I. Introduction: an infrastructure that European law can describe but not yet govern
At 12:33 Central European Time on 28 April 2025, the Iberian electricity system separated from the synchronous area of continental Europe and collapsed. Some sixty million people in Spain and Portugal lost supply, in what remains the most severe European power incident in more than two decades. The final report of the expert panel convened by the European Network of Transmission System Operators for Electricity (ENTSO-E), published on 20 March 2026 after an eleven-month investigation by a forty-nine-member panel drawn from transmission system operators, regional coordination centres, national regulatory authorities and the Agency for the Cooperation of Energy Regulators (ACER), attributed the event to a compound of failures in voltage and reactive-power control, oscillatory instability, and cascading disconnections of generation – and pointedly not to the penetration of renewable generation as such.Footnote 1 Of the twenty-two recommendations it issued, several concern the behaviour of connected facilities under system stress: the adequacy of their protection settings, the transparency of their disconnection thresholds, and the accuracy with which system operators can model what they will do when the frequency or the voltage moves.Footnote 2
The Iberian event was not caused by data centres. It is nonetheless the most instructive available lens on the problem this article addresses, because it demonstrates at continental scale the class of risk that arises when the aggregate behaviour of connected facilities under stress is imperfectly known to the entity responsible for the integrity of the system. Hyperscale data centres are the fastest-growing category of very large connected load in Europe.Footnote 3 Their artificial-intelligence workloads move between idle and full draw within a second or less; their protection systems are engineered to disconnect from the grid in response to disturbances that conventional industrial protection would ride through; and they cluster geographically to a degree that converts a facility risk into a common-mode system risk.Footnote 4 Union law knows all of this. What it has not decided is what to do about it.
That indecision is the subject of this article. The argument is not that Union law treats hyperscale data centres incoherently; the claim of incoherence, which has begun to circulate in the commentary, over-describes the problem and prescribes accordingly.Footnote 5 The argument is that the Union has completed the first half of a risk-regulation cycle and has not begun the second. It has built, with some sophistication, an apparatus of risk identification and assessment: mandatory reporting under Article 12 of the recast Energy Efficiency Directive,Footnote 6 a common Union rating scheme in embryo under Delegated Regulation (EU) 2024/1364,Footnote 7 and resilience and incident-reporting duties under the NIS2 and CER Directives.Footnote 8 What it has withheld is risk management: no binding minimum performance standard, no load-class categorisation in the connection codes, no conditionality between industrial-policy advantage and demonstrated performance, and no operationalised cost-causation discipline governing who pays for the reinforcement that hyperscale growth occasions.
The distinction between assessment and management is the organising distinction of European risk regulation itself, inherited from food and chemicals law and generalised across the internal market.Footnote 9 Its normative point is the separation of what is known from what is to be done. Its pathology, familiar to readers of this journal, is that the separation can become a resting place: assessment is performed, management deferred, and the deferral presented as caution when it is in substance abdication. Union law on hyperscale data centres is presently in that condition, and the empirical demonstration of it arrived in July 2025, when the Commission published the first assessment of the data collected under Article 12.
Section II sets out the method, situates the argument in the legal and technical literature, explains what “hyperscale” denotes here and the vocabulary the article uses, and acknowledges an evidentiary limitation. Section III characterises the hazard. Section IV examines the assessment apparatus and its first empirical yield. Section V turns to the counter-current: the Cloud and AI Development Act proposed on 3 June 2026, which accelerates deployment without conditioning it on performance. Section VI examines the connection and market-design instruments; Section VII the fragmentation that Union abstention has produced across Ireland, Germany, the Netherlands and Spain; and Section VIII the distributional dimension. Section IX asks what the precautionary principle can be made to bear, and Section X concludes.
II. Method, scope, literature and definitions
1. Method and sources
This article is doctrinal in method and comparative in ambit, and it reads the acquis against a framework drawn from risk-regulation scholarship rather than from energy law alone. The Union sources examined are the recast Energy Efficiency Directive and its first-phase delegated act; the electricity market design reform of June 2024;Footnote 10 the Demand Connection Code;Footnote 11 the NIS2 and CER Directives; ACER Recommendation 01/2025 on a Network Code on Demand Response;Footnote 12 the Artificial Intelligence Act;Footnote 13 the European Grids Package of December 2025; the Cloud and AI Development Act proposal; and the European Water Resilience Strategy of June 2025.Footnote 14 For each national jurisdiction the sources are the statutory and regulatory instruments adopted between January 2023 and the time of writing. Because the hazard at issue is a matter of electrical engineering before it is a matter of law, the article also draws on the technical literature of system operators and of the operators of training clusters themselves, and Section III is explicit about which propositions rest on it.
The organising framework is the standard sequence of the risk-regulation cycle – identification, assessment, management, communication and review – and the question put to each instrument is which stage it occupies.Footnote 15 The framework is diagnostic rather than merely descriptive: it surfaces a failure mode that a purely doctrinal reading does not, namely that an acquis may be internally consistent, correctly transposed and faithfully applied and yet fail, because every one of its instruments sits on the same side of the assessment–management line.
Four jurisdictions are examined, and unequally. Ireland receives the fullest treatment because the underlying pressure is most acute there and the regulatory response most developed. Germany is examined because its transposition of Article 12 is the most stringent in the Union, and therefore the best available test of what happens when a Member State supplies the management stage that Union law omits. The Netherlands is examined because its congestion-prioritisation framework poses, more sharply than any other, the question whether a data centre may lawfully be placed behind other users in a connection queue. Spain is examined because its firm/flexible access distinction is the most doctrinally transparent operationalisation of the new Article 6a of Directive 2019/944. The United Kingdom is set aside: its post-Brexit divergence is interesting, but the question of what Union law should do is not much illuminated by a jurisdiction no longer bound by it.
