EU AI Act Infrastructure Requirements: Deadlines and Duties After the AI Omnibus

The AI Act Timeline After the AI Omnibus

The AI Omnibus entered into force on 27 July 2026 and moved the high-risk deadlines, while general application of the AI Act started on 2 August 2026 as planned. The table lists every milestone of the post-Omnibus timeline from the Commission's AI Act Service Desk, plus the Omnibus's own entry into force. Two of them set the high-risk infrastructure work: 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products.

The High-Risk Classification and Infrastructure Strategy

The AI Act uses a risk-based classification. Assess the intended use against Article 6 and the Annex III categories and exceptions; not every model or startup is automatically high-risk. The revised Annex III obligations apply from 2 December 2027, with the Article 6(1) product-related requirements applying from 2 August 2028. Article 9 requires a lifecycle risk-management process. Infrastructure should support the testing, monitoring and evidence that process needs, but the Act does not prescribe low-latency GPU provisioning or a particular hosting location. Evaluate each supplier's controls and documentation against your system requirements.

The Lifecycle of Risk Management

Article 9 requires providers to identify foreseeable risks, evaluate them and adopt suitable controls throughout the system lifecycle. Test the system under relevant conditions and keep the risk-management evidence current. Lyceum GPU VMs can host those tests and are billed for allocated runtime. A European deployment may reduce some international-transfer complexity, but does not eliminate foreign-jurisdiction or access risks by itself. Verify actual processing locations, contracts and safeguards.

• Continuous Testing Infrastructure must support automated regression testing for bias and accuracy.
• Lifecycle Management Documentation must track the model from training through deployment and post-market monitoring.
• Hosting and transfers: the AI Act sets no general EU-only hosting requirement. Where personal data is involved, assess GDPR and any applicable sectoral rules, including safeguards for international transfers.

The Financial and Operational Costs of Non-Compliance

The EU AI Act backs its obligations with administrative fines set in Article 99, and for engineering leaders the penalty tiers are the clearest argument for budgeting compliance work now. Each ceiling is a fixed amount or a share of total worldwide annual turnover for the preceding financial year, whichever is higher, and national authorities set the actual fine case by case within it.

• Prohibited practices (Article 5): up to €35 million or 7% of worldwide annual turnover.
• Breaches of the obligations of providers (Article 16, which covers the Articles 8 to 15 requirements for high-risk systems), deployers (Article 26) and the Article 50 transparency rules: up to €15 million or 3%.
• Supplying incorrect, incomplete or misleading information to notified bodies or national authorities: up to €7.5 million or 1%.

Operational Disruptions and Market Access

The fine is often not the largest cost. Where a high-risk system presents a risk and does not comply, market surveillance authorities can require the operator to take corrective action, withdraw the system from the market or recall it (Article 79). For an ML team that can take months of engineering work out of production, and retrofitting data provenance and logging into a system built without them usually costs more than building them in from the start.

Proportionality for SMEs and Small Mid-Caps

Article 99 builds in proportionality: for SMEs, including start-ups, each fine is capped at whichever of the two amounts is lower. The AI Omnibus added the same lower cap for small mid-cap companies on the €15 million and €7.5 million tiers, and extends further SME simplifications, such as simplified technical documentation, to them. Even a capped fine is a serious hit for an early-stage company, which is why compliance work belongs in the roadmap rather than in a later clean-up.

Data Governance and Residency under Article 10

Article 10 requires appropriate governance for the training, validation and testing datasets of covered high-risk systems, including relevance, representativeness and, to the extent possible, freedom from errors and completeness in view of the intended purpose. Document data origin, preparation and quality controls as applicable. It does not impose a universal requirement to record every data point or mandate EU-only hosting. A GDPR DPIA is required where processing personal data is likely to create a high risk under Article 35; AI Act classification alone does not replace that assessment. Document the processing location and any applicable transfer safeguards.

