The Debate on Datacenters and AI Computing Capacity in the Netherlands

The rapid rise of artificial intelligence, and in particular Large Language Models (LLMs), has sparked a global race. Behind the scenes of this virtual intelligence, however, lies a tangible, physical reality: datacenters. The Netherlands, traditionally one of Europe's most important digital hubs thanks to exchange points like the AMS-IX and an excellent power grid, finds itself at the center of a complex debate. How do we balance the massive demand for AI computing capacity with the hard limits of our overloaded electricity grid and spatial planning?

The discussion about datacenters is not new, but the arrival of advanced generative AI has fundamentally changed the parameters. While traditional cloud computing already required significant resources, AI demands an exponential leap in power consumption and cooling capacity. This article dissects the technical necessity of these datacenters, the impact of grid congestion on the business climate, and the broader geopolitical implications surrounding digital sovereignty.

Why AI Training and Inference Drive Up Power Demand

To understand why AI's impact on infrastructure is so significant, we must distinguish between the two primary phases of an AI model: training and inference. Both processes are extremely computationally intensive, but they have different profiles when it comes to hardware and energy consumption.

The Difference Between Training and Inference

Training a state-of-the-art foundation model (such as GPT-4 or similar open-weights models) is a process that can take months. During this process, billions of parameters are optimized by processing massive datasets. This requires tens of thousands of specialized GPUs (Graphics Processing Units) running continuously at maximum load. In theory, these training tasks can take place anywhere in the world, as long as cheap and abundant power is available.

Inference, on the other hand, is the process where the trained model is actually used by end-users to answer questions or generate text. Although a single query consumes less power than training the model, the massive and simultaneous use by millions of people worldwide quickly adds up. Ideally, inference datacenters need to be close to the end-user to minimize latency (delay), which further drives the demand for datacenters around densely populated areas, such as the Randstad.

The Impact on Power and Cooling

The transition from traditional CPUs to the latest generation of AI GPUs has drastically increased the "power density" per rack (a cabinet containing servers). A traditional server rack in a datacenter consumes an average of 5 to 10 kilowatts (kW) of electricity. In contrast, a rack fully equipped with advanced AI chips can consume 40 to 100 kW or more. This immense concentration of heat means that traditional air cooling is often no longer sufficient, leading the industry to increasingly switch to liquid cooling. This, in turn, raises questions about water consumption, particularly during periods of drought—an argument frequently cited by opponents of datacenter expansions.

The Physical Reality: Hyperscalers versus Colocation

In the Dutch landscape, "datacenters" are often spoken of as a single uniform category. However, the dynamics surrounding AI force us to distinguish between different types of facilities. The interests, spatial footprint, and social impact differ greatly depending on the type of datacenter.

What are Hyperscale Datacenters?

Hyperscalers are the giant facilities that are mostly built and managed by the world's largest tech companies, such as Google, Microsoft, and Meta. These datacenters, often located in areas like Eemshaven or Wieringermeer, are primarily designed for massive scale. They cover dozens of hectares and require a direct, large-scale connection to the high-voltage grid. For large-scale training of AI models, hyperscalers are essential due to their ability to run tens of thousands of chips in a single interconnected network.

The Role of Smaller Colocation Datacenters

On the other side of the spectrum are colocation datacenters. These are facilities where the owner (for example, Equinix or Digital Realty) rents out space, power, and cooling to various clients, including local businesses, governments, and smaller AI startups. These datacenters are often located closer to urban hubs like Amsterdam. For local innovation and running smaller, specialized AI models, these facilities are crucial. Read more about the development of such more compact networks in our article on open-source LLM trends, where we discuss the shift toward more decentralized AI.

Grid Congestion Fundamentally Changes the Discussion

The biggest bottleneck for the expansion of AI infrastructure in the Netherlands is currently not the availability of land, but the capacity of the electricity grid. The Dutch power grid is "creaking at the seams". The national grid operator TenneT, as well as regional operators, have been warning for years about grid congestion: the situation where the demand for electricity transmission is greater than what the physical cables and transformer stations can handle.

