Berlin Blinks, But Frankfurt Still Waits
On April 9, 2026, Germany’s Federal Ministry for Economic Affairs and Energy published a draft bill amending the Energy Efficiency Act, the Energieeffizienzgesetz, known in the industry by its acronym EnEfG. The announcement arrived with the administrative framing of a routine legislative update. Its actual content was a systematic retreat from the most demanding requirements that Germany had placed on its data centre sector when the original EnEfG passed in September 2023, requirements that the German Data Center Association had warned, at the time of passage, constituted a data centre prevention act. Three years of industry pressure, a European Commission infringement proceeding against Germany for failing to meet the EU Energy Efficiency Directive’s October 2025 transposition deadline, and the arrival of AI hardware generations whose physical characteristics made the original requirements arithmetically impossible to meet had collectively produced a bill that walked back PUE thresholds, restructured waste heat reuse obligations, and granted exemptions that the original legislation had explicitly refused.
The industry welcomed the amendment carefully. The law firm Taylor Wessing described the raised PUE limits and significant simplifications regarding waste heat utilisation as resulting in a considerable reduction in implementation burden. Orrick’s analysis noted that the draft proposes higher target values for PUE and introduced the option of counting in-house heat utilisation toward the energy reuse target. Bird & Bird observed that the annual data submission requirement for waste heat has been abolished, replaced by a purely voluntary notification to the Federal Office for Energy Efficiency. On each of these specific points, the 2026 amendment represented a genuine easing relative to what the 2023 legislation had required.
What none of these careful legal analyses could obscure is the nature of the easing. Germany had legislated requirements that were genuinely impossible to meet with the hardware arriving in its data centres, not impractical or economically challenging, but physically incompatible with the operating characteristics of the systems that Frankfurt’s hyperscale operators were being asked to install. The PUE target of 1.2 for new facilities commissioning from July 2026 was written into law before anyone in the Federal Ministry had run the arithmetic on what happens to a facility’s power usage effectiveness when it installs several hundred NVLink-interconnected GPU racks drawing over one hundred kilowatts each, served by direct-to-chip liquid cooling systems, surrounded by coolant distribution units and secondary heat rejection infrastructure that themselves consume substantial electrical power. The amendment is not a sign of bureaucratic goodwill. It is a record of a government discovering, three years late and under EU infringement pressure, that it had legislated against physics.
The EnEfG 2026 Amendment and the Arithmetic of Bureaucratic Retreat
What the PUE Numbers Actually Reveal About AI Hardware
Power Usage Effectiveness is a ratio. It measures the total energy consumed by a data centre divided by the energy consumed by the IT equipment the facility exists to operate. A PUE of 1.0 is physically impossible, representing a data centre in which every joule of electricity input is used by servers with nothing consumed by cooling, lighting, power distribution losses, or any other overhead. A PUE of 1.2, which the original EnEfG required of all new facilities commissioning from July 2026, represents a facility where for every unit of IT power consumed, only an additional twenty percent is consumed by everything else. This is an extraordinarily demanding target under conventional data centre operating conditions, requiring highly optimised cooling infrastructure, minimal power distribution losses, and stable workloads that allow the facility to operate near its designed efficiency point consistently.
The specific failure mode that AI hardware creates for a PUE 1.2 target involves several compounding mechanisms that the EnEfG’s framers did not adequately model. A GPU training cluster does not operate at a uniform, predictable load profile. It cycles between different computation phases, network communication phases, and checkpoint storage operations, creating power draw variations that air-cooled infrastructure handles by varying fan speeds but that liquid-cooled infrastructure manages through coolant flow regulation and heat rejection adjustments. Coolant distribution units that are running at full cooling capacity for a peak training workload and then throttling back during checkpoint operations represent a dynamic overhead load that varies in a way that PUE measurements taken over annual averages may not accurately capture at the instantaneous operating points where efficiency is worst.
The Gleiss Lutz analysis of the April 2026 draft noted that the relaxation of PUE ratios is moderate at best, and that factors such as different cooling technologies continue to be disregarded in the assessment methodology. This is the specific point that the amendment fails to resolve: the PUE metric, regardless of whether its threshold is 1.2 or 1.3, does not distinguish between a facility whose overhead power consumption reflects efficient thermal management of conventional cloud workloads and one whose overhead reflects the thermodynamic cost of managing heat from two-hundred-kilowatt AI racks in a building whose original design assumed heat rejection at one-fifth that density. The metric conflates these situations into a single number and then applies a regulatory threshold to that number, producing compliance or non-compliance assessments that say nothing useful about whether the facility is appropriately engineered for its actual workload.
