. If your model is used in hiring, credit scoring, medical diagnostics, or critical infrastructure, you face a new tier of obligations under Article 10 (Data Governance) and Article 11 (Technical Documentation). These requirements go beyond the privacy focus of the GDPR and move into the realm of system safety and reliability. High-risk systems must be developed using datasets that are relevant, representative, and, to the best extent possible, free of errors. This requires a rigorous Data Protection Impact Assessment (DPIA) before training begins to identify and mitigate potential risks to fundamental rights.

Compliance Obligations for High-Risk AI Systems

Article 10 of the AI Act specifically mandates that training, validation, and testing data sets shall be subject to appropriate data governance and management practices. This includes an evaluation of the data for possible biases that could lead to discriminatory outcomes. Furthermore, Article 11 requires the creation of a Technical File that provides a detailed description of the model, its training process, and its performance metrics. This documentation must be kept up to date and made available to national competent authorities upon request. For many startups, the administrative burden of these requirements can be overwhelming, but failing to comply can result in fines of up to EUR 35 000 000 or 7% of total worldwide annual turnover, whichever is higher - a top tier that applies only to breach of the Article 5 prohibitions.

Lyceum platform is designed to facilitate this by providing full transparency into the underlying hardware stack. We utilize vLLM and NVIDIA Dynamo rather than the black-box proprietary engines found in many US-based alternatives. This allows developers to have full visibility into how their models are being trained and executed, which is a key requirement for meeting the transparency and documentation standards of the AI Act. By providing a sovereign and transparent environment, Lyceum can support the technical file you build for conformity assessments, although the obligations for placing a high-risk AI system on the European market rest with the system's provider and cannot be discharged by a supplier of compute.

Infrastructure Economics: Hyperscalers vs. Sovereign Clouds

For AI startups transitioning off hyperscaler credits, the cost of compliance often collides with the reality of GPU pricing. Hyperscalers frequently require block-reservations for H100s, and their egress fees can take a large bite out of a training budget. The real exposure is jurisdictional rather than geographic: the major clouds do operate EU regions, and the harder problem is that a US-incorporated provider stays within reach of US legal process wherever the bytes sit. When you factor in the legal costs of managing complex data transfer agreements and the risk of regulatory fines, the true cost of using a non-EU provider becomes significantly higher than the sticker price of the compute.

Optimizing Training Costs in a Regulated Environment

Lyceum offers a structural cost advantage by running GPU workloads in European data centres and charging no egress fees. Our on-demand H100 VMs are listed at $2.79 per GPU-hour ($3.59 per GPU-hour for dedicated inference and serverless training), billed per second with no base fee; current rates are on the pricing page. This pricing model is designed to be transparent and predictable, allowing you to scale your training jobs without worrying about hidden costs. When combined with Lyceum's scheduling product, which predicts memory use and runtime within a node, teams can place jobs on the cheapest GPU that actually fits instead of the largest one free. This scheduler allows for more efficient resource allocation, ensuring that you are not paying for idle GPU time.

By moving to an EU-native platform, you avoid the compliance tax of operating the GDPR's Chapter V transfer tools - adequacy decisions, standard contractual clauses or binding corporate rules - which are the lawful routes for US-based hosting rather than legal workarounds. You get raw GPU access via SSH, provisioned on demand, with GPU compute and storage running in European data centres. This allows your engineering team to focus on model performance rather than jurisdictional mapping. In the competitive landscape of 2026, the ability to train models efficiently and compliantly is a major advantage. Lyceum provides the high-performance hardware you need at a price point that makes sense for growing companies, with GDPR-compliant processing in European data centres. Lyceum publishes no EU AI Act conformity position; the Act's duties sit with the provider of the AI system.

Data Provenance and Technical Documentation Requirements

Under Article 11 of the EU AI Act, the concept of data provenance has moved from a best practice to a mandatory legal requirement. This involves maintaining a detailed record of the origin of all data used in the training process. For AI developers, this means you must be able to trace every data point back to its source and demonstrate that it was collected and processed in accordance with both the GDPR and the AI Act. This traceability is essential for ensuring the quality and reliability of the model, as well as for providing the necessary documentation for regulatory audits.

Maintaining Technical Documentation and Traceability

The technical documentation required for high-risk systems must include a description of the data collection processes, the data cleaning and preparation steps, and the metrics used to evaluate the quality of the datasets. This is not just a one-time task but an ongoing requirement throughout the lifecycle of the AI system. If the model is updated or fine-tuned with new data, the technical documentation must be updated accordingly. This level of detail is necessary to ensure that the AI system is transparent and that its behavior can be understood and explained by human overseers. Because the stack is open (vLLM, NVIDIA Dynamo, TensorRT-LLM), you keep visibility into the inference path and can export the execution records your documentation requires.

