The chance that a country like the Netherlands will actually get a serious AI factory is small. That’s according to Vladimir Prodanovic, principal program manager at Nvidia. The problem, he says, is that existing data centers in the Netherlands as well as many other European countries are built on air cooling and cannot be converted to the liquid cooling required for AI infrastructure.
The Dutch cabinet recently pledged €70 million for the construction of an AI factory in the northern province of Groningen. Together with €60 million from the Northern Netherlands and €70 million from Europe, this should result in a total investment of €200 million. The expertise center is scheduled to open in mid-2026, with the supercomputer fully operational in early 2027. But Prodanovic believes the factory will not be built.
Technically impossible
The Nvidia manager, who previously worked at Microsoft for many years, is adamant in his analysis during a Vertiv event in Italy. According to him, the major players based in the Netherlands are not interested in building an AI factory because their existing data centers are not suitable for this purpose. Instead, Dutch hyperscalers are looking to Spain, Italy, and Norway to purchase AI capacity from other parties.
He believes the problem lies in the fundamental architecture of the existing data centers. Dutch data centers are built for air cooling, while AI infrastructure requires liquid cooling. “In the Netherlands, we simply don’t have the capacity to build AI factories,” he says. “Because all those old data centers run on air conditioning and are not suitable for liquid cooling. And they cannot be converted either.”
Why conversion doesn’t work
Prodanovic’s statement is supported by technical analyses from the sector. According to Data Center Dynamics, existing data centers were often never designed to bear the static weight of manifolds and extra cabling for GPU clusters. In older data center spaces, the space under the floor is also already filled with decades of legacy cabling and air handling infrastructure.
CyrusOne, a major data center operator, emphasizes that converting existing data centers to liquid cooling requires major modifications, such as installing coolant pipes, reinforcing floors to support the additional weight of the installations, and creating space for distribution systems. This is costly and time-consuming. McKinstry points out that immersion cooling—where servers are literally immersed in a coolant—requires major renovations, especially in older facilities.
Data Center Frontier states that the cost of installing liquid cooling systems is significantly higher than traditional air cooling, especially in environments where existing infrastructure needs to be modified. According to the organization, retrofitting, modifying, or upgrading the infrastructure of existing data centers can be particularly challenging due to limited space, infrastructure constraints, and the requirement to maintain uptime.
The industry largely agrees: for AI workloads that require 30-60 kilowatts per rack or more, air cooling is simply inadequate. Traditional air cooling can handle at most 10-15 kilowatts per rack, according to multiple sources.
Billions in investment needed
The scale of the investment required only adds to the challenge. According to Prodanovic, a 100-megawatt data center would cost €6 billion: €1 billion for the data center itself and €5 billion for the AI component. “The government has just provided a significant amount of funding,” he notes, but the question is whether €70 million is sufficient for what is actually needed.
The Nvidia manager says that Microsoft uses liquid-to-air systems with a capacity of 150 kilowatts per rack, but he believes that this is insufficient for a real AI factory. The problem: these systems cannot dissipate enough heat and create overheated areas in the data center. For AI applications, you need to be able to place servers as close together as possible—preferably over multiple floors—because every meter of distance between servers slows down the speed. “Every meter is 5 nanoseconds of latency,” he explains.
Alternatives
Prodanovic also has little confidence in Belgium and Germany for building an AI factory. “I’m not really positive about the Netherlands, Belgium, or Germany,” he says. However, he does see opportunities in Southern Europe and Scandinavia. Deutsche Telekom is currently building an enterprise AI cloud in Munich, aimed at the small business market with clusters of 250 GPUs.
He cites Denmark as a positive example. There, the government has actively invested in a sovereign AI supercomputer and brought all Danish companies together. The country built the Gevion supercomputer through the Danish Center of AI Innovation and is now expanding further to create a sovereign cloud in which large Danish companies are participating.
Parties such as Nebul are now demonstrating that AI inferencing is possible in the Netherlands, albeit on a different scale than the large-scale AI factories Prodanovic refers to. The question is whether the plans for Groningen take sufficient account of this technical reality.
Prodanovic’s conclusion is clear: for a serious 100-megawatt AI factory, you have to build from scratch, which requires an investment of around €6 billion. The Dutch 200 million is a drop in the ocean. And even if the money were available, the fundamental problem would remain. “In the Netherlands, we simply don’t have the capacity to build AI factories,” he says.
Also read: How data centers are making the giant leap to 1 megawatt per rack