Germany’s March 2026 National Data Center Strategy, published on March 18, 2026, recognised this structural inadequacy indirectly by linking expansion to technological innovation across the data centre technology stack, including cooling technology, energy efficiency, and waste heat use. The ambition to at least double data centre capacity by 2030 and at least quadruple AI-related capacity over the same period, set alongside regulatory requirements that the same government was simultaneously being forced to walk back due to physical impossibility, illustrates the fundamental tension in German data centre policy: ambitious capacity goals calibrated to competitive necessity sitting alongside efficiency requirements calibrated to political sustainability targets, with no mechanism to reconcile them when the hardware connecting these two policy domains refuses to comply with both simultaneously.
The Thermodynamics of Low-Grade Waste Heat and the ERF Illusion
The Temperature Gap That No Regulation Has Resolved
The Energy Reuse Factor obligation embedded in the original EnEfG and preserved in modified form in the 2026 amendment is, in engineering terms, the most technically demanding requirement the legislation places on data centre operators, and the one that most directly exposes the gap between the policy’s intent and the physical realities of liquid-cooled AI infrastructure. The ERF requires data centres commissioned from July 2026 to achieve a minimum of ten percent energy reuse from July 2026, rising to fifteen percent from July 2027 and twenty percent from July 2028. The requirement is grounded in a legitimate sustainability objective: data centres generate substantial waste heat, and that heat could in principle substitute for fossil-fuel-generated heat in Germany’s municipal district heating networks, reducing both the facilities’ environmental footprint and the carbon emissions of the residential and commercial heating sector they serve.
The thermodynamic obstacle that makes this objective structurally problematic for liquid-cooled AI facilities is the temperature differential between the heat that modern cooling systems produce and the temperature that district heating networks require to function. Germany’s municipal district heating infrastructure, the Fernwärme networks that serve residential and commercial buildings in Frankfurt, Berlin, Munich, and Hamburg, operates at supply temperatures of sixty to eighty degrees Celsius, a range that allows the hot water flowing through pipes to heat buildings without additional energy input at the point of use. An air-cooled data centre that uses chilled water loops and computer room air handlers rejects heat at temperatures that can, in favourable conditions, be elevated to levels compatible with district heating injection, because the heat rejection process in an air-cooled system operates at temperatures determined by ambient air conditions, chiller setpoints, and the design of the heat rejection plant.
A direct-to-chip liquid cooling system serving a rack drawing one hundred and fifty or two hundred kilowatts operates at a fundamentally different temperature regime. The coolant enters the cold plates at approximately thirty to forty degrees Celsius and exits at approximately forty-five to fifty degrees Celsius, having absorbed heat directly from the GPU package surface. This exit temperature of forty-five to fifty degrees Celsius is the temperature at which the waste heat leaves the facility’s internal cooling loop and enters the heat rejection system. The gap between this exit temperature and the sixty to eighty degrees Celsius minimum required for district heating injection is not a gap that better thermal management at the rack level can close, because it is determined by the operating temperatures of the chip package itself. A GPU running at the thermal design power it was engineered for produces heat at the temperature that its internal thermal resistance and coolant flow rate dictate, and that temperature is substantially below what district heating infrastructure requires.
The industrial heat pump is the engineering solution that bridges this gap. A water-to-water heat pump can take the forty-five-degree-Celsius output of a liquid cooling loop and elevate it to sixty-five or seventy degrees Celsius, making it compatible with district heating injection. This is technically correct. The economic and energetic cost of doing so is what makes the ERF mandate, as applied to liquid-cooled AI facilities, a structural contradiction within the EU’s Energy Efficiency First principle rather than an implementation of it. A heat pump consumes electrical power to operate. The ratio of heat output to electrical input for a water-to-water heat pump at this temperature lift is described by its coefficient of performance, typically in the range of three to four, meaning three to four units of heat energy are delivered for each unit of electricity consumed. When the electricity consumed by the heat pump is added to the facility’s total power draw, it increases the PUE by a meaningful increment, potentially enough to push a carefully optimised facility back above the threshold that the EnEfG’s relaxed 1.3 standard requires.