2. The literature and the contribution of this article
The legal literature on data centres has grown quickly, and its centre of gravity is sustainability rather than system stability. The most complete doctrinal map of the Union’s data-centre acquis is that of Commins and Irion, who trace the instruments directed at climate-neutral, energy-efficient data centres by 2030 – the reporting regime of Article 12 at their centre – and conclude that an information regime of this kind should be converted into binding sustainability targets, both to contain the rebound effects of a growing industry and to counter the “waterbed” effect by which capacity refused in one jurisdiction reappears in another.Footnote 16 Hacker’s foundational treatment of sustainable AI regulation identified environmental impact as the blind spot of an AI governance built around trustworthiness and fundamental rights, and proposed an escalating ladder of disclosure, sustainability impact assessment and, ultimately, consumption limits;Footnote 17 Ebert, Alder, Herbrich and Hacker then joined the facility layer of the Energy Efficiency Directive to the model layer of the AI Act, showed that the two regimes fail to meet – the one regulates the building and cannot see the workload, the other regulates the model and stops at training – and recommended binding PUE and renewable-energy targets at Union level on the German model.Footnote 18 Beyond the law, the energy-systems literature has established that the growth of data-centre electricity consumption was for a decade offset by efficiency gains, and that the headline estimates on which policy relies are of uneven quality;Footnote 19 the environmental-science literature has quantified a water footprint that the electricity acquis does not see;Footnote 20 and the engineering and reliability literature – the work of the operators of training clusters on the power oscillations those clusters produce, and the incident investigations of the North American Electric Reliability Corporation (NERC) and, since 2025, of ENTSO-E – documents the behaviour that Section III describes.Footnote 21 Within energy law, the closest neighbours are the work on grid congestion and connection prioritisation and the emerging literature on co-located generation for data centres.Footnote 22 This article shares the legal literature’s diagnosis that reporting alone will not suffice and differs from it in treating the hazard as one of system stability rather than of aggregate footprint – how the facility class behaves in the seconds after a fault, on which the legal literature has been silent – in locating the deficiency in the structure of the regulatory cycle rather than in the absence of a particular target, and in bringing into the legal analysis a body of technical evidence that the legal literature has not yet absorbed.
3. Types, scale and workload: what “hyperscale” denotes here
Union law defines a data centre but not a hyperscale one. The definition, borrowed from the energy-statistics Regulation and incorporated by reference in the recast Energy Efficiency Directive and in the Cloud and AI Development Act proposal alike, is a structure “used to house, connect and operate computer systems/servers and associated equipment for data storage, processing and/or distribution”: it is blind to size and to workload.Footnote 23 The 500 kW threshold of Article 12(1) is not a definition of anything; it is the perimeter of a disclosure obligation, and the Commission’s consultants have since recommended extending reporting down to 100 kW. The first-phase delegated regulation adds a typology, but it is one of ownership and service model rather than of scale – “enterprise,” “colocation” and “co-hosting” data centres, according to whether the operator serves itself, houses its customers’ equipment or supplies equipment and building alike as a service.Footnote 24 A 500 kW server room and a 300 MW training campus are reported under the same three labels, and the legal vocabulary is therefore no guide to the physical reality.
That reality is described by the industry and the energy agencies in terms of size and operator. Enterprise facilities are typically measured in hundreds of kilowatts to a few megawatts of IT power; colocation facilities, in which space and power are let to many customers, in single to low double-digit megawatts; and “hyperscale” facilities – a term coined by market analysts for the sites of the twenty or so largest cloud and internet-platform companies – at 100 MW and above, where a single site draws as much electricity as 100,000 households.Footnote 25 The artificial-intelligence build-out has added a further order of magnitude: the IEA now records individual AI data centres of 600 MW to 5 GW, campuses of 1 GW and more are planned in Portugal and France, and the “AI gigafactories” to be co-funded under the Union’s InvestAI initiative are each to house around 100,000 advanced processors.Footnote 26 The European stock is, by comparison, small and old: ENTSO-E counts more than 10,500 facilities of at least 50 kW across Europe with a combined IT capacity of about 12.7 GW, and of the facilities that reported to the Commission for 2023 and specified their size, only 55–8% – exceeded 10 MW, while 61% were below 2 MW.Footnote 27 The category that the Commission’s consultants label “very large” begins at 10 MW and cannot distinguish a 12 MW colocation hall from a training campus thirty times its size. The hyperscale segment is thus a few dozen facilities, growing fast, carrying a disproportionate share of load, and invisible as a category in every Union instrument. “Hyperscale data centre” therefore denotes here a facility or campus whose installed IT power demand is of the order of 100 MW or more – two orders of magnitude above the reporting threshold – and whose principal workload is the training of large machine-learning models or their operation at scale; the 100 MW figure marks the scale at which the behaviours described in Section III become system-relevant, not the point at which they begin, and the analytical claim – that the Union assesses this class of facility without managing it – holds for any facility above the threshold.
Nor are all large data centres alike in what they do. A cloud-computing facility runs many independent workloads on conventional servers, at a typical density of around 9 kW per rack, and its load is in aggregate smooth. An artificial-intelligence facility is built around accelerators – processors designed for the massively parallel arithmetic of machine learning – packed into liquid-cooled racks of 100 kW and more, an order of magnitude denser.Footnote 28 Within that class, training and inference behave differently. A training run is a single computation spread across tens of thousands of accelerators that must repeatedly synchronise; it alternates, many times a second, between phases in which every processor draws close to its rated power and phases in which they exchange results and draw far less, and it stops and restarts as the model’s state is saved. Inference – the use of a trained model to answer requests – is steadier and more distributed; the reliability literature is explicit that the rapid ramps of training are “not observed in current AI inference methods,” even though inference now accounts for the larger share of the lifetime energy of a model.Footnote 29 Policy debate and legal text alike tend to treat “data centres” as one thing. The hazard analysed in Section III is specific to large training facilities and to the protection architecture that AI and cloud facilities share; the distributional analysis of Section VIII applies to both.
A note on vocabulary is owed to readers who are not specialists in artificial intelligence or power engineering. “Training” is the process by which a model’s parameters are fitted to data, and a “training run” is one such computation, which for frontier models occupies tens of thousands of processors for weeks; “inference” is used in its computational sense – the operation of a trained model to produce outputs – which is narrower than the “capability to infer” by which the AI Act defines an AI system. An “accelerator” is a processor specialised for machine-learning arithmetic, the graphics processing unit (GPU) being the dominant form. “Power usage effectiveness” (PUE) is the ratio of a facility’s total energy consumption to that of its IT equipment, so that 1.0 is the theoretical ideal, and “water usage effectiveness” (WUE) the ratio of water input to IT energy. “Fault-ride-through” is the capability of equipment to remain connected through a short voltage dip caused by a fault elsewhere on the network; an “uninterruptible power supply” (UPS) is the battery-backed system through which a facility’s IT load is ordinarily fed and to which it transfers when grid supply is disturbed; and a “common-mode” failure is one in which many units fail at once from a shared cause.Footnote 30
4. An acknowledged limitation
One limitation must be acknowledged, and it is not minor. The empirical evidence on the system-stability behaviour of hyperscale loads in specifically European conditions remains thin and largely post-dates the 2024 reform package. Much of what is known is drawn from North American incident investigation, from the engineering literature on training-cluster power oscillation, and from post-incident analysis of the Iberian event – none of which was designed to answer the legal question. The article is explicit throughout about the boundary between what is established and what is inferred, and Section IX takes seriously the objection that a regulator confronting evidence of this quality is entitled to be cautious about acting on it.
