Energy and AI: the state of Dutch data centers

The rapid rise of artificial intelligence brings far-reaching consequences for the physical infrastructure in the Netherlands. Where data centers used to be assessed mainly on their network connectivity, latency, and storage capacity, the absolute energy demand is now central to every siting question. Due to the exponential increase in compute power needed for training and large-scale running of complex language models, power demand in the digital sector is growing rapidly. This article offers an in-depth analysis of the state of affairs surrounding Dutch data centers, the persistent nationwide grid congestion, and the measures being taken to maintain a balance between digital innovation and the limits of the national energy transition.

The physical footprint and power density of modern AI clusters

Traditional cloud services and web hosting facilities consume considerable amounts of energy, but the arrival of advanced generative AI places an unprecedented and fundamentally different strain on the power supply. Processing advanced queries, performing matrix calculations, and continuously updating weights in gigantic neural networks requires specialized hardware accelerators that run permanently under high load. In the background article on energy consumption and compute power the fundamental relationship between algorithms and kilowatt-hours is explored in more depth, making clear that efficiency gains on the software side only partly offset the enormous growth in total data volume and model computations.

In the Netherlands, this physical load traditionally concentrates in the Amsterdam metropolitan region (the so-called AMS hub) and around strategic hubs in Noord-Holland and Flevoland. Data center operators are noticing that thermal load and the required power density per server rack have risen dramatically. Where a standard enterprise rack used to require a power draw of a few kilowatts, modern AI clusters with densely packed accelerators quickly consume tens to more than forty kilowatts per cabinet. This forces data center operators into drastic redesigns of their power distribution grids and advanced cooling systems.

Grid congestion and the queues on the electricity network

The most pressing bottleneck for the further expansion of AI infrastructure on Dutch soil is not so much a lack of investor capital or available square meters, but structural grid congestion. Regional grid operators and national transmission system operators repeatedly report that the capacity of both the high-voltage grid and the underlying medium-voltage grid is fully utilized in numerous locations. New data centers requesting a heavy grid connection for large-scale compute capacity end up in long-running queues. This creates a tension in which the national ambition to remain a digital frontrunner in Europe clashes with the physical limitations of the electricity grid.

The impact of this grid congestion extends beyond just the siting of new facilities. Existing data centers looking to scale up to meet rising market demand run into hard limits when obtaining additional transport capacity. This is driving a search for alternative solutions, such as spreading load over time and optimizing network traffic. Although software-based routing increases efficiency, it does not solve the fundamental shortage of megawatt-hours at the source.

The Jevons paradox and the dynamics of energy efficiency

A frequently heard argument in the tech sector is that newer generations of chips and optimized models automatically become more efficient per computation. While it is true that energy efficiency per floating-point calculation steadily improves, in practice this phenomenon often leads to the Jevons paradox: as a technology becomes more efficient and cheaper to use, total usage rises exponentially. The Power Usage Effectiveness (PUE) of Dutch data centers ranks among the sharpest in the world, indicating very efficient management of cooling and facility systems, but absolute power consumption keeps rising due to the enormous increase in the number of processed transactions.

The balance between hardware efficiency and total consumption requires constant policy adjustment. Where data centers used to be designed for a predictable, flat server load profile, the dynamics of intensive training runs and continuous inference now produce erratic power spikes. This requires operators to work closely with energy suppliers to conclude flexible contracts that respond to the availability of renewable generation.

Cooling systems, liquid cooling, and the connection to heat networks

The enormous heat output of modern AI hardware makes traditional air cooling inadequate for high-density racks. As a result, Dutch data centers are rapidly switching to advanced cooling techniques, including direct-to-chip liquid cooling and closed water systems. These technologies remove heat more efficiently, while at the same time offering an interesting opportunity for the built environment: waste heat.

Municipalities and regional heat companies are increasingly entering into agreements with data center operators to use this waste heat for urban heat networks. Implementing this, however, comes with technical and logistical hurdles, such as bringing the water to the right temperature and laying large-scale pipeline infrastructure. Moreover, not every data center is located near densely populated residential areas, meaning waste heat in peripheral areas is sometimes still released to the outside air via cooling towers.

API aggregation and software-based load balancing

To efficiently manage scarce compute power and dampen throughput peaks, modern infrastructures make intensive use of smart software layers. By intelligently balancing and distributing requests across various providers and compute nodes, the load on specific locations is optimized. This is demonstrated by the explanation of LLM API aggregators showing how requests are efficiently channeled to reduce peak load. Such architectural choices help keep operational efficiency on level, though they remain subordinate to the physical limits of the underlying electricity grid.

Software-based optimization also extends to model selection. Deploying lighter models where sufficient not only eases users' budgets but also directly lowers the thermal and electrical load within server rooms. Yet demand for heavier models keeps increasing, particularly for complex analysis and reasoning tasks.

The rise of autonomous systems and the impact on baseload

The nature of workloads in data centers is fundamentally shifting due to the transition to autonomously operating software. Where users used to query a model on an ad hoc basis, systems are increasingly deployed in ongoing automation processes. In the article on autonomous AI systems it is described how multi-agent loops allow applications to independently go through multiple thinking and action steps before a task is completed. This means background activity in data centers is less dependent on office hours, and the baseload on the power grid remains high throughout the night.

This continuous load poses an additional challenge for grid operators. Because the traditional off-peak hours are flattening out, there is less room to carry out large-scale grid maintenance without directly affecting the continuity of business-critical systems.

Legislation, the National Data Center Strategy, and European frameworks

The government has tightened the conditions for the establishment of new data centers in recent years. The National Data Center Strategy sets strict requirements for energy efficiency, societal added value, and the mandatory offtake or supply of waste heat. Only facilities that can demonstrate they play an essential role in the digital economy and are demonstrably becoming more sustainable still receive priority for scarce grid capacity.

In addition, European regulatory frameworks play a guiding role in how organizations handle their data flows and compute capacity. The broader policy timeline is set out in the overview on the implementation of the AI Act in 2026, which also addresses the mandatory transparency and environmental reporting that large-scale compute facilities and model developers must comply with.

Local resilience and behind-the-meter solutions

In response to the long queues at grid operators, data centers and regional industrial estates are increasingly experimenting with 'behind-the-meter' solutions. This includes large-scale battery storage (BESS), local microgrids, and combinations with their own renewable generation such as large-scale solar parks or wind turbines on business parks. These systems absorb short power spikes and make data centers less vulnerable to short-term fluctuations on the national grid.

Although these decentralized buffers can absorb part of the peak load, they are not a complete substitute for a heavy grid connection. For continuously training large models and powering hyperscale data centers, a direct, heavy connection to the high-voltage grid remains necessary.

Conclusion: The balance between digitalization and the energy transition

The state of Dutch data centers in 2026 shows that artificial intelligence is a driving force behind economic and technical innovation, but at the same time places a heavy burden on the country's physical infrastructure. Grid congestion, rising power density, and strict sustainability requirements are forcing all parties involved into far-reaching innovation and strict prioritization. Only through close and pragmatic cooperation between government, grid operators, and the tech sector can the Netherlands maintain its strong digital position without the national energy transition grinding to a halt.