The Pinsent Masons analysis of the 2026 amendment’s waste heat provisions noted that the draft provides a new exemption from the obligation to utilise waste heat if there is no technically and economically feasible connection to an existing or planned district heating network within a five-kilometre radius. This exemption acknowledges, in the formal language of legislative drafting, exactly the physical and economic problem that the preceding analysis describes. The problem is not that Frankfurt lacks district heating networks. The city operates an extensive Fernwärme system. The problem is that connecting a facility producing forty-five-degree waste heat to a network requiring sixty-five-degree supply heat requires industrial heat pump infrastructure whose capital and operating cost the five-kilometre exemption threshold does not address. A facility within two kilometres of a district heating main that cannot economically justify the heat pump capital expenditure required to bridge the temperature gap is not helped by an exemption that applies only to facilities with no district heating infrastructure within five kilometres.
Grid Redispatch 2.0 and the TSO Maturity Wall
Two Hundred and Seventy Gigawatts of Applications, One Overloaded Network
Germany’s transmission system operators, the four TSOs responsible for managing the high-voltage grid that moves bulk electricity between Germany’s generation zones and its consumption centres, entered 2026 facing a queuing crisis of a scale that their institutional processes were not designed to manage. The Baker McKenzie analysis of the maturity evaluation procedure that Germany’s four TSOs jointly proposed on February 5, 2026 provided the quantitative baseline: as of the end of Q3 2025, the four TSOs had received 717 grid connection applications totalling approximately 270 gigawatts in requested capacity. The German Association of Energy and Water Industries survey published in November 2025 found that distribution system operators had received applications for an additional 569 gigawatts of distribution-grid capacity. Together, these queues represent requested connections totalling over eight hundred gigawatts, in a country whose total installed generation capacity is approximately 250 gigawatts.
The structure of this queue reflects the same speculative application problem that ERCOT identified in the American context: applicants submit grid connection requests before they have secured financing, before they have confirmed hardware delivery timelines, and in some cases before they have finalised site selection, preserving optionality across multiple potential sites by maintaining parallel applications. The December 2025 amendment to the Power Plant Grid Connection Ordinance, which abolished the first-come-first-served principle for projects above one hundred megawatts, was specifically designed to address what the Federal Ministry described as zombie battery storage facilities, projects with grid connection requests that have no realistic prospect of implementation. The same logic applies to speculative data centre applications, though the legislative language did not specifically address them.
The maturity evaluation procedure that the TSOs proposed as a replacement for first-come-first-served operates through annual process cycles. During a designated window, all submitted applications are collected, assessed against a maturity framework that evaluates the credibility and readiness of each project, and ranked for connection allocation based on that assessment rather than on application date. The procedure was expected to be implemented as early as April 2026, subject to Federal Network Agency approval. Its practical consequence for data centre operators is a shift from a system in which the primary competitive advantage was submitting an application early to one in which the primary competitive advantage is demonstrating project readiness through binding financial commitments, hardware procurement orders, and site preparation evidence.
For large data centre projects in Frankfurt’s Rhine-Main corridor, this maturity framework interacts with the grid’s structural imbalance in a specific way that exacerbates the connection timeline problem rather than simply reshaping the queue. Germany’s renewable energy generation is concentrated in the north and northwest, where onshore wind generation is abundant. Germany’s primary computing and industrial consumption is concentrated in the south and west, where Frankfurt, Bavaria, and Baden-Württemberg draw the bulk of the commercial and industrial electricity that the grid serves. Moving power from northern generation to southern consumption requires the high-voltage transmission lines that cross Germany’s geographic centre, and those lines are chronically congested, forcing TSOs to redispatch, paying northern generators to reduce output and southern generators to increase output, at a cost that reached billions of euros annually before the current data centre buildout began adding further load concentration to the southern corridors.
