Baseline and Trajectory of AI – Ten Key Takeaways
PDF

To understand the strategic implications of AI for the EU and the Netherlands, we outline ten foundational insights gained from our expert conversations that capture the political, technical and societal dynamics shaping the current AI landscape.[11]

1.
Great powers – notably the US and China, and to a lesser extent the EU and some Member States, particularly France – regard AI as a strategic race. Access to infrastructure, talent and funding are considered core requirements for AI leadership.
2.
Uncertainty in time horizons increases. While AI timelines seem to accelerate, it is unclear where we are on the so-called S-curve of innovation – that is, how mature AI technologies are and where there is still space to improve. This makes prediction impossible and scenario planning very difficult, but indispensable.
3.
The accessibility paradigm is shifting – on the one hand, new entrants like xAI have rapidly approached top-tier capabilities based on an intense capital investment; on the other hand, DeepSeek has shown that capabilities are not a linear function on compute, and that efficiency using resources is also an element to consider.
4.
The real risk is not AI itself, but how governments, companies and citizens embed AI into society.[12] Its use can lead to significant efficiency gains, but also amplifies important existing problems, such as the spread of mis- and disinformation, inequality and social instability. For instance, traditional media that had already been under pressure from a changing media landscape are losing additional income as Google’s AI-generated search results limit redirects to news pages, which generated an income through advertisements.[13]
5.
Proliferation of AI inevitably comes at a security cost. Wider availability of powerful AI through open models increases the risk of misuse by malign actors, including in the cyber and biotech domains. The importance of AI safety and oversight of those trends will only increase.
6.
The rise of agentic AI changes the scale and autonomy of AI use. These systems can plan and act independently, increasing unpredictability.
7.
There is AI beyond LLMs. Future breakthroughs from AI will also come from less-discussed model types (for example, computer vision, human biology models, semantic and logic models), increasing the complexity and uncertainty in AI.
8.
Questions around model dependencies and trust arise as winners and losers emerge. The use of foreign models, or models from adversaries, raises concerns over software supply-chain integrity, hidden vulnerabilities and embedded biases.
9.
As AI improves, compute-based risk classification,[14] such as for General Purpose Artificial Intelligence (GPAI) in the EU’s AI Act, will be rendered useless within a short time span. Regulatory benchmarks based on computing power are quickly outdated because of fast-evolving optimisation methods. In other words, equally powerful models can emerge using a decreasing computational volume.
10.
Ecological and energy burdens are a structural constraint. Training and operating advanced AI consumes vast amounts of electricity, water and rare materials. These pressures are not only a side-effect of proliferation but an inherent feature of scale, creating new dependencies on energy systems and critical resources.[15]

Key Uncertainties: AI Access and Ubiquitousness

The baseline above shows the multi-pronged nature of the challenge for the EU and its Member States in relation to AI. Against this backdrop, this report assesses the potential impact of AI on national security based on two main uncertainties: (1) the degree to which the Netherlands, the EU and the public have access to cutting-edge AI capabilities, for example via open source AI proliferation or innovation; and (2) the ubiquitousness of the technology, for example depending on its incorporation into devices and the extent to which agentic AI lives up to the promises made by tech companies. These uncertainties arise as key determinants because they directly shape the power dynamics between state and non-state actors and the extent to which the Netherlands can act independently or must rely on external actors for AI capabilities.

In short, these are key ‘known unknowns’ with which the EU and its Member States must reckon. While these uncertainties make it difficult to plan for a single future, strategic foresight can help in preparing for a range of plausible developments and outcomes. As such, this report relies on input from a scenario workshop organised in May 2025 to gain insights from experts in AI. For more information on this research method, see the appendix.

The next section will dive into three Plausible Tomorrows, over a three- to five-year horizon. These Plausible Tomorrows, while speculative, engage in different ways with the two key uncertainties highlighted above: access to the latest AI technology and the extent to which AI becomes ubiquitous.[16] Figure 1 shows how the three Plausible Tomorrows interact with these uncertainties. An analysis of the risks and threats follows each scenario.

Figure 1
Graphical representation of how the three Plausible Tomorrows interact with the two key uncertain-ties identified: access to the latest AI technology and the extent to which AI becomes ubiquitous
Graphical representation of how the three Plausible Tomorrows interact with the two key uncertain-ties identified: access to the latest AI technology and the extent to which AI becomes ubiquitous
Please note that the items on this list are not ranked by importance.
Compounding the problem, the AI systems that Google and others deploy have been trained on data sourced from the same media outlets now under threat. This essentially happens without the media outlets’ consent, which has led, for instance, the New York Times to sue OpenAI and Microsoft. This development presents a serious threat for access to information and, ultimately, democracy itself.
Compute-based risk classification refers to the practice of assessing AI risk based on the amount of computational power (e.g. processing capacity or training compute) used to develop or operate a model.
Several technology firms have signed long-term deals to use nuclear energy to supply power to data-centre operations. While many projects are not yet fully operational, this trend suggests that corporations are exploring alternative energy-supply sources, with potential geopolitical and power balance implications. See, for example, Reuters, ‘Google Announces Tennessee as Site for Small Modular Nuclear Reactor’, 19 August 2025.
Of course, many other futures can be envisioned, for example: an AI bubble bursting, geopolitical dominance by either China or the United States, breakthroughs in quantum computing that relegate the current AI paradigm to a secondary role, or shifts along the open–closed model spectrum such as open weights. Moreover, a detailed discussion on aspects such as the importance of data ownership, interoperability and AI governance lie beyond the scope of this report because of space constraints, but they merit further analysis elsewhere.
Previous
Taking Stock: Where Does AI Stand?
Next
Plausible Tomorrow 1: AI Lights Up the Port of Rotterdam