III. Characterising the hazard
The regulatory categories through which European law admits large consumers to the electricity system were built for an industrial demand profile that hyperscale artificial-intelligence facilities do not exhibit. Commission Regulation (EU) 2016/1388, the Demand Connection Code, distinguishes transmission-connected demand facilities, demand units providing demand response, and closed distribution systems, and calibrates its requirements – frequency withstand, voltage withstand, fault-ride-through, reactive control – for facilities whose operational profile is recognisably that of a steel mill or a chemical plant: slow ramp rates, rotating-mass inertia, response time constants measured in seconds.Footnote 31 For the overwhelming majority of large European consumers this taxonomy was, until very recently, empirically accurate. Three features distinguish the new class.
1. Ramp behaviour and the synchronisation of load
A large training run synchronises tens of thousands of accelerators across many racks, alternating between computation-heavy phases in which every processor draws close to its rated power and communication phases in which the processors exchange results and draw far less; because the phases are synchronous by design, the campus load moves between the two states in unison and within fractions of a second. The International Energy Agency’s Energy and AI report describes power swings an order of magnitude faster than conventional industrial loads at comparable capacity, and the engineers of Microsoft, OpenAI and NVIDIA, reporting on their own training clusters, place the resulting oscillations at tens to hundreds of megawatts within a single data centre, concentrated in a band of roughly 0.2 to 3 Hz that lies close to the resonant modes of turbine-generator shafts and long transmission lines.Footnote 32 Google’s engineers report fluctuations of the same order recurring every few seconds for weeks;Footnote 33 NERC has measured a ramp of 1.9 per unit per second at one facility and a fall from about 450 MW to about 40 MW within thirty-six seconds at another;Footnote 34 and ENTSO-E’s 2026 assessment describes fluctuations of 30 to 60% of load within milliseconds, warning that persistent periodic patterns of this kind can excite the inter-area oscillation modes of the continental system and have already produced sub-synchronous oscillations in practice.Footnote 35 Two qualifications are necessary: the phenomenon belongs to training rather than to inference, which is steadier; and it is invisible at the timescale on which adequacy is assessed, where the daily profile of a data centre is flat.Footnote 36 The hazard lies in the sub-second dynamics, not in the daily curve. What makes it a system-level rather than a facility-level phenomenon is synchronisation: the oscillation is not stochastic noise averaging to zero across a portfolio of consumers but a coherent, phase-aligned swing, and coherent swings at several campuses in the same electrical region compound rather than cancel.
2. Protection logic and the economics of disconnection
A hyperscale campus houses computing hardware whose replacement value runs to billions of euros – global investment in data centres reached half a trillion dollars in 2024Footnote 37 – and its protection and control systems are engineered to shield that hardware from voltage and frequency excursions that conventional industrial protection tolerates. Two designs are in use. In the conventional design the entire IT load is fed through uninterruptible power supplies which, on detecting a voltage dip, transfer the load instantaneously to batteries and thereby remove it from the grid; ENTSO-E observes that such systems remain dominant even in new facilities and may transfer “at the first sign of a voltage dip.” In the newer design used for training clusters, UPS protection of the IT equipment is frequently omitted and the computation is protected instead by periodic checkpointing, so that a disturbance produces an outright loss of load rather than a transfer.Footnote 38 From the operator’s standpoint the logic is impeccable: the cost of damaged silicon, or of a training run lost, vastly exceeds the cost of a brief interruption. From the system operator’s standpoint, a disturbance that ought to be locally containable instead produces the simultaneous disconnection of a large block of load – and, where such facilities are clustered, the disconnection is common-mode.
This is no longer conjecture. On 10 July 2024 a fault on a 230 kV line in the Eastern Interconnection of the United States, cleared normally by the protection system, was followed within seconds by the loss of approximately 1,500 MW of load that NERC’s incident review found to be “exclusively data center-type load,” disconnected “on the customer side by customer protection and controls” as the facilities transferred to their backup systems; some 1,260 MW did not return for hours.Footnote 39 A second event of similar size occurred in the same region in February 2025, and on 22 July 2026 a third removed approximately 3,800 MW – the largest such event in the history of the PJM system – prompting the chair of PJM’s operating committee to observe that “this was a normally cleared fault. They should not disconnect from the grid.”Footnote 40 Texas has recorded twenty-six such events since 2023 and has adopted the first dedicated ride-through requirements for large electronic loads;Footnote 41 NERC, having found that roughly three-quarters of the load models supplied for data centres are insufficient to represent their dynamic behaviour, issued in May 2026 a Level 3 alert – its highest tier – on the modelling, study and commissioning of computational loads, and its guidance records both that ride-through standards exist for generators and none for large loads and that a voltage dip at a single facility can trigger “a synchronized demand reduction across all participating data centers” that share a workload.Footnote 42 The European record is thinner but not empty. The Irish transmission system has experienced a 204 MW drop in data-centre demand following a single fault, documented by the system operator’s own engineers;Footnote 43 ENTSO-E reports that “multiple incidents across different grid systems have already demonstrated load losses ranging from several hundred megawatts to over one gigawatt following routine transmission faults,” identifies the further risk of “flapping” as UPS systems return load to the grid within seconds and depress frequency back towards the disconnection threshold,Footnote 44 and describes a “tendency to switch to island operation mode during minor disturbances” that “may potentially remove several GW of load.”Footnote 45 The engineering literature locates the cause in a mismatch between equipment standards, which define what IT equipment will tolerate, and grid codes, which define what the system needs from connected load.Footnote 46 This is the class of behaviour the ENTSO-E expert panel identified as central to the cascade of 28 April 2025, albeit in relation to generation rather than load; its recommendations on the transparency of protection settings and the accuracy of facility models translate to the demand side without strain.Footnote 47
3. Clustering
Facilities co-locate around fibre backbone nodes, availability zones and tax-favourable jurisdictions. The five “FLAP-D” markets – Frankfurt, London, Amsterdam, Paris and Dublin – are the principal European clusters and face connection lead times ranging from a few years to more than a decade; ENTSO-E estimates that individual campuses now reach several hundred megawatts and that clusters in concentrated regions exceed one gigawatt, “comparable to the dimensions of European frequency reserves,” and the IEA expects more than half of Europe’s data-centre capacity to sit in such clusters by the end of the decade.Footnote 48 In Ireland, where data centres consumed 22% of national metered electricity in 2024 and 23% in 2025, and roughly half of regional demand is concentrated around Dublin, a common-mode disturbance affecting that cluster is not analogous to the loss of a single industrial customer.Footnote 49 It is closer to the loss of a large generator, with the inversion that the loss creates a surplus rather than a deficit and calls for rapid curtailment rather than rapid dispatch.
4. The legal consequence: differentiation as requirement, not merely permission
What follows in law from a difference in electrical behaviour? The conventional reading of the non-discrimination principle in Article 6 of Directive 2019/944 is that system operators must treat comparable users alike.Footnote 50 It is not usually read as requiring that materially different users be treated differently. But the converse formulation is the operative one: identical regulatory treatment of materially different categories of user is itself a discrimination problem where that treatment imposes disproportionate systemic costs on third parties who are not party to the connection. The Court’s reasoning on differentiation by objective characteristic translates without undue strain into electricity.Footnote 51 Where the technical and operational characteristics of one demand class differ materially from another, uniform treatment is at best permitted and, where third-party costs are engaged, questionable.