AlgorithmWatch’s November 2025 investigation of Frankfurt’s data centre energy situation documented the operational consequence of this structural imbalance with unusual directness: data centres in Frankfurt are using fossil gas to respond to energy bottlenecks that occur due solely to the high energy requirements of the industry. A facility that cannot receive the renewable electricity it has contracted to purchase via the high-voltage grid because transmission congestion prevents that electricity from reaching its connection point is a facility that is operating on gas generation regardless of what its power purchase agreement says. Germany’s hundred percent renewable electricity requirement for data centres from January 2027 applies to the electricity that facilities purchase on a balance sheet basis. It does not, and cannot, guarantee that the physical electrons flowing through the connection at any given moment originate from renewable sources rather than from the gas turbines that southern German grid operators dispatch to manage transmission constraints.
The BESS Gambit and Its Real Cost
The German data centre industry’s institutional response to the grid connection queue, and the pathway that the maturity evaluation procedure implicitly rewards, is increasing reliance on behind-the-meter battery energy storage systems as a mechanism for demonstrating grid compatibility and shortening effective connection timelines. A data centre that installs substantial BESS capacity can absorb off-peak renewable generation when the grid has surplus, store it, and deploy it during peak demand periods when the transmission-constrained grid cannot reliably deliver what the facility needs. This approach satisfies the maturity framework’s preference for projects with credible grid stability credentials, reduces the facility’s net exposure to redispatch-driven renewable supply interruptions, and can, in some configurations, allow a facility to operate at a connection capacity below what its IT load would otherwise require by using stored energy to cover demand peaks.
The cost of this approach is substantial and represents a capital expenditure category that European data centre economics have not previously needed to budget at anything approaching the current scale. Battery energy storage systems sized to serve as a meaningful buffer for a hundred-megawatt data centre campus represent capital commitments of tens of millions of euros for the storage hardware alone, before the grid interface equipment, installation, facility integration, and ongoing operations and maintenance costs are included. The four German TSOs had received nearly seven hundred grid connection requests for battery energy storage alone totalling two hundred and fifty gigawatts, as of the Energy-Storage.News analysis published in February 2026, illustrating that BESS is not a niche response to grid congestion but is becoming a standard component of serious infrastructure development across the German market.
NTT’s announced 480-megawatt data centre in Nierstein, in the Rhine-Hesse region south of Frankfurt, illustrates the scale at which the Frankfurt data centre belt is attempting to expand beyond the primary hub’s most congested grid connection points. Moving fifty kilometres from Frankfurt’s core in search of grid capacity that the Frankfurt substation infrastructure cannot provide represents an operational compromise between location optimisation, measured by proximity to DE-CIX and the enterprise customer base that Frankfurt’s internet exchange node serves, and power availability, measured by what TenneT’s transmission infrastructure can actually deliver within acceptable connection timelines. The belt expansion is a rational market response to the maturity wall’s constraints. It is also a spatial dispersal of infrastructure that was originally concentrated for latency and connectivity reasons, a dispersal whose networking cost implications do not appear in the grid connection economics that drove the site selection.
The Liquid Cooling Supply Chain and Fluid Sovereignty
European CDU Manufacturing and the PFAS Horizon
The liquid cooling transition that Germany’s EnEfG 2026 amendment implicitly acknowledges as the operational standard for high-density AI facilities has its own supply chain constraints that operate independently of the regulatory and grid problems this analysis has examined, and that add a further layer of uncertainty to the deployment timelines of the facilities waiting in Germany’s grid connection queue. Coolant distribution units, the rack-level infrastructure that delivers liquid cooling to individual server components in a direct-to-chip architecture, are not commodity hardware. They are engineered systems requiring precise thermal and hydraulic specification for each specific rack configuration, manufactured by a relatively small number of companies with the thermal engineering depth to build them reliably at the power densities that next-generation AI hardware demands.
The European manufacturing base for liquid cooling CDUs and associated manifold infrastructure is growing but remains structurally inadequate for the scale of demand that Germany’s data centre expansion pipeline represents. Vertiv, whose EMEA liquid cooling expansion was announced in May 2026, Schneider Electric, and a smaller group of European specialists represent the primary supply options, alongside American and Asian manufacturers whose European delivery timelines are extended by logistics and customs considerations that add weeks to the installation schedules of projects already under timeline pressure from grid connection delays. Germany’s Addleshaw Goddard legal analysis of the data centre sector specifically identified cooling technology as a strategic priority area where innovation can create competitive advantages, implicitly acknowledging that conventional procurement pathways for cooling hardware do not reliably deliver the lead times that current market conditions require.
