Climate AI and energy consumption: a tension Dutch startups cannot ignore

AI startups working on climate and the energy transition carry a built-in dilemma. The models they use to optimise heat networks, analyse satellite imagery or process energy data require computing power, and therefore electricity. As models grow heavier and applications need to scale, that energy footprint grows with them.

In the Netherlands, a cluster of startups is active that applies AI to specific climate problems, from grid congestion to vegetation risk along power lines. They typically work with specialised, narrower models rather than large generative systems. That makes their situation different from that of a generalist such as a large language model, but it does not exempt them from the question of how they account for their own consumption.

Why measuring your own energy footprint remains difficult

Anyone who wants to know how much energy their AI model consumes quickly runs into a measurement headache. Training costs can be reasonably quantified through GPU hours and data centre consumption, but inference, running a model on production data day to day, is harder to isolate. Cloud providers do not always report transparently on the energy mix per region, and shared infrastructure makes it difficult to attribute consumption precisely.

Companies such as Energyworx, which delivers cloud-based energy data processing to utilities from its base in Houten, and Withthegrid from Utrecht, which monitors and controls IoT assets, run their models on shared cloud infrastructure. They are therefore dependent on the sustainability choices of their cloud provider. Large hyperscalers such as Microsoft and Google publish renewable energy targets, but the reality varies per data centre and per moment. Without detailed API-level reporting, accurate measurement remains an approximation.

A useful indicator gaining traction is the carbon intensity of the region where a model runs, expressed in grams of CO₂-equivalent per kilowatt-hour. By scheduling computing tasks at times when the electricity grid is greener, or by choosing regions with lower intensity, startups can reduce their emissions without changing the models themselves. This does require access to real-time data on the energy mix, precisely the type of data that companies like Energyworx process for utilities.

Technical choices that influence consumption

A model's architectural choices have a direct impact on energy consumption. Gradyent from Rotterdam, which optimises heat networks through a digital twin after raising a Series B of €28 million in 2025, works with application-specific models that operate on a limited amount of sensor data. This is fundamentally different from training a large language model on text from across the entire internet. A narrower model, trained on domain-specific data, requires considerably less computing power for both training and inference.

Tibo Energy from Eindhoven, which handles energy management for businesses and helps prevent grid congestion, raised €6 million in seed capital in early 2025. For a young company at this stage, model choices already determine future operational costs. Compact, efficient models are not only more environmentally friendly but also cheaper to run in production, which strengthens the business case.

Techniques such as model compression, quantisation and knowledge distillation make it possible to train smaller variants that approximate the behaviour of a larger model. How widely these methods are already applied within the Dutch climate AI sector is not well documented, which is in itself an observation about the state of transparency in this field.

Weighing climate gains against costs

A central question is whether the climate benefit of an application outweighs its energy consumption. Overstory from Amsterdam, the best-funded climate AI startup in this cluster with €67.8 million raised, uses satellite imagery and AI to identify vegetation risks near electricity networks. Timely detection can prevent power outages, which in turn saves energy that would otherwise be lost during failures and restoration. The net climate balance depends on how frequently the model runs, how large the datasets are, and how significant the avoided losses are.

Sympower from Amsterdam, active since 2015 and funded with €74.5 million, links the energy flexibility of businesses to the balancing of national power grids. The premise is that smarter management of supply and demand reduces the use of peak energy, often from fossil sources. If that works, the avoided emissions are many times greater than what the underlying AI systems consume. But that trade-off requires figures that are rarely publicly available.

52impact from Amersfoort, which uses geospatial data for environmental and sustainability risk management and closed a €15 million Series B in 2022, operates on the analytics side. Companies that map their environmental exposure can make better-informed decisions about reduction. The AI component processes large volumes of spatial data, which is computationally intensive, but the decisions that follow from that analysis can have substantially larger effects.

What investors and policymakers can expect

For investors in climate AI, the energy footprint of a portfolio startup is rarely an explicit due diligence factor. That is beginning to change, partly because the EU taxonomy for sustainable finance and upcoming reporting requirements around digital infrastructure will demand greater transparency. Startups that already measure and report today are building a head start in that accountability process.

Policymakers looking to encourage climate AI face a comparable trade-off. Subsidies and investment programmes for sustainable digital innovation could attach conditions to the energy performance of supported systems, similar to how energy efficiency requirements already apply to hardware and buildings.

Within the Dutch and broader European climate tech ecosystem, the number of startups using AI as the core of their product is growing. Whether they systematically measure and minimise their own footprint is not merely a technical detail, it goes to the heart of the credibility of the climate mission they espouse. Transparency on this point remains scarce, and that is a gap that will eventually be closed, either by regulation or by the market.