This is not to say that the Demand Connection Code is unfit for purpose. It is to say that the categorical question – what kind of facility, for regulatory purposes, is a hyperscale artificial-intelligence campus – can no longer be deferred. ENTSO-E itself has now acknowledged that data centres, like storage, electric vehicles and heat pumps, fall outside the specific requirements of the connection network codes in force; that the revision of those codes recommended by ACER in December 2023 “has been deprioritised, and no specific timeline has been communicated”; and that transmission system operators should therefore introduce national connection requirements for data centres – fault-ride-through profiles, rate-of-change-of-frequency withstand, ramp-rate limits – in the interim, at the price of the very fragmentation that Union network codes exist to prevent.Footnote 52 ACER’s Recommendation 01/2025 extends demand-response participation to electric vehicles, heat pumps and power-to-gas installations, and is a genuine advance.Footnote 53 It does not address the prior question of how the largest, most clustered and most rapidly varying class of demand on the European system is to be classified for connection and ancillary-service purposes. Risk identification is substantially complete. Categorisation – the first act of risk management – has not begun.
IV. The assessment apparatus and its first yield
1. The architecture
Article 12 of the recast Energy Efficiency Directive, read with Delegated Regulation (EU) 2024/1364, is the Union’s most ambitious intervention to date on data-centre performance, and it is an assessment instrument. Operators of facilities whose installed information-technology power demand exceeds 500 kW must report annually – by 15 May, following the first cycle in September 2024 – the information specified in Annex VII: energy consumption, power usage effectiveness (PUE), water usage effectiveness, waste-heat utilisation and renewable-energy share.Footnote 54 The delegated regulation defines four headline indicators and a longer schedule of reportable indicators across three blocks, and treats facility-level data as commercially confidential within the meaning of Regulation (EC) No 1049/2001 and Directive 2003/4/EC, releasing only the aggregate.Footnote 55 Article 26(6) adds a waste-heat obligation for facilities above 1 MW total rated energy input, qualified by a technical-and-economic-feasibility carve-out subject to cost–benefit analysis under Annex X.Footnote 56
The informational reach of this architecture is impressive. Its regulatory consequence is another matter. Article 12 imposes no minimum performance standard: a facility may report a PUE far above the European average with no direct legal consequence whatever, and the Code of Conduct on Data Centre Energy Efficiency, to which Article 12(4) refers for facilities above 1 MW, is voluntary. The Directive itself contemplated the next step: Article 12(5) required the Commission to assess the reported data by 15 May 2025 and to accompany its report, “where appropriate,” with legislative proposals “including establishing minimum performance standards”; the reports were delivered, in July and October 2025, and no proposal accompanied them.Footnote 57 The legal literature is unanimous that reporting alone will not suffice: Commins and Irion call for the reporting requirements to be translated into binding targets, and Ebert and his co-authors propose binding PUE and renewable-energy targets at Union level, observing also that the delegated regulation measures IT energy at the uninterruptible power supply rather than at the server, so that the ratio it produces is blind to the efficiency of the computation itself.Footnote 58 Whether the present arrangement amounts to inadequate regulation depends, in the end, on what one expects a reporting regime to do.
2. What the first assessment disclosed
On that question the first empirical cycle is now informative, and it is the principal reason for revisiting the problem in 2026. The Commission’s first technical report, published in July 2025 on the basis of the 2023 reporting cycle, covered 770 data centres, which it estimated to be some 36% of the 2,161 facilities in the Union above the reporting threshold; six Member States had no reported data at all. The reporting population consumed about 14.1 TWh and showed a weighted average PUE of 1.36.Footnote 59 The aggregate database that the Commission opened in April 2026, covering the 2023 and 2024 cycles, lists 776 facilities in 18 Member States with an installed IT power demand of 6.4 GW and consumption of 16.7 TWh.Footnote 60 Three features of that result deserve attention.
The first is the participation rate. The Commission’s own estimate of 36% is corroborated by independent analysis of the reported population against compiled facility inventories, which puts participation in Ireland – the jurisdiction under the greatest system stress in the Union – at approximately one in six.Footnote 61 A reporting obligation with a one-in-three response rate is not an assessment instrument calibrated conservatively; it is an assessment instrument that has substantially failed at the collection stage, and the aggregates it produces are correspondingly unsafe as a basis for management decisions. The cause is structural and familiar: the obligation binds the operator, the enforcement competence rests with the Member State, and the Union institution that needs the data has no direct relationship with either.Footnote 62
The second is the composition of the reporting population, which bears directly on the definitional problem discussed in Section II. Only 8% of the reporting facilities had an installed IT power demand above 10 MW and 61% were below 2 MW; colocation facilities made up 54% of the population, enterprise facilities 29% and co-hosting facilities 16%.Footnote 63 The aggregate PUE of 1.36 is thus dominated in number by small and medium facilities and in load by a few dozen large ones – which performed markedly better, at 1.21, than the smallest band at 1.64 – so that the indicator most likely to be read as a verdict on the sector says least about the segment that matters most for the system, and nothing about how it behaves. The Commission’s October 2025 assessment of next steps proposes to extend reporting downward to 100 kW and to add indicators for accelerator performance and liquid cooling: a deepening of assessment, not a turn to management.Footnote 64
The third is what the aggregate conceals. The confidentiality regime in Article 5 of the delegated regulation means that facility-level performance is not disclosed. A market-facing disclosure regime works, where it works, because disclosed information is translated into competitive pressure by counterparties able and motivated to act on it; the Union’s experience with sustainability-related disclosure in financial services shows that the translation can occur.Footnote 65 The hyperscale cloud market does not present that structure. End-customers of cloud services are largely indifferent to facility-level energy performance and, by reason of the confidentiality regime, could not act on it if they were not. Commins and Irion make the same point from the side of the researcher, for whom aggregate publication permits no comparison across sites, and Ebert and his co-authors add that facility-level metrics cannot in any event surface the variation between workloads that determines the energy cost of a given model.Footnote 66 Article 12 therefore generates information for which there is at present no market counterparty and no regulatory consequence – a condition fairly described as reporting without consequence.
3. The transposition record and the German exception
The picture is completed by the transposition record. In November 2025 the Commission sent letters of formal notice to twenty-six Member States for failure to transpose the recast Directive by the deadline of 11 October 2025, and by June 2026 had moved to reasoned opinions against two of them.Footnote 67 One transposition went considerably further than the Directive required. Section 11 of the German Energieeffizienzgesetz imposes binding thresholds: data centres commencing operation on or after 1 July 2026 must achieve a PUE no higher than 1.2, existing facilities must reach 1.5 by July 2027 and 1.3 by July 2030, waste heat must be reused at a rate rising from 10% for facilities commissioned from July 2026 to 20% for those commissioned from July 2028, and electricity must be sourced half from renewables from 2024 and wholly from renewables from January 2027, with fines for non-compliance.Footnote 68 Germany has, in short, supplied the management stage that Union law omitted.