The regulatory horizon that represents the most consequential forward uncertainty for the European liquid cooling supply chain is the EU’s restrictions on per- and polyfluoroalkyl substances, the PFAS chemicals used in certain dielectric fluids that are employed in immersion cooling systems. European PFAS restrictions under the REACH regulation, driven by the EU’s Chemicals Strategy for Sustainability, are progressing toward what some regulatory timelines suggest could be the most significant chemical restriction in European industrial history. Several of the dielectric fluids currently used in single-phase and two-phase immersion cooling deployments contain PFAS compounds, and the trajectory of European PFAS regulation creates genuine uncertainty about whether the coolants being specified in current facility designs will remain commercially available, legally usable, and procurable at acceptable prices within the operational lifetime of the facilities that are being designed around them today.
German data centre operators planning immersion cooling deployments for AI facilities that will commission in 2026 and 2027 and operate through the 2030s are making material selections today that will interact with a regulatory framework that is still being written. The Pinsent Masons and Taylor Wessing analyses of the EnEfG amendment both note that different cooling technologies continue to be disregarded in the assessment methodology, a characterisation that applies with equal force to the PFAS regulatory question: Germany’s energy efficiency compliance framework has been developed and modified without specific engagement with the chemical and material constraints that are beginning to constrain the cooling systems the framework is intended to govern.
What the Gridlock Actually Costs
Frankfurt’s Position in a Competitive European Landscape
The practical consequence of Germany’s regulatory-grid-cooling compound constraint is visible in the competitive repositioning that European data centre investment is already undergoing. The Frankfurt Rhine-Main corridor remains Europe’s largest data centre market by installed colocation capacity and internet traffic volume, anchored by DE-CIX, the world’s largest internet exchange point by traffic volume, and by the enterprise customer concentration of Germany’s financial and industrial heartland. The infrastructure advantages that created this position are real, durable, and not easily replicated by competing markets. They are, however, not sufficient to guarantee Frankfurt’s share of the AI-era investment wave if the regulatory and grid conditions governing new development systematically extend deployment timelines beyond what hyperscalers and neoclouds racing to serve enterprise AI demand can accept.
Spain’s Madrid market, Nordic locations in Stockholm, Oslo, and Helsinki, and emerging secondary European markets in Poland and the Czech Republic are all capturing investment that the Frankfurt corridor’s congested grid connection queue and complex regulatory compliance environment is delaying or deterring. A hyperscaler that can commission a facility in Spain or the Nordics within twelve to eighteen months of a grid connection application, against a Frankfurt maturity evaluation procedure that extends effective timelines to seven to ten years for facilities without pre-existing connection rights, is not making a preference for those markets. It is responding to the gap between what Frankfurt’s regulatory and grid environment offers and what the AI hardware deployment cycle requires.
Germany’s March 2026 National Data Center Strategy acknowledged this competitive dynamic explicitly, linking expansion to faster grid access alongside sustainability and technological innovation. The strategy document explicitly states that innovation in the practical operation of data centres can create competitive advantages and strengthen European value creation. The tension between this recognition and the grid connection maturity framework that the same government’s TSOs were implementing simultaneously, between the data centre strategy’s aspiration and the grid policy’s operational reality, defines the Frankfurt gridlock that this analysis has examined. Germany is not failing to understand the AI infrastructure race. It is failing to resolve the specific contradiction between a regulatory framework designed for a previous era of data centre development and the physical requirements of the hardware generation that is arriving to occupy the grid capacity that the regulatory framework was built to govern.
The EnEfG amendment is a recognition of that failure, not a resolution of it. Walking back a PUE threshold from 1.2 to 1.3, restructuring waste heat reuse requirements that were thermodynamically impossible for liquid-cooled AI facilities to meet, and granting exemptions to operators without district heating access within five kilometres corrects the most immediately untenable elements of the original legislation. It does not address the grid redispatch congestion that caps how much renewable electricity Frankfurt’s data centres can physically receive regardless of what their power purchase agreements say. It does not resolve the temperature gap between liquid cooling outputs and district heating requirements that makes the ERF mandate a net energy cost rather than a net energy saving for the facilities it applies to. It does not eliminate the PFAS regulatory uncertainty that affects the cooling fluid specifications of immersion systems being designed today for a decade of operation. Germany has retreated from the most visible element of its data centre regulatory overreach. The deeper structural constraints remain intact.
