The EU AI Act entered into force on 1 August 2024 and begins phased application from February 2025 onwards. It is the first comprehensive legal framework governing artificial intelligence systems in any major economy. For infrastructure engineers, its implications extend well beyond software and algorithms. The Act creates direct obligations around the physical systems that host, power, and cool AI workloads, particularly for systems classified as high-risk under Annex III. Understanding those obligations, and engineering infrastructure that meets them, is now a core competency for any consultancy advising clients who deploy AI at scale in, or who are selling into, the European market. This article examines how the Act interacts with the EN 50600 series of data centre standards, and what that combined regulatory and technical environment means in practice for engineers, facilities managers, and technical procurement leads. | The Act classifies AI systems by risk level. High-risk systems, which include AI used in critical infrastructure management, education, employment, essential private and public services, and law enforcement, carry the heaviest compliance burden. Providers and deployers of high-risk systems must implement risk management systems, ensure data governance, maintain technical documentation, enable human oversight, and achieve appropriate levels of accuracy, robustness, and cybersecurity. Article 9 requires a risk management system that is continuous throughout the lifecycle of the AI system. For infrastructure engineers, this means the physical layer, including power distribution, cooling architecture, network topology, and monitoring systems, must be designed to support that continuous risk management. An AI system that becomes unpredictable due to thermal throttling, power fluctuation, or network latency is not meeting the robustness requirements of Article 15, which mandates that high-risk AI systems be resilient to errors, faults, and inconsistencies. Infrastructure instability is therefore a compliance risk, not merely an operational inconvenience. | Article 17 requires providers of high-risk AI systems to establish a quality management system covering the design, development, and deployment of the AI, including the resources used. In practice, this means the compute and data centre infrastructure. Engineers designing racks, power feeds, cooling circuits, and monitoring instrumentation for AI workloads must now produce documentation that can feed into an Article 17 quality management file. This is a genuine change in the engineering deliverable. The Act also introduces, at Article 51, obligations for providers of general-purpose AI models with systemic risk, defined partly by training compute above 10 to the power of 25 floating-point operations. The infrastructure required to train such models, large GPU clusters drawing megawatts of power, falls squarely within scope. Providers must conduct adversarial testing and report serious incidents, both of which depend on reliable infrastructure monitoring and logging at the physical layer. | The EN 50600 series, published by CENELEC and progressively updated, is the European standard suite for data centre facilities and infrastructure. It covers power supply systems (EN 50600-2-2), environmental control (EN 50600-2-3), telecommunications cabling (EN 50600-2-4), security systems (EN 50600-2-5), and management and operation (EN 50600-3-1), among others. The series defines four classes of availability (Class 1 to Class 4), allowing engineers and clients to specify infrastructure resilience proportionate to the criticality of the workload. For AI workloads running high-risk systems under the EU AI Act, the relevant question is which EN 50600 availability class is appropriate. A high-risk AI system governing critical infrastructure, for example a predictive maintenance system for an energy grid, would logically require Class 3 or Class 4 availability (concurrent maintainability and fault tolerance respectively), because downtime or degraded performance would directly undermine the robustness requirements of Article 15. Choosing Class 2 (redundancy) for such a workload may be technically defensible in some contexts, but engineers must be able to justify that choice within the quality management documentation required by Article 17. | EN 50600-3-1, covering management and operation, specifies requirements for key performance indicators including Power Usage Effectiveness (PUE), Water Usage Effectiveness (WUE), and Energy Reuse Factor (ERF). These are not optional metrics for AI infrastructure in the EU context. The EU Energy Efficiency Directive (EED), recast as Directive 2023/1791 and in force from October 2023, requires data centres above 500 kW IT load serving the EU market to report PUE, WUE, ERF, and the Renewable Energy Factor (REF) annually to a European database managed by the Commission. Non-compliance carries regulatory risk for operators. According to the European Commission's Joint Research Centre, data centres in the EU consumed approximately 76.8 TWh of electricity in 2018, and projections under various AI adoption scenarios suggest that figure could rise significantly by 2030. That trajectory is precisely why the EED recast has tightened reporting obligations and why EN 50600 metrics are now embedded in regulatory compliance rather than being solely a voluntary efficiency benchmark. | AI training and inference workloads are characterised by high, dense, and often variable power draw. A rack of GPU accelerators used for large model inference may draw between 15 kW and 80 kW per rack depending on the hardware generation and configuration, compared with 5 kW to 10 kW for conventional enterprise compute. This density drives specific engineering decisions at the power distribution unit (PDU), uninterruptible power supply (UPS), and upstream transformer level. For EN 50600 Class 3 or Class 4 compliance, power paths must be concurrently maintainable or fully fault-tolerant. That means 2N or 2(N+1) UPS configurations, dual