That fact cuts three ways. It demonstrates that binding minimum standards for this facility class are legislatively feasible and administratively workable: the objection that the sector is too heterogeneous or too fast-moving to bear performance standards is answered by the fact that a Member State has imposed them. It also demonstrates the cost of Union abstention, since a binding standard of 1.2 in one Member State and no standard at all in another is a difference in operating cost that will express itself in siting decisions. And it demonstrates, in the sequel, the fragility of a national standard exposed to that difference. After sustained criticism from the industry associations – which argued that the thresholds were unattainable without excessive water use, that the waste-heat quotas presupposed heat networks that do not exist, and that the law would drive investment abroadFootnote 69 – the federal government adopted in June 2026 a draft amendment raising the PUE ceilings for existing facilities to 1.6 and 1.4, giving new facilities four years rather than two to reach 1.2, deferring the renewable-electricity obligation to 2030 and adding waste-heat exemptions.Footnote 70 The only Member State to have converted assessment into management was, at the time of writing, relaxing every numerical value it had adopted. That is the “waterbed” dynamic the literature predicts for unilateral standards in a mobile industry,Footnote 71 and the strongest argument for setting the standard at the level at which the industry cannot arbitrage it. Whether the asymmetry constitutes an internal-market problem in the technical sense cannot be pursued here. That it constitutes a risk-management asymmetry, which the second-phase delegated act under Article 33(3) will either close or entrench, is clear enough.
That instrument is now in view: the Commission consulted on a draft rating-scheme regulation between 27 March and 23 April 2026, with adoption anticipated in mid-2026. The draft would generate, from August 2027 and annually thereafter, an electronic label classing each reporting facility by its power and water usage effectiveness; it attaches no consequence to the class.Footnote 72 A rating scheme is an assessment instrument with management potential; it becomes a management instrument only if something is attached to the rating. The next section concerns the most consequential thing that could have been attached to it, and was not.
V. Acceleration without conditionality: the Cloud and AI Development Act
On 3 June 2026 the Commission proposed the Cloud and AI Development Act, the centrepiece of its Tech Sovereignty Package and the legislative expression of the AI Continent Action Plan of April 2025.Footnote 73 Its animating objective is the multiplication of European compute: an at-least-tripling of Union data-centre capacity within five to seven years and the closing of the “capacity gap” by 2035, pursued through the mandatory designation by each Member State of at least one data-centre “acceleration zone,” the automatic treatment of projects in such zones as strategic projects for permitting purposes, a twelve-month ceiling on permit-granting, the Commission’s designation of “strategic projects” eligible for preferential access to Union funding, and a cloud-sovereignty framework built on graduated assurance levels conditioning public-sector procurement.Footnote 74 The underlying policy rationale – the erosion of the European cloud position, with European providers’ share of their own market having fallen from roughly 29% in 2017 to approximately 15%, where it has since remained – is neither contrived nor disreputable, and the competence basis in Article 114 and Article 173(3) TFEU is secure.Footnote 75
The difficulty is not the objective. It is the relationship between this instrument and the energy-efficiency-first principle in Article 3 of the recast Energy Efficiency Directive, which is not a slogan but a legal obligation binding on the Union institutions in policy, planning and major investment decisions.Footnote 76 A Union industrial policy that promotes the multiplication of very large electricity consumers, and confers real legal and financial advantages on those that participate, without conditioning those advantages on demonstrated energy performance, stands in evident tension with that principle.
The tension is sharpened by comparison with the Union’s own recent practice. The Net-Zero Industry Act conditions strategic-project status on contribution to specified industrial and climate objectives, and lists the technologies eligible;Footnote 77 data centres are not among them. The Critical Raw Materials Act conditions strategic-project status on environmental and social due diligence.Footnote 78 The 2022 Climate, Energy and Environmental State Aid Guidelines confine indirect emission-trading-cost compensation to a closed list of energy-intensive sectors that does not include data processing.Footnote 79 Each of these instruments embeds performance conditionality into the architecture of industrial advantage. The Cloud and AI Development Act does something weaker, and it is worth being precise about what. Article 10(1) lists eight “aspects” that Member States “shall consider” when designating acceleration zones – among them grid capacity, the scope for on-site storage and clean generation, waste-heat reuse, a preference for brownfield sites and “the ability of the site or area to function sustainably” – and Article 10(2) asks them, “where appropriate,” to analyse each zone’s energy needs and feed the analysis into network development plans.Footnote 80 These are planning considerations addressed to the Member State at designation, not obligations of the operator that follows. Article 11(1), the only provision that speaks of “sustainability requirements” for the facilities themselves, is conditional in form: “when setting” such requirements, Member States “shall use” the key performance indicators of the first-phase delegated regulation – and it cross-refers to points (a) to (n) of Annex II, the measured inputs, leaving out the renewable-energy indicators in points (o) to (r) and the derived sustainability indicators of Annex III, PUE among them.Footnote 81 Article 13 then confers the benefits of the zone – strategic-project status for permitting purposes, an aggregated baseline permit and the twelve-month ceiling – on every data-centre project in the zone, without condition.Footnote 82 Article 14 allows the Commission to designate strategic projects that satisfy any two of five criteria, only two of which – “highly sustainable or innovative features” and a contribution to the stability of the electricity grid – concern performance; a project may therefore be designated, and its designation withdrawn only for ceasing to meet the criteria on which it was granted, without any sustainability commitment at all.Footnote 83 The proposal, in short, harmonises the vocabulary in which Member States may describe performance and leaves each of them free to demand none.
It might be objected that a “sustainability criterion” can be every bit as binding as a “minimum performance standard,” and that the distinction drawn here is one of nomenclature. The objection is fair, and the reply is that the deficiency lies not in the label but in three structural features of the proposal. The first is the addressee: every sustainability provision in Title III is addressed to the Member State, at the level of the zone, and none creates an obligation that a facility must satisfy or a benefit that a facility must earn, so that the operator – the only entity whose behaviour the system operator needs to change – is the subject of none of them. The second is the relationship between criterion and advantage: in the Net-Zero Industry Act the advantage is the consequence of meeting the criterion, whereas here the advantage is the consequence of location and the criterion, where a Member State chooses to set one, runs alongside it. A criterion that conditions nothing is an aspiration, whatever its grammatical mood. The third is fragmentation: Article 11 contemplates twenty-seven sets of “sustainability requirements,” or none, expressed in a common metric but at no common level – the arrangement the Directive already produces, whose results Section VII documents. An instrument that draws a mobile industry into designated zones across the Union while leaving the performance conditions of those zones to national choice does not merely fail to close the assessment–management gap; it institutionalises the regulatory competition that the gap produces.