Datacenters require a so-called "baseload" power supply: they need a constant and massive supply of electricity 24 hours a day, 7 days a week. This is in contrast to renewable energy sources such as solar and wind, which are weather-dependent. This creates a mismatch on the grid. For an in-depth look at how smart grids and dynamic energy consumption can help, we refer to our comprehensive analysis on AI and energy.

Current grid congestion acts de facto as a moratorium on the construction of new datacenters in many Dutch regions. Developers face long waiting lists for a power connection, leading to frustration in the tech sector, which fears that the Netherlands is losing its competitive position in the digital economy to countries like France or the Scandinavian nations, where grid capacity and (cheaper) nuclear or hydropower are still available.

The Location Debate: Proponents and Opponents

The decision to allow new datacenters, especially on the scale required for AI, invariably leads to fierce local and national political debates. The arguments on both sides of the table are fundamental and reflect a broader search for the balance between economic growth and spatial quality.

Arguments of Proponents

Proponents, including industry associations such as NLdigital and parts of the business community, emphasize that datacenters form the foundation of the modern economy. Without the necessary infrastructure, the Netherlands cannot play a meaningful role in the AI revolution. They argue that:

• Innovation and Employment: Although a datacenter itself generates relatively few jobs, the presence of high-quality digital infrastructure creates an attractive ecosystem for tech companies, AI researchers, and startups.
• Data Sovereignty: If we do not build the infrastructure domestically, our data, and the added value that AI derives from it, will flow to facilities abroad.
• Sustainability: Modern datacenters are becoming increasingly efficient. Furthermore, they offer opportunities for utilizing waste heat, which can be used to heat surrounding residential areas or greenhouse farming, provided the infrastructure is properly constructed.

Arguments of Opponents

Opponents, ranging from local action groups to environmental organizations, point to the heavy toll these facilities take on the living environment. Their primary objections are:

• Landscape 'Boxification': Hyperscale datacenters are giant, windowless buildings that dominate the horizon and come at the expense of scarce agricultural or natural land.
• Grid Crowding Out: As long as the grid is full, the arrival of a datacenter means that other local businesses cannot transition to sustainable energy or expand, and housing projects face delays.
• Environmental Impact and Water Consumption: In addition to indirect power consumption, many large facilities still use drinking water for cooling, which can put local drinking water supplies under pressure during hot, dry summers.

This friction also affects regulation. The way governments supervise these developments and issue permits is rapidly evolving. You can read more about this in our section on AI supervision in the Netherlands, where we discuss how national frameworks attempt to guide the rollout of AI in the right direction.

Digital Autonomy and the European Context

The discussion about AI datacenters transcends Dutch borders. At the European level, there is a growing realization that dependence on non-European technology companies poses a strategic risk. The concept of 'digital autonomy' (or Sovereign AI) is high on the agenda in Brussels and The Hague.

Currently, the hardware and cloud market for AI is heavily dominated by American companies (such as NVIDIA, Microsoft, AWS, and Google), and a significant part of the supply chain resides in Asia. If the European Union wants to be able to enforce its own values—as laid down in the GDPR and the AI Act—in language models, an independent technological stack is required. This means European models, trained on European data, running on European servers.

The Netherlands faces a dilemma here. On the one hand, we want to be a leader in responsible AI and stimulate independence. On the other hand, facilitating this independence requires massive infrastructural investments in our own country. For organizations and developers who want to set up infrastructure for models themselves, we offer in-depth technical guides. For instance, on our external knowledge portal, you can read all about the specific hardware requirements for running LLMs locally.

Conclusion: Seeking a New Balance

The debate on datacenters and AI in the Netherlands will only intensify in the coming years. AI is not a passing trend; it is a fundamental shift in how we work, think, and solve problems. The computing capacity required for this is massive and cannot be ignored.

The solution will likely not lie in a binary 'yes' or 'no' to new datacenters, but in stricter governance and smarter technology. This includes requirements for spatial integration, mandatory supply of waste heat, the use of flexible AI training schedules (which pause when the power grid is overloaded), and a firm prioritization of facilities that directly contribute to European digital autonomy. The Netherlands must choose: either we consciously build the infrastructure of the future with strict societal conditions, or we accept that control over the most important technology of this century will be exercised elsewhere.