power feeds from separate substations where possible, and automatic transfer switching with transfer times that do not exceed the hold-up time of downstream equipment. GPU servers typically have shorter hold-up tolerance than traditional servers, so the UPS specification must account for this. Engineers must also consider harmonic distortion introduced by large numbers of switch-mode power supplies operating simultaneously, which can affect power quality and accelerate UPS battery degradation if not mitigated at the design stage. | Conventional air-cooled data centre design, optimised for 5 kW to 10 kW racks, is not adequate for AI infrastructure at 30 kW to 80 kW per rack. At those densities, traditional raised-floor or overhead air delivery systems cannot maintain inlet temperatures within the ASHRAE A2 or A3 envelope (which EN 50600-2-3 references) without prohibitive airflow volumes that are themselves energy-intensive. The engineering response is a shift toward direct liquid cooling (DLC), rear-door heat exchangers (RDHx), or full immersion cooling. Each has distinct implications for PUE. A well-designed DLC system can achieve PUE values approaching 1.03 to 1.10, compared with 1.4 to 1.6 for legacy air-cooled facilities. Since PUE is a mandatory EED reporting metric, the cooling architecture choice is simultaneously a thermal engineering decision and a regulatory compliance decision. According to the Uptime Institute's Global Data Center Survey 2023, the average PUE for data centres globally was 1.58, with newer hyperscale and AI-optimised facilities achieving considerably lower values through liquid cooling adoption. For EU-market clients facing EED reporting obligations, a PUE of 1.58 is likely to attract regulatory scrutiny as the Commission's reporting database matures and benchmarks become more transparent. | Article 15 of the EU AI Act requires that high-risk AI systems achieve appropriate levels of accuracy and be resilient to errors and inconsistencies that could lead to prohibited outputs or dangerous behaviour. At the infrastructure layer, this translates to a requirement for real-time monitoring of the environmental and power conditions under which AI compute operates, with logging sufficient to demonstrate post-incident that the infrastructure was within specification at the time of any system fault. EN 50600-3-1 specifies a monitoring framework that includes DCIM (Data Centre Infrastructure Management) systems tracking power, cooling, and environmental parameters at rack and row level. For AI infrastructure under the Act, this DCIM data should be retained and linked to the AI system's operational logs so that any performance anomaly can be correlated with physical infrastructure state. This is not standard practice in most existing data centres and represents a genuine integration engineering challenge between IT operations and facilities management teams. Digital twin technology is increasingly being used to create live, validated models of data centre physical infrastructure that can be queried alongside IT system logs. A digital twin of the cooling and power infrastructure enables engineers and operators to replay historical conditions, identify correlations between thermal events and AI model performance degradation, and support the continuous risk management cycle required by Article 9 of the Act. | The Article 17 quality management file for a high-risk AI system must include documentation of the computational resources used. For bespoke or co-located AI infrastructure, this means the engineering record of the physical facility must be audit-ready. Building Information Modelling (BIM) at the appropriate level of information need, structured to ISO 19650, is a practical method for producing and maintaining that infrastructure documentation in a way that is both machine-readable and version-controlled. ISO 19650 information management is not yet explicitly cited in EU AI Act guidance, but its structured approach to information containers and delivery milestones maps naturally onto the documentation lifecycle that Article 17 implies. Engineers scoping AI infrastructure for EU-market clients should ensure that all engineering deliverables are structured within an ISO 19650-compliant BIM environment from the outset of the project, so that the as-built and as-operated records are accessible for regulatory audit, change management, and ongoing quality management purposes. | Five practical steps follow for engineers scoping AI infrastructure projects in the UK and EU context. First, classify the AI workload by EU AI Act risk tier before finalising the infrastructure specification, because high-risk classification drives EN 50600 availability class selection and documentation requirements. Second, confirm whether the facility is subject to EED Directive 2023/1791 reporting obligations (IT load above 500 kW serving the EU market), and if so, ensure PUE, WUE, ERF, and REF are measurable and reportable from day one of operation. Third, assess rack power density against cooling technology options at schematic design stage, noting that air cooling above 20 kW per rack carries efficiency and reliability risk for AI workloads, and that liquid cooling options should be evaluated early. Fourth, integrate DCIM infrastructure monitoring with AI system operational logging to support Article 15 robustness evidence and Article 9 continuous risk management requirements. Fifth, structure all engineering deliverables within an ISO 19650-compliant BIM environment to support Article 17 quality management documentation throughout the facility lifecycle. The combination of EU AI Act compliance obligations and EN 50600 technical requirements creates a multi-disciplinary engineering challenge that spans power, cooling, structural, BIM, and digital systems engineering. Addressing each of those disciplines in an integrated manner, from schematic design through to operational monitoring, is the foundation of infrastructure that is both technically sound and demonstrably compliant.