None of this is to say that a binding minimum standard is a sufficient instrument, and the proposal made here is not for one alone. The economics of instrument choice has long established that a uniform standard, once met, gives the regulated firm no reason to go further, and that the incentive to innovate beyond the threshold is weaker under such standards than under price-based or graduated instruments;Footnote 84 that the form of a standard matters as much as its level, since a ratio such as PUE can be improved at the margin by making its denominator – the IT load – less efficient;Footnote 85 and that standards fixed by legislation lag the technology by the length of the legislative cycle, which in the Union’s Ecodesign framework and Japan’s Top Runner programme alike has averaged five to six years.Footnote 86 The literature’s answer is not to abandon the floor but to pair it with a gradient: minimum standards combined with graduated labels, as in Union appliance regulation, or a benchmark that moves with the best performers, as in Top Runner, whose combination with labelling has been found to raise producers’ research expenditure.Footnote 87 Instrument mixes in which the floor admits and the ladder rewards are the settled recommendation of the regulatory-design literature, which has also shown that firms cluster above a legal floor only where something – a market, a licence, a reputation – gives them reason to.Footnote 88 The second-phase rating scheme is precisely such a gradient, and the Cloud and AI Development Act is precisely the instrument to which it could be attached. A minimum class could condition admission to the acceleration zone; higher classes could condition strategic-project designation and the funding that follows from it; and the class boundaries could be indexed, on the Top Runner model, to the performance of the best-rated facilities in each reporting cycle, so that the benchmark rises with the technology rather than waiting for the legislator. The indicators already reported under the delegated regulation include the functions a facility provides to the grid and the battery capacity it offers to it,Footnote 89 so that a rating could reward the system-friendly behaviour that Section III shows to be lacking and not only energy efficiency – a connection of which the drafters were not unaware, since Article 14(1)(c) counts a contribution to grid stability among the grounds for strategic-project status. This is the conditionality design that the industrial-policy literature recommends – public advantage in exchange for verifiable commitments, with the advantage graduated to the commitmentFootnote 90 – and it answers the objection that a floor blunts the incentive to exceed it: the floor admits, the ladder rewards, and neither is left to twenty-seven national choices.
The omission is the more striking because the instrument that would supply the missing benchmark is being drafted in the same institution in the same year, and because the Directive itself, in Article 12(5), had already named minimum performance standards as the sequel to the first assessment. A conditionality linkage between acceleration-zone designation and a threshold rating under the second-phase delegated act would cost the Union nothing in legislative innovation: it would replicate, in the digital-infrastructure field, a technique the Union has already adopted twice in the industrial field. Its absence is not an oversight of drafting. It is the choice to keep assessment and management apart, made a third time, and made at the point where the Union’s leverage over the facility class is at its maximum – because a hyperscale operator seeking accelerated permitting and preferential funding is an operator with something to lose. The legislative window remains open: at the time of writing the proposal was before the Parliament’s committees and the Council had yet to adopt a position.Footnote 91
VI. Connection, flexibility and the treatment of large loads in the 2024 market design
The electricity market design reform of June 2024 is the most substantial recasting of Union electricity law since the Third Energy Package, and its preoccupations are recognisably those of 2022: insulating consumers from wholesale volatility, stabilising renewable revenue, improving flexibility. It inserts new Articles 7a (peak-shaving products), 9 (non-fossil flexibility support) and 19a–19e (contracts for difference and power purchase agreements) into Regulation (EU) 2019/943, and new Articles 6a (flexible connection agreements), 15a (energy sharing), 28a (vulnerable customers) and 31 (locational tariffs) into Directive 2019/944.Footnote 92
None of these provisions singles out hyperscale facilities, and most are indifferent to the size or character of the demand-side participant. That neutrality repays examination. The peak-shaving products under Article 7a are crisis instruments, triggered by a declaration under Article 66a, rather than structural mechanisms for the routine participation of large loads. The non-fossil flexibility support under Article 9 is technology-agnostic, but its principal beneficiaries, given the cost dynamics of European demand response, are aggregated residential and small commercial loads, not facilities whose operators have overwhelming commercial incentives to maximise utilisation. The contract-for-difference provisions operate on the supply side.
The provision that comes closest is Article 6a of Directive 2019/944, which permits Member States to authorise flexible, non-firm connection agreements in congested areas. Doctrinally this is the right tool: a flexible connection admits a new load at a constrained node on terms that acknowledge the constraint, rather than deferring admission until reinforcement is complete. But the authorisation is permissive, not mandatory; the recitals are careful to preserve reinforcement as the structural solution; and the provision supplies no criteria for allocating flexible-access capacity where applicants compete for it. The allocative work – which is to say the risk-management work – is left entirely to the Member States.
The Commission’s European Grids Package of December 2025 confirms the pattern. Its guidance on grid connections names data centres among the large connection requests for which grid reinforcement “is usually necessary,” recommends “first-ready, first-served” queues and “grid acceleration areas” for, among others, “the data industry,” and concedes that there is “as yet no example of an explicit and direct prioritisation framework based on economic criteria” or on contribution to policy objectives such as the deployment of data centres.Footnote 93 It is guidance, addressed to Member States; the allocation decisions it describes remain, as under Article 6a, national decisions within a Union framework that neither requires nor harmonises them.Footnote 94
The Union is not unaware of the consequence. ACER’s 2025 report on network-tariff methodologies identifies the proliferation of ad hoc national tariff regimes for particular user groups as a threat to the integrity of price signals, and warns in terms against exemptions and bespoke tariffs that inhibit cost causation and shift costs to other users.Footnote 95 The tariff principles it articulates – cost-reflectivity, non-discrimination, avoidance of cost-shifting – are a useful piece of soft law. They are not operationalised in binding form, and until they are, the cost-allocation dimension of hyperscale integration remains the least governed part of the architecture.
VII. What Member States do when the Union does not: four responses
The predictable consequence of a Union framework that assesses without managing is that the management is performed elsewhere. Four national responses illustrate both the ingenuity and the incoherence that results.
1. Ireland: conditionality by regulatory decision
Ireland is the jurisdiction in which the pressure is most acute. Data centres accounted for 22% of national metered electricity consumption in 2024 and 23% in 2025, up from 5% in 2015, and the system operator’s adequacy assessment projects a rise towards 31% by 2034.Footnote 96 Roughly half of the regional demand sits around Dublin.