Data Integrity and Bias Mitigation

The requirement for relevant, representative datasets that are, to the extent possible, free of errors and complete for the intended purpose is a technical challenge that requires sophisticated data cleaning and validation pipelines. Under Article 10, providers must implement appropriate data governance and management practices. This includes examining the original design choices and the data collection processes. Infrastructure must be able to host these large-scale validation tasks without becoming a bottleneck. Furthermore, the Act emphasizes the need to identify and mitigate potential biases that could lead to discrimination. Where those checks need special categories of personal data, Article 4a, inserted by the AI Omnibus, lets providers of high-risk systems process them only where strictly necessary and other data, including synthetic or anonymised data, would not do, under safeguards such as pseudonymisation, strict access controls and deletion once the bias is corrected. Choose bias assessment and mitigation methods appropriate to the intended use, affected groups and identified risks; this does not necessarily require a compute-intensive algorithm. Teams can run these checks on Lyceum GPU VMs, or on larger clusters available on quote. Common mistakes in data governance include failing to document the data cleaning process or using unverified third-party datasets without a quality audit. Under the high-risk rules, these oversights can lead to forced system withdrawals from the market, making a robust and transparent data infrastructure a critical asset for any AI company.

Technical Documentation and Automatic Logging

Article 11 requires technical documentation that demonstrates compliance, including relevant system design and development information. Article 12 requires automatic event logging appropriate to the system's intended purpose and traceability needs, especially for risk identification and post-market monitoring. It does not universally require storing every prompt, output, model-weight hash or GPU utilization sample. Select necessary evidence, apply access and retention controls, and minimize personal data. A serving stack you operate can provide logs, while a managed service needs a clear division of evidence and responsibilities.

The following are engineering options to assess against the system's risks and documentation needs, not a universal statutory minimum. Collect only what is necessary and available:

• Request evidence: a minimized event record or appropriate reference to the input, with personal data protected and retained only as needed.
• Version evidence: model and deployment identifiers, plus weights or container hashes where you control them and they are relevant.
• Execution evidence: relevant configuration, failures and resource conditions needed to investigate system behavior.
• Outcome evidence: decisions, fallbacks and human interventions; record calibrated confidence or other quality indicators only where the system provides meaningful ones.

Traceability and Audit Trails

The requirement for automatic logging is designed to ensure a level of traceability that allows for the monitoring of the AI system's operation and the identification of potential issues. This includes keeping logs that allow for the tracking of the system's outputs and the decisions it makes. For infrastructure teams, this means implementing a logging architecture that is both scalable and secure. The logs must be protected from unauthorized access and tampering, as they serve as the primary evidence of compliance during an audit. On GPU VMs you control, you run that logging stack yourself and decide where the logs are stored and how they are exported for analysis. Additionally, the technical documentation must be kept up to date and made available to the relevant national authorities upon request. This documentation must provide a clear and comprehensive overview of the system's design, development, and operation. By using a standardized and transparent infrastructure, teams can simplify the process of creating and maintaining this documentation, reducing the administrative burden and ensuring that they are always ready for a regulatory inspection.

Cybersecurity and Robustness: The Article 15 Mandate

Article 15 is perhaps the most technically challenging requirement of all, though it is a design-and-development duty on the provider of the high-risk AI system rather than on its hosting infrastructure. It requires high-risk systems to achieve an appropriate level of accuracy, robustness, and cybersecurity. This includes resilience against unauthorized third-party interference and AI-specific attacks such as data poisoning, model poisoning or adversarial inputs, of which prompt injection is one form. Robustness often requires technical redundancy. If a GPU node fails during a critical inference task, the system must have a fail-safe plan. Dedicated inference on Lyceum supports auto-scaling with minimum and maximum replicas, so the replica count follows load. To be direct about Lyceum's own position: it holds no ISO 27001 certificate, no SOC 2 report and no BSI C5 attestation today, and it states no EU AI Act conformity position. Lyceum states that it uses European data centres, does not train on customer data and offers a DPA on request. These supplier statements and the relevant service configuration support due diligence; they do not establish compliance for the customer's AI system.

