Why is that?

Previously, the direction for new AI models was always the same: higher quality meant more parameters and, thus, more required computing power, which correlated with high energy demand. This is not the case with DeepSeek because the model is structured differently: instead of maximizing raw computational power, it uses intelligent routing through specialized expert models (Mixture of Experts). Although more computing power can bring advantages here as well, the actual quality leap stems from the smarter architecture, not sheer size.

What exactly does this mean for infrastructure planning, specifically the construction of data centers, energy infrastructure, and data networks?

It means that some fantasies are unlikely to materialize. Recently, there were voices in the industry advocating for an unlimited expansion of gigantic data centers. However, if AI models become increasingly efficient and smaller, large-scale data centers may no longer be necessary in the long term. Instead, we may see smart networks of numerous decentralized, optimized edge models that process tasks locally and energy-efficiently. The focus will then be more on data technology for dynamically managing such networks, directing AI workloads to where free capacity is available.

There have already been considerations, primarily in Silicon Valley, to address the massively growing energy demand of AI through Small Modular Reactors (SMR) or even by reactivating decommissioned power plants. Are these considerations now off the table?

They're probably not entirely off the table. After all, there are other sectors besides AI with high energy demands, such as cryptocurrencies or cloud computing. Companies like Google or Amazon have been pursuing activities in the SMR space as part of a long-term strategy to take control of their energy supply rather than as a direct response to AI. We shouldn't forget that energy costs are often one of the largest cost drivers for digital companies. Therefore, we can expect tech companies to continue exploring technologies like SMR. However, the momentum is likely to decrease significantly.

To what extent do energy providers need to adjust their strategies to prepare for the growing energy demand from AI? Do you foresee a long-term paradigm shift in energy supply for AI systems?

There will need to be a shift in thinking. Until now, energy supply for AI systems has mainly focused on large amounts of power required at a single point. This was often seen as a problem for the energy transition by many energy providers, as the high base load demands of massive data centers are difficult to meet with renewable energy sources like wind power or photovoltaics.

However, if models that are far more efficient and rely on decentralized networks become the norm, AI could transform from a massive power consumer to a driver for smart grids — with local clusters that are self-sufficient through solar energy, battery storage, and intelligent load management. Energy providers will need to adapt to this and accelerate their smart grid initiatives to actively shape the change.

What new challenges arise from the combination of AI systems and intelligent energy networks?

First of all, this offers great opportunities. Smart grids require real-time processing of vast amounts of data to balance supply and demand and intelligently manage millions of small energy units. AI is precisely the right technology for this. However, new risks also emerge: when AI models make decisions, there must be robust mechanisms to prevent erroneous conclusions, unintended network instabilities, and cyberattacks.

For network operators, this primarily means one thing: cyber security will become essential — and in the long term, there's no way around building more software expertise.

Do you see synergies between industrial automation and new approaches to energy-efficient AI?

Yes, and they could be enormous. Industrial automation increasingly relies on AI, for instance, for predictive maintenance, process optimization, or robotics. So far, this often meant: more automation required more computing power. As a result, many applications have only worked in the cloud, with all its advantages and disadvantages.

Energy-efficient AI models, however, can be executed directly on the machine or within the local network. This not only saves energy but also reduces latency times and network loads. At the same time, costs decrease, making it more realistic for companies to develop their own AI models — thus driving technological advancement themselves. The synergies are obvious.

All these considerations are based on the assumption that DeepSeek will prevail. How realistic is it that a technology, partially subject to state censorship, will be adopted globally in the short to medium term?

I think it's important to distinguish between DeepSeek as a product and the underlying technology. As far as its spread as a product is concerned, censorship by the Chinese government is naturally an obstacle. Many companies will have serious concerns about data privacy. Therefore, after an initial hype, DeepSeek is likely to establish itself primarily in China — as a counterpart to Western LLMs within a Chinese AI ecosystem.

However, the technological influence is much more significant: DeepSeek's code is freely available as open-source. Companies or research institutions can develop their own versions based on the same energy-efficient principles but without political restrictions. Even if DeepSeek doesn't succeed as a product, it could still change the market in the long term.

The AI race has only just begun and has taken an exciting turn since mid-January.