On 12 December 2025 the Commission for Regulation of Utilities published its decision on a new electricity connection policy for large energy users, superseding the direction of November 2021 under which new connections in the Dublin region had in practice been suspended.Footnote 97 The decision is, in the vocabulary of this article, an unambiguous risk-management instrument. Applicants above 10 MVA must bring dispatchable on-site or proximate generation or storage, de-rated and at least equal to their maximum import capacity, separately connected and metered and participating in the Single Electricity Market; must match at least 80% of annual demand with additional Irish renewable generation, on a six-year glide path; and are subject to locational and constraint screening. Applicants between 1 and 10 MVA must provide on-site generation sized to their full import capacity; those below 1 MVA are exempt. System operators were required to publish their engagement and connection processes, and locational capacity information, by 31 March 2026.
What is doctrinally notable is the public-law foundation on which this rests. The statutory direction power under section 34 of the Electricity Regulation Act 1999 supplies the vires; the climate-consistency duty under section 15 of the Climate Action and Low Carbon Development Act 2015, as substituted in 2021, supplies the obligation to exercise it in a particular direction.Footnote 98 The scope of that duty has been worked out in litigation: the High Court’s characterisation of it as an “imperative” obligation was narrowed on appeal, in February 2026, to a “climate sense check” – mandatory and judicially enforceable, but tolerant of a range of outcomes – without disturbing the foundation.Footnote 99 Ireland demonstrates that the materials for a robust national risk-management response can be assembled from existing statutory powers without bespoke primary legislation. The cost is regulatory complexity and litigation exposure; the benefit is a response capable of being defended.
2. Germany: conditionality by statute
The German position has been described above, including the retreat from it now in progress. Its significance here is comparative: where Ireland conditions admission to the system, Germany conditions operation within it. Both are management instruments; neither is required by Union law; and an operator choosing between the two jurisdictions faces materially different obligations arising from the same Directive – obligations which, in the German case, the legislature is now revising downward under precisely the competitive pressure that Union-level standards would remove.
3. The Netherlands: prioritisation and its limits
The Dutch case poses the hardest question. Faced with transmission-level connection requests of some 38 GW and waiting lists at every voltage level, the national regulator has developed a prioritisation framework for scarce transport capacity, giving precedence to congestion-relieving connections, public-safety functions and basic needs.Footnote 100 Its first version was annulled by the College van Beroep voor het bedrijfsleven in March 2025 for want of adequate substantiation, on appeals brought by, among others, the incumbent telecommunications operators and the national data-centre association; the revised framework adopted in December 2025, applicable from 1 January 2026, admits telecommunications to the “basic needs” category and data centres to none. The litigation raises squarely the question Union law has declined to answer: whether, and on what basis, a connection queue may lawfully rank users by the social value of their consumption rather than by the chronology of their applications. That question cannot be resolved by the non-discrimination principle alone, because the principle presupposes a prior determination of which users are comparable – and it is precisely that determination that Union law has not made. The Dutch dispute is the assessment–management gap surfacing as litigation, in a Member State that had no choice but to decide.
4. Spain: transparency by competition
Spain is simultaneously the destination of choice for southern-European hyperscale investment and a system under acute transmission constraint. Its response operates through demand-side capacity competitions at congested nodes of 220 kV and above, and through the distinction introduced by CNMC Circular 1/2024 between firm access – carrying conventional supply guarantees but available only at unsaturated nodes – and flexible access, available at saturated nodes conditional on the customer’s acceptance of operational restriction under defined network conditions.Footnote 101 The distinction is designed with hyperscale facilities in mind and operationalises Article 6a in the most doctrinally transparent form yet achieved: criteria published in advance, saturation maps public, decisions judicially reviewable. It is the most directly transposable of the four models to Union level – but transposition would require the legislator and ACER to perform, for the whole internal market, the allocative work a single national regulator has so far performed alone.
5. The common structure
The four regimes differ in instrument and in institutional confidence. They share three commitments. None treats a hyperscale facility as merely another industrial customer. Each conditions admission to, or operation within, the system on some commitment beyond what was demanded of the previous generation of large loads. And each rests on a national legal foundation that the Union acquis permits but does not require. It is the third that founds the argument for consolidation – and that argument must now confront the dimension in which the costs of non-consolidation fall on those who are not party to any of it.
VIII. The distributional dimension: risk borne by those who did not create it
Risk regulation is not only about the magnitude of a hazard. It is about its incidence. The hazards described in Section III are borne, in the first instance, by the electricity system; but the costs of managing them are borne by whoever the tariff methodology says shall bear them, and in most Member States that means all network users in proportion to consumption.
Two mechanisms transmit hyperscale growth into household bills. The first is the wholesale price effect: additional industrial-scale demand entering a supply-constrained short-run market produces non-trivial price effects on infra-marginal consumers, and the effect is amplified in scenarios where the incremental demand is met disproportionately by gas-fired generation.Footnote 102 The second, and legally the more tractable, is the network-cost effect. Under the socialisation model embedded in most national tariff regimes, the cost of reinforcing transmission and distribution to accommodate a new large load is recovered from all users in proportion to consumption. The result is a regressive cross-subsidy: households whose marginal contribution to reinforcement need is negligible bear a share of costs occasioned principally by industrial-scale new connections.
Union law is not without resources here. Article 18 of Regulation (EU) 2019/943 requires network tariffs to be cost-reflective and non-discriminatory, and ACER’s tariff principles operationalise that requirement through cost causation, locational signalling and the avoidance of cost-shifting.Footnote 103 Properly applied, the doctrinal implication is unambiguous: the substantial reinforcement costs occasioned by new hyperscale loads should fall principally on those loads – through deeper connection charges, locational signals and time-differentiated tariffs – rather than being socialised across the customer base. The materials for that allocation exist. They have not been applied with anything approaching the consistency ACER’s own report demands.
Nor does the Union’s principal redistributive instrument bridge the gap. The Social Climate Fund, substantial in absolute terms, is designed to address the social incidence of the ETS2 carbon-pricing instrument applicable to buildings and road transport; industrial-scale electricity demand growth driven by data-centre deployment is not within its remit.Footnote 104 Article 28a of Directive 2019/944, inserted in 2024, strengthens protection against disconnection for vulnerable customers, and Article 2(52) of the recast Energy Efficiency Directive supplies the Union’s first binding definition of energy poverty.Footnote 105 These are protective instruments, and they are welcome. They are not allocative instruments, and the problem is allocative.
There is, moreover, a second distributional axis that the legal literature on this subject has largely neglected, although the environmental-science literature has not: water. The European Water Resilience Strategy, adopted on 4 June 2025, identifies data centres among the consumption concerns it addresses, targets a water-efficiency improvement of at least 10% by 2030 and announces minimum water-performance standards for the sector.Footnote 106 Cooling water is a rival good in a way that electricity, at the margin, is not: a hyperscale facility drawing evaporative cooling water in a water-stressed basin competes directly with agriculture and with municipal supply, and the competition is invisible to the electricity acquis entirely. Water usage effectiveness is a reportable indicator under Delegated Regulation (EU) 2024/1364, and fewer than two-thirds of the facilities that reported for 2023 supplied it – which is to say that the Union assesses it, incompletely, and manages it not at all.