Resilience Against Adversarial Attacks

The threat landscape for AI systems is rapidly evolving, and Article 15 specifically highlights the need for resilience against adversarial attacks. These attacks can take many forms, from data poisoning, where malicious data is introduced into the training set, to prompt injection, where a user attempts to bypass the model's safeguards. Infrastructure must be designed to detect and mitigate these threats. This includes implementing strict access controls, monitoring for unusual activity, and using secure hardware that is resistant to tampering. To meet Article 15 requirements, teams should consider deploying across multiple replicas to prevent single points of failure and regularly stress-testing models against known vulnerabilities. By taking a proactive approach to cybersecurity and robustness, providers can reduce risk and gather evidence; these controls alone do not establish full legal compliance or guarantee safety.

The Buy vs. Build Compliance Decision

As these deadlines approach, CTOs face a critical decision: build and manage their own infrastructure or use a managed European provider. Managing your own hardware involves significant cooling, maintenance, and capacity challenges. Conversely, Lyceum bills GPU VMs per second, with no subscription or base fee. The compliance moat is becoming a competitive advantage. Startups that can evidence their own EU AI Act work have an easier answer when enterprise buyers ask about regulatory risk in their supply chain. An OpenAI-compatible API can simplify migration from hyperscaler credits, but teams must validate model features, processing locations and operational requirements. GDPR controllers retain their Article 24 responsibilities, while processors have their own duties, including under Articles 28 and 32. A supplier cannot confer compliance on an AI system. Under Article 25(4) of the AI Act, as amended by the AI Omnibus, the provider of a high-risk AI system and a third party that supplies an AI system, AI model, tools, services, components or processes used or integrated in it must, by written agreement, specify the information, capabilities, technical access and other assistance the provider needs to comply. Self-serve GPU VMs, with larger clusters available on quote, mean you do not have to trade speed for European hosting.

The Strategic Value of Data Location

Choosing the right infrastructure partner is about more than just cost and performance. It is about long-term strategic alignment with the regulatory environment. In the European Union, knowing where data is processed and meeting regulatory duties are becoming key differentiators in the market. Companies that can demonstrate a commitment to these principles are better positioned to win the trust of customers and partners. A provider with European data centres and a DPA available on request, such as Lyceum, covers part of the evidence on where data is processed, but a supplier's documentation is evidence for, never a discharge of, a startup's own legal obligations. Furthermore, the ability to scale infrastructure quickly and cost-effectively is essential for staying competitive in the fast-moving AI space. Per-second billing for GPU VMs and serverless training, with no long-term contract required, gives startups room to experiment and grow. As these deadlines near, teams should select infrastructure by the evidence, controls, performance and contractual terms their systems require.

Navigating Conformity Assessments under Article 43

Under Article 43, high-risk AI systems must undergo a conformity assessment before being placed on the market or put into service. This process ensures that the system meets all the requirements set out in the Act, including those related to risk management, data governance, and technical documentation. For many systems, this can be a self-assessment, but for others, particularly those involving biometric identification, a third-party notified body may be required. The infrastructure used to develop and host the AI system plays a critical role in this assessment. Auditors will look for evidence that the technical requirements have been met, which requires a high degree of observability and transparency in the underlying hardware and software stack. On GPU VMs and endpoints you control, your own logging captures hardware utilisation and data-processing events as that evidence.

Infrastructure as Evidence

The ability to provide a clear audit trail is a significant advantage during the conformity assessment process. If a system is hosted on a platform that does not provide granular control over the hardware layer, it can be difficult to prove that the system is operating as intended. On Lyceum GPU VMs, with raw GPU access over SSH, teams can record the hardware configuration of each run as one element of that evidence, but a supplier's documentation is evidence for, never a discharge of, their own compliance obligations. Furthermore, the Act requires that a new conformity assessment be carried out whenever a substantial modification is made to the system. Having a flexible and transparent infrastructure makes it easier to track these modifications and ensure that the system remains compliant throughout its lifecycle. This proactive approach to compliance not only satisfies regulators but also builds trust with customers and partners who are increasingly concerned about the legal and ethical implications of AI.