Furthermore, the AI Act requires that developers provide a summary of the training data to the public. This summary must be sufficiently detailed to allow third parties to understand the types of data used and the measures taken to protect privacy and intellectual property. For many companies, this represents a significant shift toward greater transparency. By using a sovereign cloud provider like Lyceum, you can ensure that your data provenance records are stored in a secure, EU-based environment, protected from unauthorized access. This not only helps you meet your legal obligations but also builds trust with your users and stakeholders, who are increasingly concerned about the ethical and legal implications of AI.

Bias Detection and Data Quality Standards

Article 10 of the EU AI Act sets out rigorous standards for the quality of datasets used in high-risk AI systems. Under Article 10(3), training, validation and testing data sets must be relevant, sufficiently representative, and, to the best extent possible, free of errors and complete in view of the intended purpose [4]. This is particularly critical for preventing algorithmic bias, which can lead to discriminatory outcomes in areas like hiring or law enforcement. To meet these standards, developers must implement robust data governance practices that include bias detection and mitigation strategies at every stage of the ML pipeline.

Ensuring Data Quality and Algorithmic Fairness

Bias can enter an AI system in many ways, from the initial selection of data sources to the way the data is labeled and processed. To mitigate this risk, the AI Act requires developers to perform statistical analyses of their datasets to identify potential biases. This involves examining the representativeness of the data across different demographic groups and ensuring that the model does not learn or amplify existing societal prejudices. If bias is detected, developers must take corrective measures, which may involve collecting additional data or adjusting the training algorithms. Article 10(5) of the AI Act already permits providers of high-risk systems to process special categories of personal data where this is strictly necessary for bias detection and correction, subject to strict safeguards including deletion once the bias has been corrected [4].

Lyceum provides the high-performance compute needed to run these complex bias detection and mitigation tasks. Our H100 VMs allow you to process large datasets quickly, enabling more frequent and thorough quality checks. By integrating bias detection into your regular training workflow, you can ensure that your models are not only accurate but also fair and compliant with the AI Act. This focus on data quality is not just about avoiding fines; it is about building better, more reliable AI systems that perform consistently across all user groups. In a market where trust is a key differentiator, demonstrating a commitment to data quality and fairness can give your AI products a significant competitive edge.

Data Subject Rights and the Right to Object

The GDPR grants individuals several key rights regarding their personal data, and these rights apply just as much to AI training as they do to any other form of data processing. Two of the most important rights in the context of machine learning are the Right to Erasure (Article 17) and the Right to Object (Article 21). If an individual objects to their data being used for AI training, or if they request that their data be deleted, you must have a process in place to honor that request. This can be technically challenging, especially if the data has already been incorporated into a trained model.

Implementing Effective Opt-Out and Erasure Mechanisms

The right to object is particularly relevant when you are relying on Legitimate Interest as your legal basis. Under Article 21, individuals have the right to object to processing based on legitimate interests at any time. If they do, you must stop processing their data unless you can demonstrate compelling legitimate grounds that override their interests. In the context of AI, this means you must provide a clear and easy way for users to opt-out of having their data used for training. This opt-out must be unconditional, meaning you cannot penalize users for exercising their rights. Failing to provide a functional opt-out mechanism is a common source of regulatory scrutiny and can lead to significant fines.

The right to erasure, or the right to be forgotten, presents even greater technical hurdles. If a data subject requests the deletion of their data, you must remove it from your training sets. However, there is an ongoing debate about whether this also requires the removal of the data influence from the trained model itself. While the law is still evolving in this area, the 2025 EDPB orientations suggest that developers should implement measures to ensure that personal data can be effectively removed from the training pipeline. Lyceum sovereign infrastructure allows you to maintain granular control over your datasets, making it easier to identify and remove specific data points when a request is made. By building these capabilities into your infrastructure from the start, you can ensure that you are prepared to meet the growing demands of data subject rights in the AI era.

Sources

[1] EDPB, Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, adopted 17 December 2024, read 3 August 2026; [2] EDPS, Generative AI and the EUDPR, orientations for ensuring data protection compliance when using generative AI systems, version 2, 28 October 2025, read 3 August 2026; [3] European Commission, AI Act regulatory framework, dates of application, read 3 August 2026; [4] Regulation (EU) 2024/1689, Article 10 (Data and data governance), AI Act Explorer reproduction of the official text, read 3 August 2026.