IX. What precaution can and cannot be made to bear
An objection must now be met. If the evidence on the system behaviour of hyperscale loads in European conditions is as thin as Section II concedes, is the Union not entitled – perhaps obliged – to withhold management measures until the evidence improves? Is the assessment-first posture not simply good regulatory practice?
The objection has force and deserves better than dismissal. But it misconceives the structure of the precautionary principle and misidentifies the uncertainty at issue. The principle, as developed in Union law, addresses scientific uncertainty about whether a hazard exists or how severe it is.Footnote 107 It licenses protective action in advance of conclusive proof; it does not license inaction pending proof, and the Court has insisted that a precautionary measure be founded on a risk assessment as complete as the circumstances permit rather than on a purely hypothetical approach to risk.Footnote 108
Applied here, the principle cuts in an unexpected direction, for three reasons. First, the uncertainty is not about whether the hazard exists. Fast-ramping, protectively disconnecting, geographically clustered multi-hundred-megawatt loads are not a hypothesis; they are a description of facilities now operating. What is uncertain is their aggregate behaviour under stress in a given European electrical region – an uncertainty that is not scientific in the relevant sense but informational, and one the Union has itself perpetuated by building a reporting regime that captures roughly a third of the population and publishes only the aggregate.
Second, precaution has an asymmetry that the assessment-first posture ignores. The cost of an unnecessary management measure – a load-class categorisation that proves over-cautious, a connection condition that proves redundant – is a marginal increase in the cost of capital for a facility class with exceptionally high margins. The cost of an omitted measure is a cascading disconnection event in a densely clustered electrical region. These are not commensurable, and the incommensurability is precisely what precautionary reasoning exists to handle: where the potential harm is severe and the mitigation cheap, the burden of showing that mitigation is unnecessary properly falls on the party resisting it.
Third, and most fundamentally, Union abstention is not a state of nature. It is a decision with distributive consequences, and Sections VII and VIII have identified them: fragmentation of the internal market, socialisation of reinforcement costs, unmanaged competition for water in stressed basins, and the improvisation of national instruments whose lawfulness is now being litigated. A regulator that declines to manage a risk has not avoided taking a position on it. It has allocated the risk to those least able to price it.
What precaution cannot bear is the weight of a general licence. It does not entitle the Union to impose whatever conditions on data centres a Member State might wish, nor to abandon proportionality, nor to dispense with the evidence base Article 12 was designed to build. The argument here is emphatically not that Union law should be more restrictive of data centres as such. It is that Union law should decide – on the evidence it has, imperfect as it is, and with the review mechanisms that imperfect evidence demands – rather than continue to gather evidence it does not act upon.
X. Conclusion: closing the gap
The post-2024 acquis was framed in response to the wholesale-price crisis of 2022, not to the artificial-intelligence build-out that followed it. That is not a criticism; no legislature anticipates everything. The criticism is that two years on, with the first assessment cycle complete and its results in hand, the Union continues to legislate on the assessment side of the line while the management side is improvised by Member States, contested before national courts and – in the case of the Cloud and AI Development Act – quietly foreclosed by an industrial policy that confers advantage without asking for performance in return.
Three corrections follow, and none is novel. The first is categorical: a revision of the Demand Connection Code, building on ACER Recommendation 01/2025 and on the connection requirements that ENTSO-E has now asked its members to adopt nationally for want of a Union instrument, establishing a load class for very large flexible demand, with connection and ancillary-service requirements – fault-ride-through, rate-of-change-of-frequency withstand, ramp-rate limits – calibrated to its actual electrical behaviour rather than to that of the industrial facilities the Code was written for. North America, after three common-mode disconnections of 1.5 to 3.8 GW in two years, will have reliability standards for large loads filed by the end of 2026; the Union, told by its own system operators that the equivalent revision has been deprioritised, has no timetable at all. The second is conditional: a graduated linkage between the advantages conferred by the Cloud and AI Development Act – acceleration-zone designation, strategic-project status, preferential funding – and the classes of the second-phase rating scheme, on the model already established in the Net-Zero Industry Act and the Critical Raw Materials Act and with the floor-and-ladder structure described in Section V, so that admission is conditioned on a minimum class and advantage on a higher one. The third is allocative: the operationalisation, in binding form, of the cost-causation principle that Article 18 of Regulation (EU) 2019/943 already states and that ACER’s tariff principles already elaborate, so that the reinforcement costs occasioned by hyperscale connection fall on the connecting load rather than on the household bill.
Each draws on materials that already exist, and the second on a mandate that the Directive itself supplies, since Article 12(5) named minimum performance standards as the sequel to the first assessment. What they require is not new competence but the decision to cross from assessment to management – a decision the Union has deferred through two legislative cycles, and which the second-phase delegated act, the Cloud and AI Development Act trilogue and the Network Code on Demand Response each present an occasion to take.
The standard against which Union law should be judged is not whether it has produced a comprehensive answer to the artificial-intelligence data-centre question; that question arrived too quickly, and from a direction the 2022–2024 reform process could not have anticipated. The standard is whether the acquis contains the materials from which an answer can be built, and whether the institutions are willing to build it. The materials are there. The evidence, for all its imperfection, is better than it was when the reform package was adopted, and it will not improve by being left uncollected. Whether the Union crosses the assessment–management line within the present reform cycle is, in the end, a political question rather than a doctrinal one. But it should be recognised as a decision, and not mistaken for the absence of one.
Acknowledgments
The author is grateful to colleagues at the Faculty of Law of Károli Gáspár University of the Reformed Church in Hungary for discussion of an earlier draft, to participants at the 2025 Budapest Energy Law Workshop for comments on the comparative material, and to the journal’s anonymous reviewer, whose comments on the definition of the facility class, the technical vocabulary, the evidentiary basis of Section III, the treatment of the Cloud and AI Development Act and the engagement with the literature led to substantial revisions.
Financial support
This research received no specific grant from any funding agency, commercial or not-for-profit sectors.
Competing interests
The author declares none.
Use of generative artificial intelligence
The author used a large-language-model assistant solely for language polishing, structural editing suggestions and citation-format checking. The tool was not used to formulate the substantive legal arguments, to select the jurisdictions examined, to identify legal authorities or to verify legal sources. All outputs were reviewed and revised by the author, who has verified every reference against the primary source and takes full responsibility for the content and integrity of the article.
Data availability statement
This article is a work of doctrinal and comparative legal scholarship. No new empirical datasets were generated or analysed. All legal instruments, case law, regulatory materials and secondary sources relied upon are identified in the footnotes and are, unless otherwise indicated, publicly available through official sources.
Open access