Post-Market Monitoring and Continuous Compliance

The obligations of a high-risk AI provider do not end once the system is deployed. Article 72 introduces the requirement for post-market monitoring, which involves the continuous collection and analysis of data on the performance of the AI system. This is intended to identify any potential risks or malfunctions that may arise during real-world use. From an infrastructure perspective, monitoring should be proportionate to the system's nature and risks; the Act does not prescribe a universal real-time telemetry architecture. The system must be able to log performance metrics, user interactions, and any incidents that could indicate a failure to comply with the Act's requirements. On infrastructure you can observe end to end, such as self-managed GPU VMs, that telemetry stays under your control and can be exported for regulatory reporting.

Continuous Compliance in Production

Post-market monitoring is a critical component of the broader compliance strategy. It allows providers to detect and address issues before they escalate into significant problems. For example, if a model begins to exhibit bias or a decrease in accuracy after deployment, the monitoring system should trigger an alert, allowing the team to take corrective action. This might involve re-training the model on more representative data or adjusting the system's parameters. The infrastructure must be able to support these updates without causing significant downtime or disruption to the service. Dedicated inference endpoints on Lyceum auto-scale between a minimum and a maximum number of replicas, which a team can set to the capacity a rollout needs. Additionally, Article 73 requires reporting qualifying serious incidents under its specified conditions and deadlines; not every ordinary malfunction is automatically reportable. Having a detailed and accurate log of the system's operation is essential for fulfilling this reporting requirement and demonstrating that the provider has taken appropriate steps to mitigate the issue. This ongoing commitment to transparency and accountability is a key requirement for high-risk AI systems under the AI Act.

Implementing Human Oversight through Technical Design

Article 14 of the EU AI Act mandates that high-risk AI systems must be designed and developed in a way that allows for effective human oversight. The goal is to prevent or minimize the risks to health, safety, or fundamental rights that may arise when an AI system is used. This oversight can be achieved through technical measures built into the system, as well as through the provision of clear instructions and tools for the human operators. From an infrastructure and design standpoint, this means that the system must have interfaces that allow humans to monitor its operation, intervene when necessary, and even shut the system down if it poses a risk. Oversight measures must be proportionate to the risks, autonomy and context of use. Article 14 includes intervention or interruption through a stop control or similar procedure that brings the system to a safe state; abruptly powering off hardware is not necessarily the appropriate implementation.

Technical Measures for Oversight

The technical implementation of human oversight requires a deep understanding of the system's decision-making process. This is often referred to as interpretability or explainability. Infrastructure leads must ensure that the tools used to deploy and manage the AI system provide the necessary transparency to support human intervention. For instance, the system should be able to provide reasons for its outputs, allowing a human operator to verify the accuracy and fairness of the decision. An open-source serving stack keeps the serving layer inspectable, so inputs, outputs and system state can all be logged. Interpretability of the model itself remains a modeling problem rather than an infrastructure one. Furthermore, the infrastructure must support the implementation of constraints and safeguards that prevent the system from taking unauthorized or harmful actions. This might include setting limits on the system's autonomy or requiring human approval for certain high-stakes decisions. By building these oversight mechanisms into the infrastructure from the ground up, providers can ensure that their systems remain under human control and comply with the strict requirements of Article 14. This not only meets regulatory demands but also enhances the safety and reliability of the AI system in real-world applications.

A Practical Compliance Framework for Engineering Teams

With the 2 August 2026 general application date passed and the Annex III high-risk obligations applying from 2 December 2027, ML engineering teams have a fixed window to build compliance into their pipelines rather than bolt it on. In short, the work runs in five steps: inventory every AI system and classify it against Article 6 and the Annex III use cases; put data governance controls such as dataset versioning and bias checks into the training pipeline; audit your infrastructure provider on the country of each instance, a data processing agreement and exportable logs; generate the Annex IV evidence pack as code; and monitor production continuously so human oversight and post-market monitoring have data to act on. Our EU AI Act compliance checklist for developers walks through each step in detail.